Untitled

A Cognitive Approach to the Verb
WDE
G
Cognitive Linguistics Research
16
Editors
Rene Dirven
Ronald W. Langacker
John R. Taylor
Mouton de Gruyter
Berlin · New York
A Cognitive Approach
to the Verb
Morphological and
Constructional Perspectives
Edited by
Hanne Gram Simonsen
Rolf Theil Endresen
Mouton de Gruyter
Berlin · New York 2001
Mouton de Gruyter (formerly Mouton, The Hague)
is a Division of Walter de Gruyter GmbH & Co. KG, Berlin
© Printed on acid-free paper

which falls within
the guidelines of the ANSI
to ensure permanence and durability.
Die Deutsche Bibliothek — Cataloging-in-Publication Data
A cognitive approach to the verb : morphological and construc-

tional perspectives / ed. by Hanne Gram Simonsen ; Rolf Theil
Endresen. - Berlin ; New York : Mouton de Gruyter, 2001
(Cognitive linguistics research ; 16)
ISBN 3-11-017031-0
© Copyright 2000 by Walter de Gruyter GmbH & Co. KG, D-10785 Berlin
All rights reserved, including those of translation into foreign languages. No part of this book
may be reproduced or transmitted in any form or by any means, electronic or mechanical,
including photocopy, recording, or any information storage and retrieval system, without
permission in writing from the publisher.
Printing: Werner Hildebrand, Berlin
Binding: Lüderitz & Bauer, Berlin
Printed in Germany
Contents
Introduction
Hanne Gram Simonsen and Rolf Theil Endresen
Section I: Basic issues
Topic, subject, and possessor 11

Ronald W. Langacker
Language as a cognitive network 49

Richard A. Hudson
Section II: Morphological aspects of the verb
The Norwegian verb 73

Rolf Theil Endresen and Hanne Gram Simonsen
Building blocks or network relations:

Problems of morphological segmentation 95
Helge Gundersen
Past tense acquisition in Norwegian:

Experimental evidence 129
Hanne Gram Simonsen
Individual variation in past tense inflection:

Experimental data from Norwegian SLI children 153
Kirsten Meyer Bjerkan
vi Contents
A connectionist model of past tense acquisition

in Norwegian 165
Anders N0klestad
Regular inflection in the mental lexicon:

Evidence from Italian 189
Paolo. F. Matcovich
Section 111: Constructional aspects of the verb
Malefactive by means of GIVE 203

Ole Torfinn Fagerli
The Norwegian and Russian reflexive-middle-passive

systems and Cognitive Grammar 223
Hans-Olav Enger and Tore Nesset
Norwegian reflexive-middle-passive marker -s:

Homonymy/polysemy and morphology 243
Hans-Olav Enger
Control and transitivity: A study of the

Norwegian verb love 'promise' 259
Kristian Emil Kristoffersen
References 281
Index 305
Introduction
Hanne Gram Simonsen and Rolf Theil Endresen
1. Background
In June 1998, a small conference took place at Sanner Hotel, Gran,

Norway, with the title "The Verb in Cognitive Grammar:
Morphological, Semantic, and Syntactic Aspects". The bulk of the
participants at the conference had all for a longer or shorter time
been affiliated to a Cognitive Linguistics research group based at the
Department of Linguistics at the University of Oslo. This group,
named The Oslo Project in Cognitive Grammar (OPCoG), is headed
by Rolf Theil Endresen and Hanne Gram Simonsen, and has been
active since 1992. The conference was financed through a prize of
NOK 75.000 awarded to the research group by the Faculty of Arts,
University of Oslo.
There were two invited keynote speakers: Ronald W. Langacker
from the University of California at San Diego, and Richard A.
Hudson from University College, London. In addition, 18 other
papers were presented. The present book consists of a selection of
the papers, which have been through a thorough reviewing and
rewriting process.1
Given the topic of the conference, "The Verb in Cognitive
Grammar"—or, rather—"The Verb in Cognitive Linguistics", the
papers concentrated on a central theme and addressed this theme
from different angles, illustrated through different languages, but all
within a cognitive linguistics perspective. A majority of the papers
described different aspects of Norwegian—a language not earlier
investigated at such a scale within cognitive linguistics. Norwegian
offers valuable contributions through being closely related to
English, yet differing in certain important and interesting ways. In
addition, several other languages were included, providing a wider
typological range.
2 Hanne Gram Simonsen and Rolf Theil Endresen
In their papers, the two keynote speakers provided a broad

perspective and a general theoretical background for the main theme.
Richard A. Hudson drew mainly upon his own theory of Word
Grammar in explaining his view of language as a network, while
Ronald W. Langacker discussed and developed new ways of
characterising semantic and grammatical organisation within his
theory of Cognitive Grammar.2 We have placed these two papers
together in Section One of the book. The rest of the papers fall
naturally into two groups: those mainly addressing morphological
aspects of the verb—placed in Section Two, and those discussing
constructional aspects of the verb—placed in Section Three.
However, the division into three main sections does not mean that
there is a strict border between the sections, as several of the papers
touch upon themes crossing the section borders. In this way all the
papers can be seen as being in a network relation to each other ...
2. Section I: Basic issues
The first section starts with Ronald W. Langacker: Topic, subject,

and possessor. This paper explores the special, intimate relationship
between the grammatical notions topic, subject and possessor,
illustrating this relationship with examples from a wide range of
languages, and analysing it within the framework of Cognitive
Grammar. The notions are discussed in terms of reference point
relationships, where they are presented as prominent reference points
with respect to propositions, processes, or things, creating
asymmetric relationships with an inherently dynamic, temporal
dimension. The differences between them are further discussed as
related to different levels of organisation: the lexical, the
grammatical, and the discourse levels. This way, the paper presents a
new and integrated, conceptually based view of these phenomena.
Then follows Richard Hudson: Language as a cognitive network.
As indicated by the title, it develops the central idea of language
being a network, and discusses several consequences of this view,
relating them to cognitive theories like Langacker's Cognitive
Introduction 3
Grammar (Langacker 1987a, 1990, 1991), Goldberg's Construction

Grammar (Goldberg 1995a), Bybee's morphology (Bybee 1985,
1995a), and most thoroughly to his own theory, Word Grammar
(Hudson 1984, 1990, 1995, 1998). The basic similarities between
these theories, in spite of different terminologies and and starting
points, are underlined. He illustrates his arguments with examples
from two areas of language analysis: the polysemy/homonymy
distinction, and regular/irregular morphology. The examples from
morphology include a Word Grammar analysis of Norwegian verbs,
pointing forward to subsequent papers in the book.
3. Section II: Morphological aspects of the verb
This section contains six papers, most of them concerning

Norwegian. To give an overview of some relevant aspects of this
language which may function as necessary background knowledge,
we begin with Rolf Theil Endresen & Hanne Gram Simonsen: The
Norwegian verb. This paper gives a short survey of the Norwegian
verbal system, starting with some phonological notes, followed by an
outline of the morphological and periphrastic constructions and their
use, and summing up by an overview of the verb classes, their
relative sizes, and their phonological and morphological make-up. To
be able to evaluate different theoretical approaches to morphology
detailed knowledge both of phonological patterns and of frequency
patterns in the relevant language is needed.
Then follows Helge Gundersen: Building blocks or network
relations: Problems of morphological segmentation. Here two
morphological models are contrasted: the morpheme or building
block model, and the network model as presented in cognitive
linguistics. It is demonstrated how the requirement of exhaustive
segmentation of the building block model creates problems which do
not appear within a network model, illustrated through a special type
of Norwegian verbs, derived by the suffix -er. Gundersen argues that
while segmentation may be an important part of morphology, it is
neither necessary nor sufficient to account for morphological
4 Hanne Gram Simonsen and RolfTheil Endresen
relations, and he posits Langacker's and Bybee's theories as

alternative, superior accounts.
The remaining papers in this section all concern the acquisition of
past tense inflection of verbs, addressing the role of input factors
(phonological properties and type and token frequencies) during the
acquisition process. In this connection they also address the ongoing
debate as to which kind of model is better suited to account for the
acquisition and processing of morphology: a single mechanism
model or a dual mechanism model. According to the dual mechanism
model, there is a qualitative difference in processing between regular
and irregular verbs, while the single mechanism model assumes
essentially the same kind of processing for both regular and irregular
verbs, input factors playing an important role for both types of
inflection. These questions are investigated through different kinds
of experimental data in the following papers.
Past tense acquisition in Norwegian: Experimental evidence by
Hanne Gram Simonsen investigates the performance of Norwegian
4-, 6-, and 8-year-olds in an experimental task eliciting past tense
forms of verbs from different verb classes, and also comparing their
performance with that of adults. The results clearly indicate that
input factors play an important role, both for correct performance
and for the errors made, and also that they play a role for all verb
classes. Furthermore, the relative importance of each factor changes
in the course of development, in such a way as to lend support to a
single mechanism, exemplar-based model for morphological
acquisition.
Kirsten Meyer Bjerkan: Individual variation in past tense
inflection: Experimental data from Norwegian SLI children reports
on results from the same elicitation task applied to 6- and 8-year-old
Norwegian children with specific language impairment. The different
response patterns in the children are discussed as results of different
kinds of schema formation and entrenchment. Anders N0klestad: A
connectionist model of past tense acquisition in Norwegian presents
a connectionist model of past tense acquisition of Norwegian,
comparing the model's behaviour to results of the past tense
elicitation experiments presented in the preceding papers. The
Introduction 5
modelling results show that a single mechanism is able to learn to

produce past tense forms of Norwegian verbs, on the basis of the
corresponding present tense forms. The developmental profile of the
model shows striking similarities with the other experimental data,
but also some very interesting differences, which are given a
thorough discussion, both in the light of the simulation architecture,
and in the light of the different models proposed.
The final paper in this section is Paola F. Matcovich: Regular
inflection in the mental lexicon: Evidence from Italian. Closely
related to the preceding papers, it investigates the acquisition and
processing of Italian verb inflection in 4-, 6-, and 8-year-old children
and adults through two sets of experiments—eliciting past participle
forms of existing verbs as well as nonce verbs. Again, the results
offer support for a single mechanism model, since no qualitative
difference is found between regular and irregular inflection with
respect to the influence of phonological or token frequency factors.
4. Section III: Constructional aspects of the verb
The second section contains four papers, starting with Öle T. Fagerli:
Malefactive by means of GIVE. Fagerli discusses semantic aspects of
certain African languages—more specifically the benefactive and the
malefactive, both in serializing languages using a word identical to
the verb GIVE, and in languages using verbal extensions to express
these semantic relations. In both language types malefactive events
may be marked by expressions used also for benefactive events.
Seeing both types of expressions as special instances of a cognitively
and semantically basic GIVE relation, the author argues that the
malefactive meaning may be seen as an extension of the originally
benefactive marker, in combination with the semantics of the host
verb and its complements.
In the next paper, Hans-Olav Enger and Tore Nesset: The
Norwegian and Russian reflexive-middle-passive systems and
Cognitive Grammar, the authors give an extensive analysis and
comparison of the reflexive-middle-passive systems in Norwegian
6 Hanne Gram Simonsen and RolfTheil Endresen
and Russian, and show how they can be presented in Cognitive

Grammar networks with Schemas and extension relations. Their
analysis illustrates the value of networks of this type both as heuristic
devices and as valuable tools for gaining new insight, particularly in
typological research.
Hans-Olav Enger: The reflexive-middle-passive marker -s in
Norwegian: Homonymy/polysemy and morphology singles out one
Norwegian suffix marking a set of the meanings examined in the
previous paper, discussing whether it should be counted as one
polysemous suffix or several homonymous ones. Arguing in favour
of the polysemy position, the author examines the status of the suffix
as inflectional or derivational in different functions, relating his
discussion to Bybee's (1985) concept of relevance. Both in its
content and its discussions, this paper also ties back to Hudson's
paper.
Finally follows Kristian Emil Kristoffersen: Control and
transitivity: A study of the Norwegian verb love 'promise'. Like the
preceding papers, this one also exploits the notion of polysemy, but
this time not of lexical items, but of constructions. The author
investigates the relation between transitivity and argument control in
one Norwegian verb in a Construction Grammar perspective. It is
argued that argument control should be treated as a property not of
the individual verbs, but of the constructions in which the verb may
occur. The specific, different patterns of transitivity and argument
control concerning the verb love are explained through the notion of
multiple inheritance from different construction types. Furthermore,
the frequency of use of the different love constructions is seen as a
result of differences in type frequency of the constructions.
5. Final remarks and acknowledgements
All the authors have profited from the fruitful and active discussion
of all the participants during the conference. Of particular value was
the active participation of the two keynote speakers, Richard A.
Hudson and Ronald W. Langacker, who deserve special thanks.
Introduction 7
The editors of the Cognitive Linguistics Research series gave very

useful advice and comments during the publishing process.
We are also grateful to the Faculty of Arts, University of Oslo, for
providing financial support. Finally, as convenors of the conference,
we would like to thank our conference secretary, Ame Martinus
Lindstad. He has been an invaluable coworker at all stages, both at
the conference and through the finalising of the book manuscript.
Notes
1. Reviewing is a time-consuming job. We wish to thank the following persons

for taking of their time to review the papers in this volume, giving valuable
suggestions, comments and evaluations: Hans Basb011, Elizabeth Bates,
Andrew Carstairs-McCarthy, Bruce Derwing, Wolfgang Dressier, Lena
Ekberg, Hans-Olav Enger, Richard Hudson, Laura Janda, Suzanne Kemmer,
Ronald W. Langacker, Virginia Marchman, Tore Nesset, Eric Pederson, Trond
Trosterud, Äke Viberg, Wolfgang Wurzel.
2. This paper by Ronald Langacker was first presented at the Bielefeld
Symposium in April 1998, and will be published in the proceedings from that
symposium. We are grateful that the arrangement committee for the Bielefeld
Symposium has agreed to our publication of Langacker's contribution in the
present book.
Section I
Basic issues
Topic, subject, and possessor
Ronald W. Langacker
Language offers ample evidence that the grammatical notions topic,

subject, and possessor have special affinities for one another. The
basis for their close relationship is explored from the standpoint of
cognitive grammar. They are all described as instantiating a basic
aspect of dynamic conceptualization involving successive foci of
attention: a salient reference point affords mental access to a context
(its dominion,) in which a target is located. A possessive construction
is characterized schematically as specifying a reference point
relation between two things. A topic construction establishes a
reference point relation between a thing and a proposition. Subject
and object are analyzed as reference points with respect to the
relationship designated by a verb or a clause (called a process): they
represent the initial and secondary focal participants accessed in
building up to the full relational conception. Evidence is provided for
their focal prominence, for their asymmetry in this regard, for this
asymmetry having a temporal dimension, and for its assimilation to
reference point phenomena. The relation between topic and subject is
discussed in terms of function and levels of organization.'
1. Evidence of special affinities
I start from the observation, not at all a novel one, that the gramma-
tical notions topic, subject, and possessor display certain affinities
for one another. Numerous phenomena, in particular languages and
cross-linguistically, suggest that these constructs have a special and
intimate relationship. My goal is to find a precise description of the
abstract similarity that is presumably responsible for their affinity.
This supports the semantic characterization of subject and object
12 Ronald W. Langacker
proposed in cognitive grammar (Langacker 1991:7.3). It also bears

on the internal conceptual organization of verbs and clauses.2
Let me start by offering some empirical evidence for topic,
subject, and possessor being specially related in some fashion. I will
merely list certain phenomena that have come to my attention. The
following inventory is not intended to be exhaustive, nor have I
undertaken a systematic survey. Other linguists could no doubt easily
add to the list.
The close connection between subject and topic is of course well
known, if not notorious. Linguists often have difficulty determining
whether a given nominal element is best analyzed as a subject, a
topic, or both. Moreover, topic status is commonly ascribed to
subjects as part of their characterization, at least prototypically.
Givon, for instance, describes subject and object as primary and
secondary clausal topic (1984:138). I basically agree with this
description and will try later to render it less vague.
Diachronically, topics often evolve into subjects. Li and
Thompson (1976:484) describe this recurrent path of gramma-
ticization as follows: "...it seems clear that subject and topic are not
unrelated notions. Subjects are essentially grammaticalized topics; in
the process of being integrated into the case frame of the verb (at
which point we call them subjects), topics become somewhat impure,
and certain of their topic properties are weakened, but their topicness
is still recognizable...That is why many of the topic properties are
shared by subjects in a number of languages. For example, some
[subject-prominent] languages do not allow indefinite subjects.".
A second observation concerns the notion pivot. I define a pivot as
the entity within a proposition that corresponds to a topic and thus
enables the proposition to function as a comment with respect to that
topic. In (1), for example, the topic is your friend, and the pivot is the
coreferential pronoun in the following clause.
I am reasonably confident that systematic investigation would
reveal a strong tendency for the subject of a comment clause to
function as pivot. Secondarily, the object of the comment clause
often serves as pivot, or else the possessor of the subject or object.
The expressions in (1) are, I believe, quite typical.
(1) a. Your friend, he will never amount to anything.

b. Your friend, I really like him.
c. Your friend, his mother just died.
d. Your friend, I really liked his mother.
Data showing an intimate relationship between subject and topic in

Russian is presented in Nichols, Rappaport, and Timberlake (1980).
The authors discuss a variety of control phenomena—involving
reflexivization, adverbial participles, predicate agreement, and case
agreement of predicate nominatives—in which the controller is
typically the grammatical subject. For each, they claim that topic
status (signaled by preverbal position) is a factor that affects the
possibility of control in less typical instances. The authors conclude
that subject and topic status (also animacy) are special cases of
nominal prominence, which favors control. This is quite compatible
with the account I will offer below.
Let us next consider phenomena indicating a special relationship
between subject and possessor. Some of these are well known. For
instance, when a verb or a clause is nominalized, it is extremely
common cross-linguistically for its subject or object to be specified
periphrastically by means of a possessive construction:
(2) Booth's assassination [of Lincoln];

Lincoln 's assassination [by Booth]
Another such phenomenon, found in French, Norwegian, and

numerous other languages, is for a subject nominal to simultaneously
function as possessor (in this case of a bodypart object) even in the
absence of any specific possessive marking, consider (3):
(3) a. II leve la main. / Han rekker opp handa.

'He raises his hand.'
b. J'ouvre la bouche. /Eg opnar munnen.
Ί open my mouth.'
Phenomena specifically relating topic and possessor may be less

common, but they do exist. One example is the prevalence of non-
subject possessor pivots, as in (Id). Another interesting case is found
in spoken French, as described by Lambrecht (1996). It is possible
for multiple topics to occur preposed to the comment clause, as seen
in (4a). Sentence (4b) is slightly different, however, since only the
sister has any direct role in the comment clause. Strictly speaking,
then, only sa soeur 'his sister' functions as topic with respect to that
clause. The first preposed nominal, Pierre, instead functions as
possessor for the second, sa soeur. While this looks just like a
double-topic construction, in actuality it represents the combination
of a possessive and a topic relationship.
(4) a. Nicole, Pierre, eile ne I'aimepas.

'Nicole, Peter, she doesn't like him.'
b. Pierre, sa soeur, je la deteste.
'Peter, his sister, I hate her.'
There are also phenomena where topic, subject, and possessor all
figure crucially, suggesting that all three have some special affinity.
Li and Thompson (1976) report that in Indonesian the pivot in a topic
construction can only be the subject of the comment clause or a
possessor of the subject:
(5) a. Ibu anak itu membeli sepatu.

mother child that buy shoe
'That child's mother bought shoes.'
b. Ibu anak itu, dia membeli sepatu.
mother child that she buy shoe
That child's mother, she bought shoes.'
c. Anak itu, ibu-nja membeli sepatu.
child that mother-his buy shoe
That child, his mother bought shoes.'
d. *Sepatu itu, ibu anak itu membeli.
shoe that mother child that buy
Those shoes, that child's mother bought them.'
For Korean, Cho (1991) points out that certain phenomena which
normally involve subjects can instead be controlled by the possessor
of an inanimate subject, provided that the possessor is part of a
topicalized constituent. One such phenomenon is subject
honorification:
(6) a. Apeci-ka Kim sensayngnim-ul mana-si-ess-ta.

father-SUBJ Kim teacher-OBJ meet-HON-PAST-ASSR
'Honorable father met teacher Kim.'
b. Halmeny-uy sayngsi-eyse-nun samsiptay-ka kacang
gr.m.-POSS life-LOC-TOP thirties-SUBJ most
hayngpokha-si-ess-ta.
happy-HON-PAST-ASSR
'In honorable grandmother's life, the thirties were the
happiest.'
Control of the reflexive marker casin works analogously.

Finally, I note the prevalence in many languages of what are often
called "double-subject constructions" (Langacker 1999a). Illustration
is provided by Luiseno, a Uto-Aztecan language of southern
California (Steele 1977):
(7) a. Noo=n no-puush konoknish.

I=1SG:PRES my-eye green
Ί have green eyes.'
b. Xwaan=up po-toonav qala.
Juan=3SG:PRES his-basket sits
'Juan has a basket.'
c. Noo-up no-te' tiiwu-q.
I=3SG:PRES my-stomach hurt-PRES
Ί have a stomach ache.'
These expressions have the following abstract form:
(8) [ noo-up [ no-te' tiiwu-q ] ]

outer inner
subject subject
inner clause
full clause
Depending on the language, there is evidence for both the "outer

subject" and the "inner subject" having subject status at some level,
and also for the entire expression constituting a single clause. In the
case of Luiseno, for instance, the tense/agreement clitic which
regularly follows the first element of a clause appears in this
construction after the outer subject (indicating that the full clause is
in fact a single clause). Moreover, it can agree with either the outer
subject (as does -n in (7a)) or else the inner subject (as does -up in
(7c)).
I analyze the outer subject in this construction as being both a
subject and a topic with respect to the full clause. Furthermore, the
construction is strongly associated with possessive relationships,
prototypically involving bodypart relations. More specifically, the
outer subject is usually understood as the possessor of the inner
subject. Luiseno requires that this possessive relationship be
explicitly marked by means of a possessor prefix. In other languages
(e.g. Japanese and Korean) it remains implicit (cf. (3)).
I am sure that this inventory can be greatly expanded. But even
from this short, random list, it should be evident that something of
interest is going on. To what can the intimate relationship among
topic, subject, and possessor reasonably be attributed?
2. Some basic notions of cognitive grammar
Naturally, I will approach this question from the standpoint of

cognitive grammar (Langacker 1987a, 1990, 1991). Its central claim
is that grammar is symbolic in nature. More specifically, lexicon,
morphology, and syntax form a continuum fully describable as
assemblies of symbolic structures. A symbolic structure resides in the
pairing between a semantic structure (some kind of con-
ceptualization) and a phonological structure (its semantic and
phonological poles). It follows that all elements posited in
grammatical description must have some kind of semantic value
(often highly schematic).
The possibility of a symbolic account of grammar presupposes a
conceptualist semantics that properly accommodates construal, i.e.
our multifaceted ability to conceive and portray the same content in
alternate ways. Lexical and grammatical elements incorporate
particular ways of construing conceptual "content", as an inherent
part of their conventional meanings. Some facets of construal
selected for brief illustration are level of specificity (conversely
schematicity), perspective, ana prominence.
Degrees of specificity are exemplified by lexical hierarchies such
as (9). Each expression is schematic with respect to one that follows,
which elaborates its coarse-grained specifications (i.e. spells them
out in finer-grained detail).
(9) thing>creature>bird>bird of prey>hawk>red-tailed hawk
The meanings of grammatical elements tend to be quite schematic.

To put it another way, the more schematic an element is
semantically, the more inclined linguists are to consider it
"grammatical" rather than "lexical". Naturally I view this as a false
dichotomy.
Perspective includes numerous factors, only one of which will be
exemplified:
(10) a. A big crack extends from the corner of the window up to

the ceiling.
b. A big crack extends from the ceiling down to the corner
of the window.
These sentences describe precisely the same, static situation—

nothing actually moves. Yet intuitively they exhibit some kind of
motion or dynamicity, and contrast semantically owing to its
opposite directionality in the two cases. I attribute this to subjective
motion on the part of the conceptualizer (i.e. the speaker, and
secondarily the addressee), who mentally scans through the scene in
one direction or the other in "building up" to its full
conceptualization (Langacker 1986). Subjective or "fictive" motion
is actually quite prevalent and varied (Matsumoto 1996a, 1996b;
Talmy 1996; Langacker to appear).
Two kinds of prominence are especially relevant to grammar. The
first is profiling. Within its conceptual base (i.e. the array of content
it evokes as the basis for its meaning), an expression designates
(refers to, directs attention to) some substructure, called its profile.
This is best seen from expressions that evoke essentially the same
content but nonetheless differ in meaning.
(a) Base iris pupil
(") Base admire admirer
O-O O-«O O--O tr Im
Figure 1. Base and profile
As shown in Figure l(a), for instance, the conception of the human

eye provides the common conceptual base for terms like im and
pupil, which are certainly not synonymous (even if one's knowledge
of their referents is limited to the configurational properties
depicted). They contrast semantically by virtue of profiling—
referring to—different sub-parts of the base. Observe that profiling is

indicated by heavy lines.
Expressions can profile not only things (abstractly defined) but
also relationships. An example is given in Figure l(b). The verb
admire profiles a relationship where one individual manifests a
certain mental attitude in regard to another.3 The noun admirer
evokes the same conceptual content—one cannot apprehend its
meaning without conceptualizing the process of admiring—yet it
surely contrasts semantically with the verb. They have different
meanings because the noun profiles a thing, specifically the
participant displaying the mental attitude.
Expressions that evoke the same content and profile the same
relationship can also differ in meaning due to another kind of
prominence, namely the degree of salience accorded to the
participants in the profiled relation. For instance, as illustrated in
Figure 2, above and below are semantically distinct despite invoking
the same elements and profiling the same spatial relation
(referentially, an above relationship is also a below relationship).
(a) above (b) below
?
<*>
Figure 2. Trajector/landmark alignment
Their semantic contrast resides in the fact that above takes the lower
participant as a landmark for purposes of specifying the location of
the higher one, whereas below does just the opposite. To capture such
differences, I say that a relational expression accords focal
prominence to one or more participants. It is usual for one
participant—termed the trajector (tr)—to stand out as the primary

focal participant in the sense of being the one the expression is
concerned with locating, characterizing, or assessing in relation to
others. A secondary focal participant, with respect to which the
trajector is located or evaluated, is called a landmark (Im).
An expression's grammatical class is not determined by its
conceptual content, but by the nature of its profile (Langacker
1987b). A noun profiles a thing. A verb profiles a process, defined as
a relationship whose evolution through time is rendered salient. As
shown above in Figure l(b), for example, admirer is a noun because
it profiles a thing, whereas admire is a verb because it profiles a
temporally extended relation. Classes like adjectives and prepositions
involve the profiling of non-processual relations. For instance,
although the relationships profiled by above and below typically
endure through time, this temporal dimension is not essential to their
semantic characterization.
Alice admires Bill
Alice admires Bill
admires
Figure 3. A symbolic assembly
A grammatical construction comprises a set of component symbolic

structures that are integrated to form a composite symbolic structure.
Integration depends on correspondences between elements of the
component structures. Consider the lower right-hand portion of
Figure 3, showing how the component structures admires and Bill are
integrated to form the composite structure admires Bill. Their
integration hinges on a correspondence between the profile of Bill
and the schematic landmark of admires, which it serves to elaborate.4
Corresponding elements are superimposed and merged (or "unified")
to form the composite structure. The composite structure admires Bill
is therefore specific in regard to its landmark, which inherits the
semantic specifications of the component Bill (abbreviated B). It is
usual for the composite structure to inherit its profile from one of the
components. The component structure which bequeathes its profile
upward is called the profile determinant, identified by being enclosed
in a heavy-line box. Admires is thus the profile determinant in
admires Bill, for the composite expression profiles the same
relationship as the verb. The profile determinant constitutes the
grammatical head, at a given level of organization.
Grammar resides in conventional patterns for combining symbolic
structures into complex symbolic assemblies. I claim that such
patterns are represented by symbolic assemblies that are partially or
fully schematic, i.e. templates that embody the abstract commonality
exhibited by particular complex expressions. These constructional
schemas govern the formation of novel expressions. For instance,
admires Bill instantiates a constructional schema that specifies the
integration of a verb with its object by virtue of a correspondence
between the former's landmark and the latter's profile. Constructional
schemas will not be a prime concern here, however.
A composite structure at one level of organization (in one
construction) can in turn function as component structure at a higher
level of organization (in a higher-order construction). In Figure 3, for
example, we see that the composite structure admires Bill is
integrated at a higher level of organization with the component Alice,
which elaborates its schematic trajector to form the higher-order
composite structure Alice admires BilL Symbolic assemblies with
this character thus exhibit a kind of constituency. However, I have

argued elsewhere (1995a, 1997a) that constituency of the classic sort
is far less consistent or important than is assumed in most theories.
Symbolic assemblies with the properties just described support
what I believe to be a viable characterization of the grammatical
notions subject and object: at a given level of organization, a subject
is a nominal whose profile corresponds to the trajector of a profiled
relationship, and an object, one whose profile corresponds to the
landmark. Thus, considering both the "horizontal" and the "vertical"
correspondences in Figure 3, Bill is identified as the object of
admires, admires Bill, and Alice admires Bill, for its profile
corresponds to their landmark. Likewise, Alice is identified as their
subject, since its profile corresponds to their trajector. Observe that
this characterization is not based on constituency. It is not
configurational but rather conceptual, and alternate constituencies
(e.g. a flat structure) yield the same grammatical relations. This
illustrates the notion that constituency is neither fundamental nor
essential to grammatical structure. It is a feature of symbolic
assemblies that sometimes emerges, but not always, and is often
flexible and variable when it does. The key work of grammar and
semantics is done instead by correspondences, profiling, and other
aspects of conceptual structuring.
3. Reference point phenomena
As I view it, conceptualization is inherently dynamic (Langacker

1997b). By this I mean that it develops and manifests itself through
processing time, the specific nature of its temporal development
being essential to its value. This was shown quite clearly in (10),
where the semantic contrast resides in the directionality of the
conceptualized mental scanning through the scene.
One aspect of dynamic conceptualization is our very general
cognitive ability to direct attention to some entity for purposes of
then establishing "mental contact" with another. This attention may
be perceptual, as in (1 la). A more abstract example is given in (1 Ib).
(11) a. Do you see that boat out there in the middle of the lake ?
There's an otter in the water just alongside it.
b. Do you remember the surgeon we met at your sister's
party? His wife just had quintuplets.
In such cases I speak of a conceptual reference point (R) providing

mental access to a target (T) found in its dominion (D). By serving as
an initial focus of attention, the reference point tends to activate an
array of associated entities which collectively constitute its
dominion. This enables attention to shift to a secondary focus within
it, the target (which activates its own dominion and may in turn
function as a reference point). The components of a reference point
relationship are depicted in Figure 4.
This reference point ability has many linguistic manifestations.
For instance, it plays a role in the sanction of definite articles, whose
presupposition of contextual uniqueness can be explicated in terms of
the speaker and addressee establishing mental contact with the same
conceived entity (Langacker 1991:89-103). Consider examples like
(12), where the definite article occurs even in the absence of previous
mention:
(12) a. My car is no longer dependable.

The motor is just about shot.
b. / would never buy this house. The roof leaks badly.
The noun car evokes an established frame, including the standard

knowledge that cars have motors, and that a given car has just one.
Likewise, house evokes a frame that includes the notion of a (single)
roof. Thus in (12a), my car directs attention to a specific car, which
can thereby function as a reference point. The frame consisting of our
general knowledge of cars specifies a set of entities accessible via
this reference point, among which are its basic parts (motor, steering
wheel, tires, etc.). Because the dominion defined in this manner con-
tains just a single motor, the intended referent of motor in the second
sentence is contextually unique, hence the definite article is used.
C = conceptualize!
R = reference point
T = target
D = dominion
= mental path
Figure 4. A reference point relationship
Another reference point phenomenon is metonymy, where an

expression's usual referent (R) is able to evoke its intended referent
(T). The term Chicago, for example, refers primarily to a city, but in
the proper context it might instead be construed as referring to any
number of associated entities comprising its dominion. Possibilities
include its populace, a sports team located there, the city government,
the physical embodiment of a grant application its government has
submitted, etc.
(13) a. Chicago is all excited about the Bulls.

b. Chicago has just won another championship.
c. Chicago was late in sending in its application.
d. Chicago is right here in the stack between Dallas
and Memphis.
I will mention only in passing a reference point analysis of

pronominal anaphora worked out in great detail by Karen van Hoek
in the context of cognitive grammar (van Hoek 1995, 1996, 1997a,
1997b). The basic idea is that a pronoun needs to occur in the
structural dominion associated with a coreferential reference point
salient in the current discourse space, whereas a full NP may not
occur in such a context. I cannot go into any detail here. Instead we
will concentrate on two other grammatical phenomena that I analyze
in terms of reference points: possessive and topic constructions.
A central problem posed by possessive constructions is the

extraordinary variety of relationships they are able to encode. The
expressions in (14a) are but a small sample of the kinds of situations
coded by prenominal possessives in English. While a case can be
made for ownership, kinship, and whole/part relations being
prototypical, it is by no means obvious that all the others can
reasonably be considered metaphorical extensions from these. It
seems as though virtually any kind of association is susceptible to
being expressed in this manner. Yet that is not quite adequate. In and
of itself, an associative relationship is symmetrical. As we see in
(14b), however, possessive expressions are usually not reversible.
(14) a. the boy's shoe; Jeffs uncle; the cat's paw; their lice;
the baby's diaper; my train; Sally's job; our problem;
her enthusiasm; its location; your candidate; the city's
destruction
b. *the shoe's boy; *the paw's cat; *the diaper's baby;
*the destruction's city
Apparently the semantic value of a possessive element or

construction needs to be both abstract (hence broadly applicable) and
asymmetrical (resulting in irreversibility). Elsewhere (Langacker
1993, 1995b) I have argued in some detail for a reference point
characterization (see also Taylor 1996). Specifically, in a possessive
construction the possessor can be described schematically as a
reference point, and the possessed, as a target found in its dominion.
In all the reference point phenomena described so far, the
reference point and target have each been things. In pronominal
anaphora, these two things are the profiles—taken as being
referentially identical—of a pronoun and its antecedent. For
metonymy, the things in question are alternate referents of the same
nominal expression: its usual referent, and the metonymic referent
intended in the context. In a possessive construction, the things in
question are distinct and are designated by separate nominals, though
one may be a constituent of the other. For instance, [[the boy]'s shoe]
invokes a reference point relation between a particular boy and a
particular shoe. The reference point is coded by the boy, and shoe
specifies the intended target in its dominion. Hence the full nominal
the boy's shoe singles out and profiles the shoe being referred to.
Frame-based uses of the definite article, as in (12), also involve the
referents of separate nominals (often a possessive construction could
in fact be employed to describe their relationship, e.g. my car's
motor). The difference is that the two nominals are separate and the
reference point relation remains implicit. Whereas a possessive
construction specifically encodes the reference point relationship, in
(12) it is merely presupposed as part of the background knowledge
that assures the contextual uniqueness required by the definite article.
The reference point notion is not inherently limited to things,
however. In particular, I propose that a topic construction expresses a
reference point relationship between a thing and a proposition. This
is diagrammed in Figure 5. The target proposition (given as a
rectangle) is coded by a clausal expression minimally consisting of a
thing (shown as a circle) engaging in some process (solid arrow).
Figure 5. A topic relationship
I believe this characterization properly captures the dynamic nature

of topic constructions, wherein the topic is presupposed as the basis
for interpreting the comment clause and thus has a certain conceptual
priority with respect to it. A topic is sometimes described as a
"mental address" to which the comment proposition is intended to be
delivered. In terms of the reference point model, we can say that the
topic evokes a domain of knowledge (its dominion) in which the
target proposition is to be integrated. Equivalently, we can describe
the topic as providing mental access to a proposition that can be

found in its dominion. Whether we speak of the proposition as
(already) being located in the topic's dominion, or as needing to be
incorporated into it, is simply a matter of whether we take the
speaker's perspective or the listener's.
Consider (15), which involves two reference point relations:
(15) My car, the motor is just about shot.
First, my car functions as reference point with respect to the motor,

its dominion providing the contextual uniqueness required for
felicitous use of the definite article. There are many motors in the
world, but the speaker has a particular motor in mind, uniquely
identifiable in relation to the speaker's vehicle in accordance with the
frame semantics of car. For the speaker, the motor is already
identified and already in the car's dominion. On the other hand, the
listener may not have been thinking about motors at all, let alone any
particular one. For the listener, the motor constitutes an instruction to
conceive of an instance of the motor category, and signals its
contextual uniqueness, but that alone does not identify it.
Identification comes from interpreting the motor in relation to my
car—it is a matter of integrating the conceived instance of motor in
the dominion of that reference point.
As a topic, my car is also a reference point with respect to the
comment clause the motor is just about shot. For the speaker, this
proposition is already known and ascribed to the proper domain—it
belongs to the dominion of the speaker's car. It is "identified" in the
sense that the speaker knows it pertains to a particular motor
(associated with a particular car), as opposed to all the other motors
an analogous proposition might describe. On the other hand, the
listener may have no prior knowledge of the situation. The motor is
just about shot might represent any number of distinct propositions,
depending on which particular motor it pertains to, so to identify the
specific proposition actually intended the listener needs instructions
for delivering it to the correct mental address. The topic my car
provides these instructions, enabling the listener to integrate the

proposition in the proper domain of knowledge.
Two points worth noting emerge from this example. First, a nomi-
nal put in focus by being invoked as a reference point for one pur-
pose is quite susceptible to being simultaneously exploited as a refe-
rence point for some other purpose. Second, one reference point re-
lationship may create the potential for another. In (15), for example,
it is by virtue of the motor being in the car's dominion that the com-
ment clause the motor is just about shot is applicable to the topic my
car. If the motor were not construed as the car's motor, no basis for
integrating the comment into the topic's dominion would be apparent.
4. Reference point analysis of subject and object
I hope to have made a reasonable case for analyzing both possessive

and topic constructions in terms of reference point relations. Not only
does this make good sense for each type of construction individually,
but it also accounts for the possessor-topic affinity: possessors and
topics are each characterized as nominal reference points (with
respect to things and propositions, respectively), and an entity
invoked as a reference point for one purpose is likely to be invoked
for another as well. Recall now that subjects show a strong affinity to
both possessors and topics. This suggests that subjects should also be
characterized as nominal reference points. I will therefore sketch an
analysis along these lines and try to support it.
I have described subjects and objects as nominal expressions
which respectively elaborate the trajector and (if there is one) the
landmark of a profiled relationship. A reference point analysis of
subjects and objects thus implies that a relational participant achieves
the status of trajector or landmark by virtue of serving some kind of
reference point function. This proposal has initial plausibility given
the independently developed descriptions of trajector/landmark status
on the one hand, and a reference point/target relation on the other.
Trajector and landmark are characterized as the primary and
secondary focal participants of a profiled relationship. Similarly,
reference point and target are reasonably described as initial and

secondary foci of attention, the first affording access to the second.
As theoretical constructs, the notions trajector and landmark are
first posited and strongly motivated for purposes of achieving
minimally adequate semantic descriptions. An asymmetry in
participant salience is the only evident basis for distinguishing the
meanings of pairs of expressions like above and below, which evoke
the same content and profile the same relationship (Figure 2). Many
such pairs exist: in front of/in back of, precede/follow, like/please,
etc., as well as actives and their corresponding passives.
Metaphorically, I conceive of trajector and landmark status in terms
of spotlights of focal prominence, a stronger and a weaker spotlight,
that can be directed at various elements of a complex scene. By way
of making this idea more precise, we can isolate the factors listed in
(16).
(16) a. The trajector and landmark of a profiled relationship are

distinguished from other relational elements by the focal
prominence conferred on them.
b. Their prominence is asymmetrical, trajector and land-
mark being describable as primary and secondary focal
elements.
c. This prominence asymmetry has a temporal dimension;
the trajector's primary focal prominence resides at least
partially in its role as initial point of access.
d. As successive foci, trajector and landmark are properly
described as reference point and target, or equivalently,
as first and second elements in a reference point chain.
Let me briefly consider some of the evidence supporting each of

these points, taking them in the order listed.
That trajector and landmark—respectively coded by subject and
object nominals—are distinguished by some kind of special
prominence is, I think, fairly evident. Linguistic scholars are in
virtually unanimous agreement that the subject and object are in
some sense the "core" arguments of a verb, and typically they
represent its only obligatory complements. Subject and object are

especially active grammatically, which I take as being symptomatic
of their special salience or accessibility. In particular, subject and
object are the only elements with which verbs commonly show
"agreement". I view such agreement as a matter of the greater
specificity that is naturally associated with entities placed in the
focus of attention, and the greater concern with clearly indicating
their role in the event.
Granted the focal prominence of trajector and landmark, their
asymmetry in this regard is quite apparent. For one thing, subjects are
generally more active and grammatically accessible than objects
(specific cases are mentioned below). A more fundamental
observation is that a relational expression may have a trajector
without a focused landmark, but not conversely. A full justification
of this statement requires numerous points of clarification and
analysis that I cannot go into here (e.g. with respect to ergativity, as
well as various types of impersonal constructions). For present
purposes, let us simply note the basic pattern across languages for a
single-participant clause to have a grammatical subject rather than an
object, even when the subject's semantic role is the same as that of a
transitive object:
(17) a. Alexander broke the vase, [subject and object]

b. The vase broke. [subject only]
c. *Broke the vase. [object only]
I believe one can ultimately sustain the generalization that a

landmark presupposes a trajector, whereas a trajector does not
presuppose a landmark. This follows from their respective
characterizations as primary and secondary focal participants: a focal
participant can be secondary only if there is also a primary one—
otherwise it is the primary focal participant by default.
Is it plausible in psychological terms to assume that our
apprehension of a situation may involve two foci of attention, one
primary and the other secondary? We can readily observe this ability
in a non-linguistic example. Suppose you are watching a boxing
match. You can easily focus your primary attention on one of the
boxers, observing his footwork, which hand he prefers, how well he
defends himself, how many blows he absorbs, etc. For a given action,
this focusing is possible whether the boxer in question has an active
or a passive role; in either case, the action is viewed and interpreted
in relation to him. But the other boxer has a prominent role as well. If
you focus on boxer A and see him deliver a blow to boxer B, it is
evident that B is also central to your conception. As a main
participant in the observed event, he clearly stands out as a secondary
focus (in contrast to the referee, the announcer, the audience, and so
on).
Experimental work by Tomlin (1995, 1997; Forrest 1996) yields
striking evidence for characterizing subject as primary focus of
attention. In multiple experiments with slightly different designs, the
experimental subjects were made to observe a scene and immediately
describe it with a simple sentence. The scenes consisted of either
dynamic events or static situations, each with two salient
participants. Attention was directed to one of the two participants
(e.g. by an arrow flashed on the screen) just prior to presentation of
the critical event or configuration, the interval being too brief for
attention to wander in between. Tomlin's hypothesis (specifically for
English) was that the speaker would choose the focused participant
as the syntactic subject of the utterance produced. This proved to be
the case, virtually 100% of the time.
Tomlin's experiments also bear on (16c), the claim that the
trajector/landmark asymmetry has a temporal dimension. Since
attention was directed to the focused participant prior to presentation
of the critical event or configuration, his results fit well with the
notion that the trajector's primary focal prominence resides at least
partially in its role as initial point of access. Pointing in the same
direction is some recent experimental work indicating that agents are
accessed more quickly than patients:
(18) "...When a prototypical causal relation can be perceived directly and

when the participants in the event maintain their role for an extended
time interval, agent information is accessed more rapidly than patient
information. The prominence of the agent over the patient in

grammatical phenomena is parallelled by a privileged status of the
agent at the level of perceptual processing. It is tempting to speculate
that the observation that the agent is a prototypical subject and that
the subject in canonical sentences surfaces before the object, can be
traced back to the fact that, during visual event perception, agent
information is available earlier for further processing than patient
information" (Verfaillie and Daems 1996:144).
The observation that subjects canonically surface before objects

needs a bit of elaboration. There is indeed a cross-linguistic tendency
for subject to precede object in sentences reasonably considered
unmarked (Greenberg 1963:77), and I take this as lending support to
their supposed temporal asymmetry. This universal tendency is not
however absolute. Moreover, subjects and objects have to be posited
even in non-canonical sentences where the object precedes. The
reason this is unproblematic is that temporal ordering in linguistic
processing is not limited to a single, left-to-right pass strictly
correlated with the phonological dimension of speech time. We can
draw an analogy to eye movements, which do not trace a simple,
continuous beginning-to-end path in the reading of a text. The overall
progression through the text incorporates numerous saccades, both
forward and back, resulting in momentary fixations at different
points and piecemeal scanning through local domains. By the same
token, speaking or understanding undoubtedly encompasses cogni-
tive processing representing myriad levels and dimensions of struc-
ture, each with its own dynamics and to some degree independent of
the linear progression of speech. Certainly it is common for all the
elements within a local domain to be simultaneously available for
conceptual integration whose inherent temporal directionality runs
counter to their sequencing in speech time. Thus we can easily
process either (19a) or (19b) despite the fact that the temporal
sequencing of events correlates with speech time only in the former.
(19) a. They finished the job before I got there.

b. Before I got there, they finished the job.
Observe the greater naturalness of uttering (19a) without a pause, and

(19b) with one. I interpret this as indicating that the correlation of
event sequencing with speech time makes it easier to process the
entire expression in a single "window of attention". With the reverse
order, apprehension is more difficult because the subordinate-clause
event has to be kept in mind until the main clause is processed; only
then can the entire scenario be conceptualized in accordance with its
inherent temporal dynamic. This more complex processing task is
facilitated by placing each clause in its own attentional frame.
Claiming that the trajector/landmark asymmetry has a temporal
dimension therefore does not imply that a subject must invariably
precede an object in terms of word order. It does however predict a
general tendency for the subject to precede, since apprehension of the
subject/object asymmetry can then be made to coincide with the
temporal flow of speech and thus be effected without "backtracking".
Analogous comments hold for reference point relationships in
general. If a reference point affords mental access to a target and a
basis for interpreting it, we should expect a strong tendency for this
inherent directionality to correlate with linear ordering. We do indeed
find that a topic normally precedes a comment, the antecedent pre-
cedes a pronoun, and a possessor precedes the possessed. Yet the
opposite sequence can often occur and may have its own functional
motivation (for "backwards" anaphora, see van Hoek 1997b). Presen-
ted with a pronoun, for example, we are able—within a local domain
—to defer the resolution of its reference until we encounter a po-
tential antecedent, at which point we apprehend their reference point/
target relationship in accordance with its intrinsic directionality.
Further supporting (16c) are various phenomena naturally
describable in terms of the subject providing a point of access for
interpreting other clausal elements. One of these is the asymmetry
inherent in equative sentences, as in (20):
(20) a. My wife is the mayor.

b. The mayor is my wife.
These expressions are not equivalent, despite the fact that they profile
the identity of the same two individuals. Thus (20a) would be more
appropriate if we were talking about my wife, and (20b) if we were
talking about the mayor. The difference emerges more clearly in
sentences describing class membership:
(21) a. A lynx is a feline.

b. * A feline is a lynx.
I analyze these expressions also as profiling the identity of two

individuals. Here, though, we are no longer dealing with actual
instances of the lynx and feline classes, but rather with arbitrary
instances, i.e. instances "conjured up" for a specific local purpose
and having no existence outside the mental space thus created
(Langacker 1991:2.2).5 Given that the lynx class is a proper subset of
the feline class, the well-formedness contrast between (2la) and
(21b) can be seen as a consequence of the subject functioning as
point of access for interpreting the profiled identity relation. In (2la),
we initially focus on a lynx and describe it as coinciding with a
feline; in view of the subset relationship, this will always be valid.
On the other hand, (21b) initially focuses on a feline and describes it
as coinciding with a lynx. Since the subset relationship is a proper
one, the validity of such a description cannot be assumed.
We come now to (16d), the suggestion that trajector and landmark
are properly described as reference point and target, or—since a tar-
get, once focused, can itself function as a reference point—as the first
and second elements in a reference point chain. In a way this follows
as a matter of definition. If trajector and landmark are successive
foci, and the trajector an initial point of access, the landmark must in
some sense be accessible via the trajector, hence a target in its
dominion. Certain other considerations reinforce this conclusion.
It can first be noted that reference point relations do form chains,
the target of one functioning in turn as reference point for the next.
The clearest example is with possessives, as in (22a). The construc-
tion in (22b) invokes a chain of spatial reference points leading to the
target of search.6
(22) a. my mother's cousin's friend's sister's lawyer

b. In the kitchen, on the counter, next to the toaster sat the
missing kitten.
To a limited extent, topics also participate in chains, as we saw in (4).

Beyond this, we can observe that subject and object have a privi-
leged status with respect to grammatical constructions for which a
reference point analysis is strongly motivated. One of these is pro-
nominal anaphora, analyzed by van Hoek (1995, 1997a) as requiring
that a pronoun occur in the dominion of a salient reference point with
which it can be construed as being coreferential. It has long been
recognized that subject and object status are especially important in
this regard. In van Hoek's terms, the subject functions as reference
point for all other clausal elements (including their modifiers), and an
object, for all other clausal elements apart from the subject:
(23) a. My sister lost her keys.

b. *She lost my sister's keys.
c. I observed the baboons in their natural habitat.
d. */ observed them in the baboons' natural habitat.
In short, subject and object act as initial and secondary reference

points for anaphoric purposes. Van Hoek (to appear) has also shown
that they do so for quantifier scope as well.
Another reference point construction where trajector and landmark
enjoy a privileged status is the use of possessives to specify the focal
participants of a nominalized verb, as (24):
(24) the army's destruction [of the city];

the city's destruction [by the army]
The trajector and landmark of a profiled relationship are claimed to

function as reference points for purposes of conceptualizing it. When
a verb is nominalized, the process it profiles undergoes a conceptual
reification, creating an abstract thing to which the profile is shifted;
because it profiles a thing, the derived expression is classed as a noun
(Langacker 1987b, 1991:1.2). If the trajector or the landmark of a

process is a reference point with respect to it, and the process is
reified to form an abstract thing, the reference point relation between
them then holds between two things. That is precisely how a
possessive construction is described: as coding a reference point
relation between two things.
Figure 6(a) shows the trajector of a process as a reference point,
and that process as a whole (represented by a box) as the target
accessed through it.7 The effect of nominalization is shown in Figure
6(b): by virtue of conceptual reification, the target process becomes
an abstract thing, indicated diagrammatically by changing the box to
an ellipse. Hence R and T are both things, just as for possessives (cf.
Figure 4), except that R is necessarily intrinsic to the conception of T
(which is true of many but not all possessives). Thus the proposed
characterization of subject and object neatly accounts for the
widespread use of possessive periphrasis with nominalizations. It
further predicts (correctly I believe) that such periphrasis should be
used primarily for the specification of trajector and landmark, and
only by extension (if at all) for non-focal participants.
Figure 6. Trajector as reference point and the effect of nominalization

5. Levels of organization
Even if accepted as basically correct, the characterization does still

need to be sharpened, clarified, and situated in a broader context. Let
us start by addressing the most fundamental issue. If trajector and
landmark are first and second reference points, with respect to what
experiential domain or dimension of structure do they so function? If
they serve as initial and subsequent points of access, just what is it
that they give access to?
The clearest examples of reference point organization pertain to
particular dimensions of linguistic structure and/or basic aspects of
our everyday conceptual experience. For instance, a chain of locative
reference points, as in (22b), pertains to the basic experience of
finding things in space. Specifically, it reflects the common strategy
of locating something in relation to salient landmarks, and also that
of progressively narrowing the search. Possession pertains to identi-
fication, the problem of directing attention to a particular thing out of
all those that populate our mental world. It exploits the typical cir-
cumstance that different instances of the same thing type are asso-
ciated with distinct individuals whose prior identification enables
them to function as reference points. Pronominal anaphora pertains to
reference tracking, necessitated by the practice of referring to mul-
tiple individuals in a single discourse, and to the same individual on
multiple occasions. Metonymy pertains to symbolization. In a con-
ventional form-meaning pairing, the form is an initial point of access,
hence a reference point. It prompts the listener to evoke a certain
domain of knowledge—the conceptual base, which we can identify
as the reference point's dominion—and focus attention on a particu-
lar element within it, the profile, which is thus a target accessible via
the reference point. In the case of a metonymic interpretation, the
expression's profile (its usual referent) functions in turn as a reference
point affording mental access to its intended referent. The overall
symbolizing relationship thus constitutes a reference point chain.
With respect to what facet of linguistic structure or conceptual
experience can we plausibly say that trajector and landmark function
as reference points? I suggest that they are reference points with
respect to the very process of conceptualizing a relationship. A

relationship is conceptually dependent on its participants, i.e. it
cannot be conceptualized without conceptualizing its participants, if
only schematically. Since a relational conception presupposes and
incorporates the conception of relational participants, the latter
provide mental access to it, which makes them reference points. For
reasons of space, I will mainly limit discussion to trajectors.
By definition, the trajector is the initial focus, or starting point (cf.
MacWhinney 1977; Chafe 1994). If we first conceive of a person, for
example, we can then go on to conceive of that person engaging in
some activity, or participating in some other kind of relationship. The
set of relationships that a trajector potentially "anchors" in this
fashion constitutes its dominion, and the profiled relationship—the
one accessed via the trajector—is the target.8 This was sketched in
Figure 6(a), where the dashed arrows lead to alternate relations
potentially anchored by the reference point (trajector), and a box
encloses the one that is actually conceived and profiled. The concep-
tualizer initially focuses on the trajector and uses that as a basis for
conceptualizing the full relationship incorporating it.
A resemblance can be noted between Figure 6(a) and Figure 5, the
diagram given earlier for topic constructions. The resemblance would
be even stronger if Figure 5 gave some indication of the pivot, i.e. the
reference point's role in the target proposition. The similarity is not
unexpected in view of the close connection between subjects and
topics. Each can be characterized in terms of a reference point
relation between a thing and a process. The problem then is to clarify
the nature of their connection, especially since not every subject is a
topic in the traditional sense.
The key to the matter is to recognize that comparable relationships
can be manifested at different levels of organization. It is commonly
observed that topics can function as such for discourse units of
varying size, e.g. for chapter, section, paragraph, and sentence, or for
analogous units in speech. Likewise, a trajector has to be posited as
part of the inherent semantic structure of expressions that profile
relationships, ranging in size from single morphemes to full
sentences. In Figure 3, for example, trajector/landmark alignment is
shown as part of the internal organization of admires, admires Bill,

and Alice admires Bill.
Roughly speaking, the higher the level of organization, the looser
the connection may be between a nominal reference point and a
process in its dominion. Consider first a "paragraph"-level discourse
topic:
(25) It's really hard to write a dissertation. You have to find a

subject. Then you have to come up with some ideas and do
lots of preliminary analyses. When you do the background
reading, you find that most of those ideas have already been
proposed and rejected. So you have to work for a number of
years before anything viable starts to take shape. You have
to worry about continued financial support. Then you have
to satisfy five committee members with mutually incom-
patible notions about what you should be doing. You have to
go through about seven drafts. Then ...
The first sentence establishes the dissertation as the discourse topic.

The following sentences are all integrated into its dominion, i.e.
interpreted in relation to the dissertation, even though none of them
mentions it explicitly. Structurally, the topic nominal is thus external
to each of the comment sentences. Conceptually, its referent—the
pivot—need not be a focal participant in any of these sentences, and
in some it may be quite peripheral indeed (e.g. You have to worry
about continued financial support).
Consider next a sentence-level topic:
(26) a. Jack, when I go to see him, he's never home, and he's
always complaining that his friends ignore him.
b. Jack, he's always complaining.
The comment sentence may be complex, as in (26a), in which case a

number of propositions are integrated in its dominion. Or it may be
monoclausal, as in (26b), in which case the target consists of just a
single proposition. Either way, the nominal establishing the topic is
still external to the comment from a structural standpoint. Yet their

connection is tighter than it is with a paragraph-level topic, as in (25),
if only because just one sentence functions as comment.9 More
importantly, the pivot tends strongly to be a central participant in the
comment clause; note that in every clause in (26) it is coded by a
subject, object, or possessor pronoun. There are of course occasional
exceptions, e.g. (27), cited by Lambrecht (1996).
(27) La mer, tu vois de l'eau.

The ocean, you see water.'
At the next lower level of organization is the situation where the

nominal that establishes a topic is structurally internal to the clause
expressing the comment proposition. Though different syntactic
arrangements are possible, the one with the closest topic-comment
integration finds the topic nominal functioning simultaneously as the
grammatical subject or object. A well-known example is object
preposing in English:
(28) Bill I admire.
Here clause-initial position marks Bill as the local topic, and at the
same time this nominal directly instantiates the grammatical role of
clausal object. Not only is the pivot a focal participant in the profiled
relationship, but also topic and pivot coincide in the sense that the
same nominal expresses them both. Is this option also available for a
subject nominal? It is certainly possible in principle, but since
English subjects normally occur clause-initially, that alone does not
unambiguously mark a subject as clause-internal topic. Note however
that expressions like (28)—where topic and pivot coincide—often
occur in contrastive pairs, which makes the instantiating nominal
accentually prominent. An example involving objects is (29a), where
accentual prominence is indicated by capital letters:
(29) a. BILL I admire, JACK I don't.

b. ALICE admires Bill, JANET doesn't.
The same kind of contrast is possible with clausal subjects, as in

(29b). Here at least I take the subject as establishing a clause-internal
topic. Each clause in (29b) establishes the subject's profile as a
salient reference point in whose dominion the clausal process is to be
integrated.
Although I have not investigated the matter thoroughly, I suspect
that the contrast between "thetic" and "categorical" judgments
(Kuroda 1972; Lambrecht 1987) resides in whether a focal
participant is additionally established as a clause-level discourse
topic. The clauses in (29b) represent categorical judgments
concerning their respective subjects. Personally, I see no reason not
to similarly regard the clauses in (29a) as representing categorical
judgments concerning their respective objects. The thetic/categorical
contrast is usually discussed in relation to non-contrastive examples,
e.g. the responses in (30), recently described by Lambrecht (1994) in
terms of "sentence focus" vs. "predicate focus":
(30) a. What's wrong? My neck hurts.

[thetic; sentence focus]
b. What's wrong with your neck ? My neck hurts.
[categorical; predicate focus]
Whereas the response in (30a) accords no special discourse status to

the subject, in that the entire sentence represents new information,
the response in (30b) takes the subject as a local topic and makes a
predication concerning it. I would say, then, that my neck in (30a) is
merely the clausal subject, while in (30b) it is both the subject and
the topic for that clause. The difference between the two ways of
expressing categorical judgments is that in (30b) my neck has already
been established as a topic by the preceding question, whereas in (29)
the topic is established as such through means internal to the
comment clause itself (initial position and/or accentual prominence).
In other words, in (29) the topic and pivot roles are instantiated by
the same nominal, within a single clause, while in (30b) this is done
by separate nominals, in successive clauses.
It should be evident that I am distinguishing between a clause-

internal discourse topic, on the one hand, and a subject or object, on
the other. Although they may be conflated in a single nominal
expression, as in (28) and (29), these represent distinct functions
which have to be kept separate for analytical purposes. A discourse
topic can serve as reference point for any number of comment
clauses, and the nominal establishing it is usually external to each of
them. The constructions in (28)-(29) represent the special, limiting
case where the topic holds sway for just one clause, and moreover is
established by a nominal that simultaneously has a central
grammatical role in it. For the most part, however, the nominal
expressing the topic is external to the comment clause. We can see
this as being iconic. Whether a clause is to be integrated in the
dominion of a particular reference point is basically an extrinsic
matter that depends on contingencies of the current discourse.
By contrast, a subject or object specifically functions as such only
for a single clause, as an inherent part of its internal structure. It does
serve as a reference point with respect to the clausal process, but the
reference point role of a subject or object per se—qua subject or
object—is independent of any particular discourse status. Instead, a
subject (or object) specifies the trajector (or landmark) of a profiled
relationship, which functions as reference point with respect to the
very act of conceptualizing that relationship. This is the closest
connection possible between a nominal reference point and a proces-
sual target. Access to a profiled relation via its focal partici-pants is
intrinsic rather than contingent (note the contrast between Figure 6(a)
and Figure 5).
Trajector/landmark alignment is inherent in the meanings of
relational expressions at any level of organization. At the lexical
level, it is a facet of how we categorize the world in terms of
conventionally recognized relation types. While there is of course an
interaction, this categorization function is in principle distinct from
the problem of constructing a coherent discourse, and certainly from
the factors involved in any particular discourse. Agentivity is a
particularly strong factor determining the trajector/landmark align-
ment of English verbs, for example. At the lexemic level, verbs that
profile events with a clearcut agent/patient asymmetry virtually

always choose the agent as trajector (initial focal participant). In
devising an inventory of lexically-coded categories well suited for
describing what happens around us, one natural strategy is to con-
sistently "take the perspective" of the agent in the sense of invoking
it as the initial focal participant to anchor the conception of its inter-
action with a patient.
In most any theory, it is recognized that the trajector/landmark
alignment ("argument structure") inherent in a lexical item is
projected "upward" to higher levels of grammatical organization. The
default alignment imposed by lexical categorization can however be
overridden if there is good reason to do so. That is where discourse
comes in. There is no guarantee that a participant we wish to invoke
as a starting point for discourse purposes will invariably coincide
with the one selected as starting point by inherent lexical coding.
When it does not we often resort to a special grammatical
construction that imposes its own choice of subject, thereby ensuring
that the current discourse topic also functions as point of access for
conceptualizing a clause in its dominion.
In (31), for example, the first sentence establishes my car as a
local topic, and the speaker then finds it pertinent to mention its
participation in certain events.
(31) You really should think about buying my car. It was just
repainted, and it drives very smoothly.
The lexical items repaint and drive are readily available for de-
scribing the types of events in question. Since the repainting and the
driving are to be interpreted with reference to the car, discourse will
flow most smoothly if the pivot (coded by if) has the gram-matical
role of subject, the natural starting point for conceptualizing the
clausal process. The car however functions as patient with respect to
these events, while the verbs select the agent as trajector and initial
point of access. For this reason, the speaker resorts to the passive
construction, in the case of repaint, and the "patient subject con-
struction", in the case of drive (cf. van Oosten 1977, 1986). Both
constructions confer trajector status on the semantic patient. The

pronoun it which specifies the patient is therefore both the subject,
affording access to the clausal process, and the pivot serving to
connect each clause with the topic.
Thus a clause, especially in its choice of subject, is a primary
locus for the interaction between discourse considerations and the
trajector/landmark alignment imposed by lexical categorization. As
the expression of a single proposition, a clause is the smallest unit of
discourse over which a topic holds dominion as the active reference
point. At the same time, a clause is the largest unit for which we can
posit a single overall trajector, either inherited by default from the
lexical verb or imposed by a special construction like the passive.
Whether the lexical default is sustained at the clause level depends on
discourse factors like topic/comment organization, which arise at that
level and above. In particular, discourse tends to flow most smoothly
when the participant chosen as subject is also the pivot corresponding
to a salient topic. It does so because a pivot is the point of connection
between a topic and a comment clause in its dominion, while a
clause's subject, by virtue of coding the trajector, is the point of
access for conceptualizing the process it profiles.
I have traced the progression from a paragraph-level discourse
topic, at a high level of organization, down to the lexically
determined trajector of a verb, at the lowest level. Viewed abstractly,
all these levels are susceptible to uniform characterization: they all
involve a reference point relation between a thing (the topic or
trajector) and a process in its dominion. This abstract commonality is
held responsible for the affinities noted earlier between topics and
subjects. As we move from the highest to the lowest level, the
connection between the nominal reference point and the processual
target becomes progressively closer. The span of discourse over
which the reference point holds dominion becomes progressively
smaller: from a series of sentences, to a single sentence, to a clause,
to a phrase, to just a verb. The reference point becomes progressively
more intrinsic to the target process, both structurally and concep-
tually. Naturally, as we move from larger, loosely integrated
structures to those which are smaller and more compact (at the
extreme, a single verb stem), their processing occurs on a pro-

gressively smaller time scale, and we are less and less likely to be
consciously aware of either the reference point organization or its
dynamic character. I would not expect the temporal dimension of
trajector/landmark alignment to be easily accessible to introspection
at the lowest level.
Despite this commonality and continuity, an important
discontinuity between topic and trajector has also been observed. It
stems from the fact that the reference point relations at the two
extremes have different functional motivations. A high-level topic
helps to organize a discourse and render it coherent by evoking a
single dominion for interpreting an open-ended series of target
propositions. As full-fledged discourse entities, these propositions
(generally expressed by finite clauses) make numerous specifications
concerning the process they designate: they specify its participants
(with subject and object nominals), its location in time and reality
(via tense and modals), its precise nature (through modifiers), etc.
Furthermore, the topic is often quite extrinsic to the target clause,
which could just as well be assigned to some other topic's dominion.
At the opposite extreme, the trajector/landmark alignment inherent in
lexical verbs contributes to their function of supplying an inventory
of process types with which to categorize events and situations. At
the level of lexical categorization, the profiled relationship is not a
full-fledged proposition, nor even a specific process instance, but
merely a process type. Moreover, the trajector is wholly intrinsic to
the target process, serving as a reference point with respect to the
very act of conceptualizing it.
Starting from the extremes, these two functions converge at the
level of the clause—the smallest unit expressing a full proposition,
and the largest one with a single overall trajector. A nominal that
specifies a clause's trajector is its subject, as a matter of definition. It
may or may not be a discourse topic as well. In a "thetic" judgment,
like the response in (30a), it has no such status. If it is a topic,
resulting in a "categorical" judgment, there are two possibilities. On
the one hand, the topic may already be established in the discourse.
In this case the subject nominal merely expresses the pivot, e.g. my
neck in (30b). Alternatively, the clause itself may serve to establish

the topic. In this case the subject nominal simultaneously expresses
both topic and pivot, as in (29b).
6. Conclusion
Though some will regard these proposals as speculative, I see them

as both plausible and motivated by a variety of considerations. The
notion that subject and object should be ranked as the first and
second elements on a scale of prominence or accessibility is hardly
unprecedented. Shibatani (1985) has made cogent observations
pointing in this direction. Keenan and Comrie (1977) list subject and
object at the top of their "accessibility hierarchy" describing the
possibilities observed cross-linguistically for relative clause
formation. Relational grammarians found good reason to label them
"1" and "2" (Perlmutter 1983), and the same ranking is posited in
other formal theories. Here we have concentrated on motivations for
describing the parameter in question as some kind of prominence
hierarchy, and specifically for characterizing subject and object as the
first and second foci on a reference point chain inherent in the
conception of a profiled relationship. Corroborating this account is
the fact that subject and object clearly function as first and second
reference points for purposes of pronominal anaphora.
Furthermore, the characterization of subject and object fits nicely
into a much broader view of conceptual and linguistic structure that
has been emerging of late in cognitive linguistics from the efforts of
numerous scholars working on diverse problems. There is a growing
realization that inherently dynamic aspects of conceptual structure—
involving notions like "access", "mental scanning", "fictive motion",
"viewpoint shift", etc.—are both pervasive and fundamental to
semantic and grammatical organization.10
More concretely, I have presented a substantial body of evidence
suggesting that trajector and landmark are properly described as
primary and secondary focal participants, and that the asymmetry in
their degree of prominence has a temporal dimension. It is possible,
moreover, to see their focal prominence and reference point status as

verb- and clause-level analogs of reference point phenomena
observable at higher levels of organization, where temporality and
sequenced access are more apparent. If we merely evoke the meaning
of a transitive verb, it may not be introspectively obvious that we
focus on its two participants sequentially in building up to the full
conception of the profiled relationship.11 However, focusing and
sequential access very similar to what is posited for verbs are quite
evident at the level of clause-external topics, especially with multiple
topics (cf. Lambrecht 1996):
(32) Henri, ma cousine, il ne I'aime pas.

'Henry, my cousin, he doesn't like her.'
Here there can hardly be any doubt that nominal referents are
invoked as successive foci subsequently used to anchor the
conception of a relationship in which they both participate. The same
can be observed at the discourse level when topical participants are
separately introduced and then described as interacting:
(33) Jack came home late in the afternoon. There was Jill,
working at her computer. He walked over and hugged her.
In short, the conceptual operations I am positing are quite

unremarkable even if, at lower levels of organization, they are too
compressed to be readily discernible by introspection.
A great deal of work is clearly needed before this proposal is
widely accepted as a viable and well-supported characterization of
subject and object. Required in particular are detailed analyses of
individual languages of diverse types (e.g. Cook 1988; Kumashiro in
preparation), as well as further experimental work approaching
matters from the psychological and processing standpoint. Perhaps
the proposed description will be enough to give investigators some
idea of what to look for. At the very least, I suggest it offers a
coherent, integrated view of topic-like elements at the lexical,
grammatical, and discourse levels, in both synchronic and diachronic
perspective. I further believe that it explains and clarifies the special

relation among topic, subject, and possessor, accounting for both
their affinities and their non-equivalence.
Notes
1. A preliminary version of this paper appeared in Linguistic Notes from La Jolla

19, 1-28, 1998.1 would like to thank the anonymous reviewers of this paper for
helpful comments.
2. In the discussion I will often concentrate on subjects: I believe the
characterization can however be succesfully extended to objects.
3. Note that I generally use circles to represent things, and various kinds of lines
and arrows to indicate relationships.
4. Observe that correspondences are represented with dotted lines.
5. Similarly, in the non-specific reading of Phil wants to pet a lynx, the referent of
a lynx is an arbitrary instance conjured up to characterize Phil's desire and has
no status outside the mental space representing it.
6. I analyze this locational reference point organization as being superimposed on
the one defining trajector/landmark alignment. Each reflects its own natural
path of access.
7. The box is included just for diagrammatic convenience - it is not distinct from
the trajector and the process it anchors (given as a solid arrow), but is rather a
notational device meant to represent their status as a single entity.
8. Recall that the topic my car in (15) is a reference point with respect to both the
motor and the entire comment clause the motor is just about shot. In similar
fashion, a trajector functions as reference point with respect to both a landmark
(if there is one) and a profiled relationship. A landmark can be characterized as
an initial target and a secondary reference point in a reference point chain
required for a relationship's full conceptualization.
9. This does not preclude a sentence-level topic also being adopted as a higher-
level topic that remains active through a longer stretch of discourse.
10. See, for example, Elman (1990); Cutrer (1994); Fauconnier and Sweetser
(1996); Matsumoto (1996a, 1996b, 1996c); Sweetser and Fauconnier (1996);
Talmy (1996); Sweetser (1997); Langacker (1997b, 1999b); van Hoek (1997a,
to appear).
11.1 attribute this to the small time scale inolved, the automatization of lexical
meanings, and the general non-salience of conceptual components that are
sublexical hence not individually symbolized (cf. Postal 1969).
Language as a cognitive network
Richard Hudson
The paper develops the leading idea of Word Grammar and other
"cognitive" theories of language, which is that language is a
network. It reviews some of the consequences of this view: spreading
activation, effects of conceptual distance, default inheritance, the
unity of grammar and lexicon and, more generally, non-modularity;
the unity of permanent and temporary representations, degrees of
accessibility and binary relations. It then shows briefly how these
ideas apply to two specific areas of language analysis: the contrast
between polysemy and homonymy, and the treatment of regular and
irregular morphology. The last section discusses Pinker's contrast
between "mentalese" and "connectoplasm", and argues that the
networks defined in this paper have all the symbolic qualities of
mentalese, so maybe the mind uses "networks all the way down".1
1. Networks in Cognitive Linguistics and Word Grammar
The main focus of this paper is one aspect of the approach to

language which has come to be called "Cognitive Linguistics"
(Lakoff 1987; Langacker 1987a, 1990; Taylor 1989), namely the
widespread assumption that the whole of language is a network, a
collection of units linked pair-wise by relations. These units are
concepts—concepts such as "word", "the word word' or "the
phoneme /p/"—and the leading claim of Cognitive Linguistics is that
the concepts of language are organised in the same way as all other
concepts. In short, "knowledge of language is knowledge" (Goldberg
1995: 5). General knowledge is widely believed to be structured as a
network, so this view allows language to be integrated with other
kinds of knowledge.
50 Richard Hudson
The network view contrasts most clearly with the view that
knowledge is divided into distinct modules, each with its own formal
properties and internal structure and with very limited links to other
modules (Fodor 1983; Chomsky 1986; Pinker 1994, 1997). As far as
language is concerned, the lexicon and the grammar form distinct
modules, as do phonology, semantics and so on; and in an extreme
version, each lexical item is held in a distinct "lexical entry" which is
a kind of mini-module.
Cognitive Linguistics exists as a rather general ideology about
language, but it also includes three well-developed theories of
language structure: Cognitive Grammar (Langacker 1987a, 1991,
1994), Construction Grammar (Goldberg 1995) and my own theory,
Word Grammar2, as well as work by a wide range of other linguists
among whom the most directly relevant is Bybee (1985, 1995a).
Networks have played a central role in Word Grammar since its
earliest days, as shown by the following quotation: "At the most
general level, [Word Grammar theory] consists of [the following]
generalisation: A language is a network of entities related by
propositions." (Hudson 1984: 1)
However, its historical roots are rather different from those of the
other theories. It was influenced heavily by Chomsky's work, and
especially by his stress on generativity, but it does not, in any way,
grow out of Transformational Grammar. (In personal terms, I have
never been a Chomskyan—witness Hudson 1976.) The main theories
of grammar that contributed to it are also outside the Chomskyan
mainstream: Systemic Functional Grammar (Halliday 1985; Martin
1992; Hudson 1971), Stratificational Grammar (Lamb 1966; Bennett
1992) and Dependency Grammar (Tesniere 1959/66). As will
become apparent below, some of the leading ideas come directly
from Cognitive Science.
The undoubted similarities between Word Grammar and the other
"cognitive" theories (Cognitive Grammar and Construction
Grammar) are due to independent evolution. This in itself is
encouraging for those of us who would like to believe that our
theories are driven by the facts rather than by our personal histories.
We shall see some differences between Word Grammar and the other
theories, but the most obvious differences are matters of notation and
the substantive differences are relatively unimportant compared with
the shared stress on networks. We shall now consider some of the
consequences of building a theory of language in terms of networks.
2. Networks
2.1. Networks and activation
For a cognitive linguist, one of the attractions of applying networks

to language structure is that cognitive psychologists often model
other parts of cognition as a network. One of the most widely
accepted ideas in this field is "that memory connections provide the
basic building blocks through which our knowledge is represented in
memory [...] all of knowledge is represented via a sprawling network
of these connections, a vast set of associations." (Reisberg 1997:
257-258). If language is like other parts of cognition, then we might
expect at least some parts of language to constitute a network; and if
some parts are a network, the most parsimonious theory is that they
all are. This is the basis for the Word Grammar claim quoted above:
"a language is a network of entities related by propositions."
Cognitive Grammar and Construction Grammar also present
language as a network of units (Langacker 1990: 2, 266; Goldberg
1995: 5).
What does this claim mean, and how is it distinctive? The network
view has several interesting and important consequences, but I shall
highlight just a selection. First, knowledge must be declarative,
rather than procedures which have an "input" and an "output". We
know relationships, not "rules" for "doing" things. For example, we
know that a past-tense verb has a suffix after its stem, and that the
stem and suffix together comprise the whole of the verb; these are
relationships among the concepts "past-tense verb", "its suffix", "its
stem" and "its whole" which are diagrammed (using Word Grammar
notation) in Figure 1 (see the next page). We do not know "how to
form a past-tense verb"—such knowledge cannot be accommodated,
52 Richard Hudson
as such, in a network. A network is like a map, which is simply

static; it is not like a set of instructions for getting from A to B.
On the other hand, because of its declarative nature, the network
may be used like a map as a guide to activity. When our goal is to
find a pronunciation we can move from "past-tense verb" to the
pattern "stem+suffix" by following these links in the network, and
when our goal is to find a meaning we can move in the opposite
direction. This is possible in a network because links may be
activated like the links on a circuit board. A widely accepted model
in cognitive psychology is spreading activation, whereby activation
of one node spreads (as weaker activation) to the nodes to which it is
directly attached. Spreading activation can only be assumed in a
network; it makes no sense if knowledge is envisaged as a collection
of rules or separate lexical items.
whole
verbpast
suffix
Figure 1
It is perhaps in this sense that Langacker (1990: 15) states that

"linguistic knowledge is procedural rather than declarative".
Activation is central to Cognitive Grammar, and the "structured
inventory of conventional linguistic units" (1990: 15) which are
activated is closely analogous to the network of concepts in Word
Grammar.
2.2. Conceptual distance and the networked lexicon
The conceptual distance between two concepts can be measured as

the number of intervening links—a notion which would be
meaningless if the same information was presented as a list of rules
or propositions. We shall explore this consequence below, but

meanwhile it should be noticed that distance is relevant because of
spreading activation—the further apart two concepts are, the less
likely they are to activate one another. The classic experiments on
semantic priming (Reisberg 1997: 265-267; Swinney 1979) show
that activation can spread across several links, so it does not require
strict adjacency. For example, when presented with the word ink,
subjects take less time to decide that it is an existing English word
when it has been primed by an immediately preceding word than
when it has not; and it is primed by pen, but not by king. Now the
point to notice is that the link between the words ink and pen lies in
their meaning: (the word) ink - (the thing) "ink"—"pen" - pen. This
involves two intervening concepts, and perhaps more (depending on
how direct the link between "ink" and "pen" is). Of course the
conclusion is not that ink is related to pen but not at all related to
king; after all, in a network everything is ultimately related to
everything else. The difference lies in the relative distance between
the concepts, which is shown in Figure 2.
pen MEANING
WORDS
pen ink king
Figure 2
The same kind of spreading activation which leads to semantic

priming also produces well-known effects in the recognition of
words. These effects are the basis of the cohort model of speech
recognition (Marslen-Wilson 1980; Marslen-Wilson and Tyler 1980,
1981) according to which the incoming sound-segments activate a
decreasing cohort of candidate words: Id activates all the words that
begin with /e/, then /!/ reduces the cohort to those that have initial
/el/, and so on, until only one candidate remains—which is why a
word like elephant can be recognised long before the end.
54 Richard Hudson
This evidence suggests, then, that when we hear a word it is

embedded in a network of phonological and semantic links to other
words. If this is so in speech perception, we should expect the same
to be true of production—as it seems indeed to be, to judge by the
evidence of speech errors. When a target word is replaced by
another, the two are usually similar either in phonology or in
meaning (Fromkin 1988), suggesting that these characteristics are
"networked" (Emmory and Fromkin 1988: 145).
It seems almost beyond dispute that the links from words to
meanings and sounds involve a network of links between individual
words (or lexemes) and individual semantic and phonological
patterns—a far cry from the earlier view of the lexicon as a list of
discrete lexical entries. It is less important to decide whether these
links are just relationships or "rules" (e.g. the "correspondence rules"
of Jackendoff 1997); indeed it is hard to know what the basis of such
a distinction might be.
2.3. Default inheritance
Large parts of the network are organised hierarchically on the basis

of the most important relationship of all, the isa relationship (which
is widely recognised in Artificial Intelligence—see Reisberg 1997:
280), Langacker's "categorization" (1990: 266), Goldberg's "instance
link" (1995: 79). This is the relationship that underlies all
classification, so saying that "A isa B" is the same as saying that A is
an example or subclass of B: "Fido isa dog", and "dog isa mammal";
or "dog isa noun" and "noun isa word". The importance of this
relationship lies in the fact that it allows "generalisation by
inheritance". If "A isa person", and a typical person has a heart, then
A must have a heart as well; so this fact can be stored just once, in
connection with the general category "person", and inherited each
time it is needed. A classic experiment by Collins and Quillian
(1969) showed that it takes longer to verify a fact which is inherited
than one which is not. For example, subjects take longer to verify
that a canary can fly than that it can sing, because the ability to fly is
inherited from "bird" but canaries are kept partly for their singing so
the ability to sing is a salient fact which is stored in relation to
canaries; and it takes even longer to verify that a canary has skin,
because this is inherited via "bird" from "animal". In short, the time
taken for verification depends (among other things) on the
conceptual distance between the inheriting node and the source of the
information (Reisberg 1997: 267-270), as measured simply in terms
of the number of intervening links that have to be traversed.
Inheritance is clearly one of the effects of activation which is
controlled by conceptual distance. One particular consequence is that
properties are only inherited by default, i.e. in the absence of a more
specific specification to the contrary. In short, the process is more
accurately described as default inheritance—"normal mode
inheritance" (Goldberg 1995: 73), "schematicity" with either
"elaboration" or "extension" (Langacker 1990: 266-267).
The overriding of defaults follows from the network model
because the inherited property is always the one which involves the
shorter route, i.e. the least conceptual distance. Take the canary
example just quoted. The ability to sing is directly related (in the
network) to "canary", giving a distance of just one link, whereas the
ability to fly is separated from "canary" by two links: from "canary"
to "bird", and from "bird" to "can fly". As we saw, retrieval is faster
over shorter distances. Suppose now that the bird species in question
was "ostrich". Unlike the typical bird, the typical ostrich cannot fly,
so this fact must be stored directly in relation to "ostrich". If we now
assume for simplicity that the link is labelled "flies?", with "yes" and
"no" as possible values, it is clear that the value "no" will be
retrieved before the value "yes" simply because of the difference in
conceptual distance.
56 Richard Hudson
flies? whole
bird
verb
yes
flies? whole
ostrich no 'past
Figure 3
The relevant network is shown in Figure 3, alongside a similar

network for verb morphology which shows why irregular
morphology takes priority over regular. In this case, the relationship
whose default value is overridden is "whole", which gives the word's
fully-inflected form. The small triangle in the diagram signals the isa
relationship iconically, since its (large) base rests on the
supercategory and its (small) apex points towards the subcategories;
so "ostrich" isa "bird", and GO isa
2A. The lexicogrammar continuum and "temporary competence" in

performance
If all generalisations are made (as in cognitive grammars) by default

inheritance down an isa hierarchy, the facts which other theories
express by means of general "rules of grammar" will be expressed as
facts about relatively general categories near the top of the hierarchy.
As in other cognitive theories, this leaves no place for a boundary
between the "lexicon" and the rest of the grammar (Langacker 1990:
102; Goldberg 1995: 7); so if the "lexicon" is a network, the same
must be true of "the rules of grammar". Moreover, Word Grammar
syntax uses word-word dependencies rather than phrase structure, so
the grammar generates just words and their dependencies. This
means that all higher-level generalisations apply to single-word
categories—to the category "word" itself, or word-classes such as

"noun" or "auxiliary verb".
The result is that all our linguistic competence is organised around
an isa hierarchy of word-types, with "word" at the top and individual
lexemes and word-forms at the bottom—the "lexicogrammar"
(Halliday 1985). Indeed the hierarchy is continued even further down
during performance by the addition of two extra kinds of word. First,
there are word-types which are inherited rather than stored. For
example, it is generally agreed that at least some regularly inflected
words are not stored; so a newly invented word obviously cannot
have been stored by the hearer, and yet we can all understand
inflected forms like wugs. Presumably the same is true at least for
inflected forms of regular and rare words like disambiguate. In order
to understand the form disambiguated we must assign it a new
temporary concept which isa the lexeme DISAMBIGUATE and "past-
tense verb" but which is also distinct from the word token which we
are understanding (whose properties need not be exactly the same—it
could be mis-spelt, for example). In Word Grammar notation, this
word-type is called DISAMBIGUATED.
The second kind of temporary addition to the network is the actual
word token as recognised in processing. Suppose I am reading the
following sentence:
(1) Pat shuffled uphill and Jo shufled [sic] downhill.
This contains seven word tokens, but two are tokens of the same
type: SHUFFLED (a concept which is unlikely to be stored). The
Word Grammar name for the tokens is based solely on position, so
they are called words 2 and 6, "w2" and "w6", giving three
temporary concepts: SHUFFLED and w2 and w6, each of which "isa"
the former. The isa relationship allows mismatches, so this analysis
is compatible with the mis-spelling shown. The hierarchy needed for
these two tokens of shuffled is shown in Figure 4.
In this view of language, then, there is no boundary between
grammar and lexicon, but neither is there any formal distinction
between competence and the structures created in performance.
58 Richard Hudson
Performance is the production of mental representations which "isa"

those of competence. Moreover we can extend this theory into the
area of learning by assuming that we build our competence by
permanently storing some of the representations that we produce
during performance (depending on factors such as frequency,
salience and so on). In short, Word Grammar is a "usage-based"
theory, just like the other two theories (Langacker 1990: 264;
Goldberg 1995: 133).
word
T
verb
v
SHUFFLE past
K
PERMANENT
TEMPORARY
W1 W2 W3 W4 W5 W6 W7
Pat shuffled downhill and Jo shufled uphill.
Figure 4
2.5. Non-modularity
More generally, there can be no modularity, in the sense of Chomsky

(1986) or Fodor (1983). A network has no natural boundaries, which
means that the network must have the same fundamental formal
properties throughout. For instance, it could not allow spreading
activation, or default inheritance, in some areas but not in others; nor
could it accommodate a phonology which was procedural and a
syntax which was declarative.
This is not to say that the network is completely uniform

throughout. On the contrary, any cognitive network will reflect the
boundaries of "nature", so that similar experiences will be more
closely related to one another than to more different experiences. The
result will be a pattern of subnetworks with more or less vague
boundaries—one network for our family members, another for
colleagues, a much larger one for people (which includes both these
subnetworks), another for telephone numbers, and so on. These
subnetworks are by no means isolated from one another—for
example, many telephone numbers belong to some person, so
numerous links go from individual numbers in the telephone-number
subnetwork to individual people in the people subnetwork.
The same organisation must surely be true of language, especially
given the importance of isa hierarchies for bringing together units
that are similar. For example, the relationships of syntax comprise a
hierarchy of dependency-types (e.g. "subject", "complement",
"adjunct") while those of morphology belong to a different hierarchy
which includes "stem", "suffix" and "whole"; there is no direct link
between these two hierarchies, and they probably each define a
different subnetwork of the total grammar. Words belong to
numerous overlapping subnetworks, but these need not correspond to
the traditional divisions into linguistic modules. For example, brain
damage can impair very specific areas of vocabulary such as all
words which name things found inside the house (Pinker 1994: 314);
no linguistic theory recognises a module for such words or their
meanings, but such cases are to be expected if knowledge is indeed
clustered into subnetworks as suggested above. Given our ability to
see connections in the world, knowledge must be compartmentalised
to the extent that our experience of the world itself is.
2.6. Accessibility
Another well-established finding of psychology (and of ordinary

experience) is that some concepts are more accessible than others.
This has been particularly well documented for word retrieval; for
60 Richard Hudson
example, accessibility depends on frequency and recency (Yelland

1994: 5012).
These notions can surely be accommodated in some way in a
network model, but they again make very little sense in terms of
rules and lexical entries. However, in a network-based theory we
have to answer an important question: is it the nodes themselves, or
their links, that vary in accessibility? In Cognitive Grammar, it is
nodes (e.g. structural Schemas) that are described as having various
degrees of "entrenchment" (Langacker 1987a: 59), but my own view
is that at least some variation applies to the links between nodes, with
some links being more easily activated than others. For example,
when we hear a familiar piece of music it is often hard to recall its
composer, regardless of how "entrenched" that composer may be;
what is inaccessible is not the composer, as a person, but the link
from the tune to that person. This is easy to explain in terms of a
network in which each connection between two nodes has a "weight"
which decides how much activation it passes from one node to the
other (Luger and Stubblefield 1993: 517). If this is the explanation
for all kinds of quantitative variation, then entrenchment will have
been replaced by weighted links, but of course it may turn out that in
some sense both nodes and links need to be weighted. This is clearly
an important question for research.
There is an interesting connection between these ideas and the
findings of quantitative dialectology. This work (largely led by
William Labov, cf. Labov 1972a, b, 1994) has uncovered a mass of
quantitative links between speakers' social characteristics (sex, age,
social class and so on) and their use of selected linguistic variables
(e.g. the pronunciation of a particular phoneme or word). These
findings have so far not been successfully integrated into any
mainstream theory of language structure, but the network approach
promises to be more helpful. Suppose a speaker S alternates between
overt III and no /r/ in words \\kefourth ana floor, but regularly uses
/r/ in about 35% of all tokens where /r/ is possible—a very typical
finding of Labovian research (Hudson 1996, chapter 5). This
behaviour must be based on S's knowledge of the local social
structures, so let's assume that S believes /r/ is typically used by
members of social group G. We now have a simple subnetwork in S's

mind which links the following concepts: speaker S, group G and
pattern /r/. We can now explain S's behaviour as the result of the
weightings of these links—for example, a weighting of 35% on the
S-G link and 100% on the /r/-G link. It remains to be seen whether it
is possible to build a single network model of this kind which
integrates the figures for a larger number of linguistic variables
(Hudson 1996: 248-252).
2.7. Relationships
A concept's characteristics—its "definition"—all involve its

relationships to other concepts, so there are no "one-place
predicates". This follows from the claim that a network expresses all
of knowledge, and not just those bits of knowledge that involve
traditional two-place predicates. It is possible to imagine a cognitive
system in which two-place predicates are expressed by network links
while one-place predicates are attached directly to nodes as some
kind of label; but this is not compatible with a pure-network theory
such as the present one. This conclusion applies of course throughout
the grammar, but its consequences are probably most controversial in
semantics.
In traditional logic, a verb such as yawn is a one-place predicate,
not a relationship, so its argument structure is "yawn (x)" and the
logical structure of a sentence such as John yawned, is "yawn
(John)" where a fundamental distinction is made between
"predicates" (yawn) and their arguments (John). This distinction is in
any case undermined by various facts such as the possibility of
nominalisations (John's yawning, or even John's yawn) which turn
the putative predicate into an argument; but in any case it is
incompatible with network models since "yawn" and "John" are both
concepts which need to be represented in the same way, as nodes in
the network. (To emphasise their similarity we shall use the
nominalisation "yawning" as the name for the node which
corresponds to yawn.) In this semantic analysis of John yawned, the
62 Richard Hudson
verb refers to a particular "act of yawning", which we may name

simply "y", and John's role in the yawning is shown by a link from
the "John" node to the "y" node. Similarly, traditional logic allows a
general category such as "linguist" to act as a predicate with a single
argument, so John is a linguist expresses a proposition based on a
one-place predicate: "linguist (John)". In a network, however, it
involves two distinct concepts linked by the "isa" relationship: "isa
(John, linguist)".
Adjectives and adverbs raise rather special problems because
there is no well-developed Word Grammar analysis of them, but in
principle the same remarks apply to them. Thus an adjective like
bald cannot be treated as a one-place predicate but defines a
relationship. In this case the argument is supplied by the noun to
which the subject is applied—e.g. by John in John is bald—and the
value is some degree of baldness, namely a degree which is above
average which we can call simply "bald". In traditional notions, then,
the meaning of John is bald is "baldness (John, bald)", and that of
bald linguist is "x, isa (x, linguist) & baldness (x, bald)". Similarly
for adverbs like loudly, whose argument is the verb to which they are
applied as adjuncts. Take the following example:
(2) The bald linguist yawned loudly.
By assumption the semantic structure of this sentence involves the

pairs of concepts and relationships shown in Table 1.
Table 1
argument value relationship
x (= the yawner) linguist isa
y (= the yawning) yawning isa
y yawn-er x
x baldness bald
y loudness loud
These relationships are shown in Word Grammar notation in Figure

5, which is a partial structure for the sentence. It does show part of
the syntactic structure (below the dotted line), but it gives no
morphology and the semantics is incomplete. The main point is to

show how apparent one-place predicates like "linguist", "yawning",
"bald" and "loudly" can be shown as relationships between distinct
concepts. (For more details see Hudson 1990, chapter 7, 1995.)
yawn-er
t.b.l. yawning loudly

.baldness
. . i x I loudness
bald linguist\ yawning loudly

A 1
linguist / yawning
loudly.
2.8. Conclusion
We have reached two general conclusions about the network view of

language. First, contrary to what I myself once thought (Hudson
1990: 15), it is not just a matter of notation, an alternative way of
presenting information which could have been presented as a list of
propositions. It is true that a network can be "unpacked" as a set of
propositions, but the network model places specific constraints on
propositions such as the fact that they have to be binary.
Furthermore, networks have properties that a set of propositions does
not have, including the possibility of measuring conceptual distances
in terms of intervening links. Above all, only a network can serve as
the database for a spreading-activation processor.
The second general conclusion is that there is considerable
evidence to support the network model—evidence from experi-
mentation (e.g. priming), from speech errors, from neuropathology
and even from quantitative sociolinguistics. All these sources of
64 Richard Hudson
evidence converge on a view of language which is very different

from the more familiar "box" model, in which cognition contains a
box of language which in turn contains distinct boxes of syntax,
morphology, lexicon and so on. Instead, we can see language as part
of the "sprawling network" of cognition, even though it may be
separated from other parts by regions of relatively low-density
connections; within this subnetwork we may be able to find further
sub-subnetworks and so on, but every part of the total network is
organised along the same lines (around isa hierarchies) and shares
nodes with other subnetworks.
In the rest of the paper I shall spell out some consequences of this
view for specific areas of language analysis.
3. Polysemy and homonymy
What light does the network view throw on the traditional distinction
between polysemy and homonymy? This distinction is recognised as
a major problem in lexical semantics, and indeed as a practical
problem for lexicographers, because it requires a binary distinction
where everyone agrees there is in fact a continuum. The distinction
has to be made wherever we find a range of alternative meanings
which can be expressed by a word of the same form, because we
have to decide whether or not the meanings all belong to the same
lexeme. According to the traditional view the two extremes of the
continuum are occupied by polysemy (one lexeme, many meanings)
and by homonymy (two lexemes, one form). A standard example of
polysemy is BOOK, whose sense may be either concrete (volume) or
abstract (book-contents); and a standard example of homonymy is
BANK^^ and BANKriver, whose senses are unconnected. It is clear in
these cases whether the senses are closely related (polysemy) or
extremely distant (homonymy), but in between such examples lies a
continuum of less clear cases: DRY (applied to wine or soil), CYCLE
(bicycle or abstract repetition), THROW (stone or party), and so on.
The assumption that lies behind all these discussions is that one of
the functions of a lexeme is to show semantic relatedness. By
assigning two senses to the same lexeme, the analyst is guaranteeing

that these senses are closely related. But as we saw earlier a network
analysis of meaning already shows the degree of semantic
relatedness in terms of "cognitive distance", which is a continuum.
Presumably there is just one link between "volume" and "book-
contents", contrasting with perhaps ten links in the chain from
"money-bank" to "river-bank", and some intermediate number for
the chain from the dryness of wine to the opposite of wet. When the
network already shows semantic relatedness so sensitively, it seems
perverse to use the very blunt instrument of lexeme contrasts to
impose a binary cut in the same information.
Once freed from this obligation to show semantic relatedness, the
lexeme can be used for the job that it is uniquely suited to perform:
that of expressing the coincidence of syntactic, morphological and
phonological characteristics. More strongly, lexemes must never be
redundant—they must always "do some work" in the grammar. For
example, suppose we recognise two BANK lexemes; they would be
grammatically identical—same morphology (plural: banks), same
word class, same lack of syntactic valency. Therefore one of them
must be redundant because its meaning could be assigned to the other
without losing any information at all in the grammar; so they should
be collapsed into a single lexeme. In contrast, the verb BANK (as in
He banked the money) must be a distinct lexeme in spite of its close
semantic links to the "money-bank" sense of the noun.
The result of this analysis is shown in Figure 6, where the division
into lexemes follows the grammar and not the semantics. The crucial
point about this diagram is that the meaning "money-bank" is much
nearer, in terms of links, to the meaning "banking" than it is to the
meaning "river-bank". The chain which links "money-bank" to
"river-bank" is just a place-holder for a properly justified analysis,
but it is hard to imagine a shorter chain between these two concepts;
whereas the one from "banking" to "money-bank" is both short and
well founded (banking something means putting it into a money-
bank). This analysis shows exactly the same relationships as a
different analysis in which the two traditional BANK noun lexemes are
distinguished, so this distinction must be redundant, and wrong.
66 Richard Hudson
entity
place 'thing
T
institution putting
terrain-feature shop
T
river-bank
τ/
money-bank
place result
banking
verb
BANK. BANF
N
Figure 6
The conclusion, then, is that a network is a very good way of

showing degrees of relatedness between words, whether this
relatedness is a matter of sound, of syntax or of meaning. There is no
point in trying to force a binary distinction between "similar" and
"different" meanings in terms of one lexeme versus two.
4. Regular and irregular morphology
Another area where the network view is relevant is in the treatment

of inflectional morphology (Creider and Hudson 1999; Hudson
2000). The widely held "Dual mechanism" model makes a
fundamental distinction between regular and irregular morphology
on the grounds that regular morphology is handled by rule, while
irregular morphology is expressed as default inheritance in an
inheritance hierarchy (Marcus et al. 1992; Pinker 1999). Such a view
is incompatible with network analysis because there is no formal
distinction between rules and default inheritance: "rules" are
expressed as defaults which are inherited unless blocked. Depending
on how we describe the information in a grammar, either everything

is rules, or everything is network links; but there is no division into
rules and links. The network view is therefore supported by the
research which supports a model in which a single mechanism
handles all morphology (Plunkett and Marchman 1993; Matcovich
1998). It remains to be seen exactly what kind of network model is
needed—a connectionist model, or a more "symbolic" one such as I
am proposing.
As we saw in section 2.1-2.3, Word Grammar morphology is
based on two functions, "stem" and "whole". A word's stem is the
form that it inherits from its lexeme, and its whole is its fully
inflected form; and inflectional morphology handles all differences
between the two, while derivational morphology deals with relations
between the stems of different lexemes. (A language in which every
word's form is that of its lexeme has no inflectional morphology, and
needs no distinction between "stem" and "whole".) A Word
Grammar morphology consists of a set of "equations" each of which
defines a different function:
(3) a. The whole of a verbpast is its stem + its suffix,

b. The suffix of a vert)past. is ed.
Equation (a) is the one diagrammed in fig. 1, where "verb^," is the

name of a subclass of verb. If we assume that the past tense of WALK
isa both WALK and verbpast, it follows by default inheritance that its
whole must be walked. Diagram notation is particularly helpful when
the right-hand side of the equation contains a functional definition
(e.g. "its stem").
For present purposes, the most important fact about Word
Grammar morphology is that exceptions, general defaults and sub-
generalisations all have the same formal properties. We can illustrate
this point through a tentative analysis of a fragment of the Norwegian
verb system, focussing on past-tense morphology (Simonsen this
volume). Most forms in a regular verb paradigm have a suffix, so by
default a verb consists of its stem followed by some suffix, and in a
past-tense verb, the most widely used suffix is et (e.g. KAST, 'throw',
68 Richard Hudson
past kästet). This defines the so-called "Large weak" class which we
can call simply "verb". Like English, Norwegian also has a class of
"Strong" verbs, which have no suffix and generally have a different
stem-vowel in the past (e.g. DRIKK, 'drink', past drakk)\ for these we
suppress the default suffix, and introduce another function, "stem-
vowel" which maps onto the last vowel of the stem. (We must take
this mapping for granted here.) The stem's default vowel is replaced
in the past tense by another which varies from verb to verb (with sub-
regularities that we shall ignore). This prose description shows how
both regular and irregular verbs are analysed in terms of the same
combination of functions and equations. These rules are diagrammed
in Figure 7.
whole
verbstron
stem-
vowel ROP
x
DRIKK verb t
vero strong past
v
verbsw: past
stem- suffix
vowel
Figure 7
Unlike English, however, Norwegian also has an important subclass

of weak verbs which have a number of peculiarities, one of which is
that their past-tense suffix is te or de instead of the default et (e.g.
ROP, 'call', past ropte). These verbs are also shown in Figure 7 as the
"Small weak" class (abbreviated here to SW). This sub-regularity is
treated in just the same way as the regular and irregular patterns: we
recognise a sub-class of verb and assign to it characteristics which
override the defaults. Sub-regularities such as this are a serious

problem for the dual-mechanism theory because the "rules" for them
are themselves exceptions.
Needless to say, in spite of its apparent complexity Figure 7 is a
considerable simplification. One area of analysis which it does not
try to show is the experiential basis of the generalisations. As
mentioned earlier, the total network is assumed to include not only
these stable high-level generalisations but also at least some of the
individual verb-tokens which were held in memory long enough to
yield the general patterns consisting of lexemes and general word-
types such as "verb:past". Moreover it is intended to accommodate
the relative frequency and salience of the various links, though at
present it is not clear exactly how this information can be formalised.
5. Conclusion: Symbolic networks
The main purpose of this paper has been to argue that the whole of
language (not just the lexicon) is a network with quite well-
understood and consistent formal properties based on default
inheritance and binary functions. This conclusion is completely
compatible with the Cognitive-Linguistics assumption that language
is part of the much larger network of general cognition.
The one issue which we have not considered is how these
networks fit into the contrast that Pinker (among others) draws
between "mentalese", the symbolic language of thought, and
"connectoplasm", thought as defined by a connectionist version of a
neural network (Pinker 1997: 128). Pinker believes that the mind is
divided between the two, though it may be unclear where the
boundary lies: "Where do the rules and representations in mentalese
leave off and the neural networks begin?" (1997: 112). The two
systems have quite different formal properties: mentalese is
symbolic, "digital", rule-governed, and uses concepts that are
"crisply" defined, whereas connectoplasm is an "analog" network in
which concepts are not defined (as such) at all, but related only by
degrees of similarity induced from experience. Given these formal
70 Richard Hudson
differences, it is unclear how the two parts communicate, so

alternatives should still be considered.
One alternative is that it is "networks all the way down", as
claimed in connectionism, but that at least some of the networks are
much more like Pinker's mentalese. We can call them symbolic
networks. Symbolic networks are widely used in cognitive
modelling, and indeed Pinker himself recognises the role of so-called
"semantic networks" within mentalese (1997: 87) as a way of
displaying a prepositional database. The same is true of the networks
of cognitive linguistics, including those that I have advocated in this
paper. A symbolic network is symbolic, like mentalese, in that each
concept has just one node and a finite and relatively fixed list of
properties; but it is a network because the properties are defined by
links to other concepts. It can be called the "language of thought"
because it has the same formal characteristics as human language—
namely, those of a network; thus it is the network that allows
compositionality and recursion, which Pinker picks out as especially
problematic for connectoplasm. As we have seen, it is merely a
matter of terminology whether we call the links in a network
"relations" or "rules", so we can say that a symbolic network is rule-
governed like mentalese. In short, symbolic networks have all the
properties that Pinker attributes to mentalese, so we may be able to
conclude that mentalese is a symbolic network.
Notes
1. I should like to thank Chet Creider, Joe Hilferty and an anonymous referee for
comments on an earlier version of this paper.
2. The books about Word Grammar are Hudson (1984, 1990, 1995, 1996, 1998).
The following web site holds downloadable material, including a 30-page
introduction and a 100-page encyclopedia which is updated annually:
http://www.phon.ucl.ac.uk/home/dick/wg.htm
Section II
Morphological aspects of the verb
The Norwegian verb
Rolf Theil Endresen and Hanne Gram Simonsen
This chapter aims at providing an overview of central aspects

concerning Norwegian verbs. Starting with a short description of the
Norwegian language situation and basic aspects of South East
Norwegian phonology, a survey of the main simple and periphrastic
verbal constructions and their use is given. Then follows an overview
of the system of verb classes, focussing on their prototypical
structure, and giving a description of their phonological patterning
and frequency characteristics. As will be argued in several
subsequent chapters in this book, it is necessary to take such features
of a language into account when evaluating different theoretical
approaches to morphology.
1. Introduction
Since not all readers can be assumed to know Norwegian, this paper
gives a short survey of the Norwegian verb, which plays a central
part in the majority of the papers in this book. We shall begin with a
few notes on Norwegian phonology, since morphological
generalisations often mention phonological conditions; the
description is based on South East Norwegian pronunciation (§2).
We then give an outline of morphological and periphrastic
constructions and their use (§3), followed by an overview of the verb
classes (§4).
The Norwegian verbal system is often presented in a way that
clearly reflects the classical Latin background of our grammatical
tradition. We want to present the system in a way that more clearly
reflects its prototypical structure.
74 RolfTheil Endresen and Hanne Gram Simonsen
1.1. Nynorsk and Bokmäl
But first of all we shall say a few words about the Norwegian
language situation. Our language has two official written forms,
called Nynorsk 'New Norwegian' ana Bokmäl 'Book Language'. The
official Norwegian way of classifying them is not as two languages
but as two 'language forms' (mälformer).
Nynorsk is written by <20% of the population and Bokmäl by
>80%. Let us emphasise that we are talking about written languages.
It would be very misleading to say that 20% of the population speak
Nynorsk and 80% speak Bokmäl. In fact most people speak their
local dialects, which correspond in differing degrees to the Nynorsk
and Bokmäl written standards.
Bokmäl and Nynorsk are very similar; when they differ, Nynorsk
words are preceded by a superscript N; "vere 'be' / \ori 'been',
while Bokmäl words are preceded by a superscript B; \cere 'be' /
\(Ert 'been'. Words that are the same in both written forms have no
mark: er 'am, is, are' / var 'was, were'.
Our last comment on the Norwegian language situation is that the
names of these language forms—Nynorsk and Bokmäl—are both
quite misleading. 'New Norwegian' is slightly more bookish than the
'Book Language' and the 'Book Language' is slightly newer than
'New Norwegian'.
2. A few notes on South East Norwegian phonology
2.1. Consonants
Most South East Norwegian dialects have the 23 consonant

phonemes depicted in Table 1.
Note that the relationship between /!/, /[/, and l\l is slightly more
complex than indicated in the table; a distinctive opposition between
/!/ and l\l is found only after the vowels /a(:)/ and /o(:)/. In other
contexts, [1] and [|J have a complementary distribution: [1] is found
after the laminals /t d s n/, and [|J is found elsewhere, i.e., word
initially, after all consonants but /t d s n/, and after all vowels but
/a(:)/ and /o(:)/. In these "other contexts", laterals will also be written
as/1/.
Traditionally, l\l has a marginal status in the standard dialect,
where IM is often used instead. In non-standard dialects, on the other
hand, it is more stable.
Table 1. The consonant phonemes of South East Norwegian dialects

Consonants labial laminal apical dorsal glottal
Unvoiced plosive P t t k
Voiced plosive b d et 9
Nasal m n π. D
Fricative f s J 9 h
Approximant υ 1 l j
Flap r
Tap r
2.2. Vowels
Most South East Norwegian dialects have the nine short and the nine
long vowel phonemes in Table 2.
Table 2. The vowel phonemes of South East Norwegian dialects

Short Vowels Long Vowels
i y « u i; y: «: u:
e 0 O e: 0: o:
as ae:
a a:
Eleven vowels are front, /i i: y y: u «: e: 0 0: ae ae:/; four vowels are

back, /u u: ο οι/; two vowels vary between front and back (socially,
geographically, and depending on the quantity), /a a:/; and one
vowel, Id varies between front and central (depending on stress).
2.3. Word accent
A Norwegian word normally has only one accented syllable, but

there are two distinctive word accents (or tönernes), in South East
Norwegian dialects; a low accent and a high-low accent. They are
not marked orthographically, but in our transcriptions we mark the
low accent (L) with a grave accent £} and the high-low accent (HL)
with a circumflex accent Q; the accent marks are placed above the
vowel of the accented syllable.1 E.g., in Bokmäl, the two present
indicative forms gjelder 'applie(s)' and gjelder 'castrate(s)' have
different words accents—consider (1).
(1) gjelder 'applie(s)' /jeler/

gjelder 'castrate(s)' /jeler/
When the last syllable of a word carries the accent, we can only have
an L accent, with the exception of some verb-plus-particle con-
structions not discussed in this paper.
3. Morphological and periphrastic constructions
3.1. The three finite simple forms
Norwegian has three finite simple verbal forms, where simple means
non-periphrastic. In Table 3, these forms are illustrated with the verb
D0MME 'judge, deem'.
Table 3. The three finite simple forms of D0MME 'judge, deem'

Meaning Orthography South East Norwegian
Mood Tense pronunciation
Indicative Present dimmer /domer/
Indicative Past d0mte /dornte/
Imperative — d0m /dam/
The finite forms have no person and number inflection. Notice

furthermore that the indicative forms are aspectually neutral, as seen
in (2) and (3).
(2) Present indicative ofARBElDE 'work'

a. Arbeider Eva her?
work.PRES Eve here
'Does Eve work here?' (Habitual)
b. Ser du "ikkje/'ikke at vi arbeider?
see.PRES you not that we work.PRES
'Don't you see that we are working?' (Progressive)
(3) Past indicative ofBARBERE 'shave'

a. Han barberte seg i gär.
he shave.PAST himself yesterday.
'He shaved yesterday.' (Perfective)
b. Eva barberte Adam da vi kom.
Eve shave.PAST Adam when we arrive.PAST
'Eve was shaving Adam when we arrived.' (Progressive)
3.2. The three non-finite simple forms
Norwegian also has three non-finite simple forms—cf. Table 4:
Table 4. The three non-finite simple forms of D0MME 'judge, deem'

Meaning Orthography South East Norwegian
Mood Tense pronunciation
— — d0mme /dome/
— Present ^mmande, *d0mmende 7d0mane/, Vd0mene/
— Past d0mt /d0mt/
The present participle is a marginal form which not everybody agrees

to including in the verbal paradigm. The infinitive and the past
participle will be discussed in connection with the periphrastic forms.
3.3. Types offormatives
Now, take a look at Table 5, where we have tried to show the major
types of formatives that are used in Norwegian verb inflection, by
means of two verbs that both are pronounced /jele/ in the infinitive.
The past indicative and past participle of 'castrate' has two
alternative forms in Bokmäl: /jela/ gjelda and /jelet/ gjeldet
(Nynorsk has only /jela/ gjelda). The choice between them is
sociolinguistically or stylistically conditioned. In urban areas, /jela/
gjelda is roughly a low prestige working class form and /jelet/ gjeldet
roughly a high prestige middle and upper class form. Both forms can
be used in the written language—but not surprisingly, the high
prestige forms dominate.2
Table 5. Two GJELDE verbs

GJELDE 'castrate'
Infinitive Present indicative Past indicative Past participle
/jlie/ 7jelar/, Vjeler/ /jela/, Vjelet/ /jela/, Vjelet/
HL,-e HL, -ar/-er HL,-a/-et HL, -a/-et
gjelde "gjeldar, "gjelder gjelda, "gjeldet gjelda, "gjeldet
GJELDE 'apply (to)'

Infinitive Present indicative Past indicative Past participle
/jele/ 7jel/, Vjeler/ 7galt/, /jält/ 7goli/, /jelt/
HL, -e L, -er L, -t, e ~ a L, -t, e -
gjelde "gjeld, "gjelder "galdt, gjaldt "goldi, gjeldt
Just like in English, both suffixes and vowel alternations are used as
formatives in Norwegian. In addition, as mentioned in section 2.3
above, Norwegian has a phenomenon not found in English: different
word accents. Word accent also plays a role as a formative: in the
present indicative, HL accent is typical of a weak verb, while L accent
is typical of strong verbs. In §4 below we shall show more
systematically how the formatives are employed in the different
conjugational classes.
3.4. Periphrastic forms I: The perfect
Norwegian has many periphrastic forms, and no exact number can be

given. Some highly grammatical!sed ones are those using the
auxiliary verb HA 'have', which has the paradigm shown in Table 6.
Table 6. HA 'have'
Meaning Orthography S.E. Norw. pronunciation
Infinitive ha /haV
Present indicative har /ha:r/
Past indicative hadde /hade/
Past participle hatt /hat/
This auxiliary combines with the past participle in the perfect forms,
as in Table 7. The perfect forms are used roughly as in English. For
some types of intransitive verbs—for example, motion verbs—the
auxiliary "VEREI *v/ERE 'be' can be used instead of HA—cf. Tables 8
and 9.
Table 7. Periphrastic forms: the perfect of KJ0PE 'buy'

Present perfect har kj0pt /ha:r 90pt/
'has bought' have.PRES buy.PAST.PCP
Past perfect hadde kj0pt /hade ςα
'had bought' have.PAST buy.PAST.PCP
Table 8. "VERE 1 *V£RE 'be'

Infinitive "vere, "vcere /uae:re/
Present indicative er /ae:r/
Past indicative var /ua:r/
Past participle "vori, "van 7uo:ri/, Vuaet/
Table 9. Periphrastic forms: the perfect of REISE 'travel, go, leave'

Present perfect har reist /hair raejst/
'has gone' have.PRES go.PAST.PCP
er reist /KIT raejst/
be.PRES go.PAST.PCP
Past perfect hadde reist /hade raejst/
'had gone' have.PAST go.PAST.PCP
var reist /ua:r raejst/
be.PAST go.PAST.PCP
3.5. Periphrastic forms II: the future
The future can be expressed in several ways in Norwegian; the three

presented in (4) are the most common ones in South East Norwegian.
(4) a. Using the simple present indicative

"Dei/BDe "kjem / "kommer snart.
they come.PRES soon
They'll come soon.'
b. Periphrastically; SKULLE 'shall' and VILLE 'will' as aux.

"Eg/"Jeg skal "gjere / Bgj0re del i "morgon / "morgen.
I shall do.lNF it tomorrow
shall do it tomorrow.'
Det vil bli regn ikveld.
it will become.lNF rain tonight
There will be rain tonight.'
c. Periphrastically, with KOMME 'come'

Det "kjem / "kommer til ä
it come.PRES tO.PREP tO.INF.MARKER
bli "ein/'en fin dag.
become.lNF a fine day
'It will be a fine day.'
3.6. Periphrastic forms HI: the progressive
As stated in section 3.1 above, the simple finite forms are aspectually
neutral in Norwegian. This does of course not mean that there are no
ways of expressing different aspects. For example, there are several
ways of expressing a progressive meaning. One common way is
found in (5).
(5) The progressive expressed by a position verb

Han Nligg/ligger og Nh0yrer /h0rer pa ei / "en ny plate.
he lie.PRES and hear.PRES on a new record
'He is (lying) listening to a new record.'
N
Jenta/"Piken "sit/sitter og les/leser.
girl.the sit.PRES and read.PRES
The girl is (sitting) reading.'
Mennene star og diskuterer problemet.
men.the stand.PRES and discuss.PRES problem.the
The men are (standing) discussing the problem.'
From a grammaticalisation perspective, the use of position verbs to

express progressive meaning has evident connections with the fact
that position verbs are used much more in Norwegian than in
English. Consider the sentences in (6), where position verbs are used
in describing the location of inanimates. In English, it is more natural
to use the verb BE in such cases, while this often sounds rather
artificial in Norwegian.
(6) Position verbs used with inanimate subjects

Boka/'Boken "Hgg/ligger pa hylla / "hyllen.
book.the lie.PRES on shelf.the
The book is (lying) on the shelf.'
Koppen star pä bordet.
cup.the stand.PRES on table.the
The cup is (standing) on the table.'
3.7. Passive and related meanings
A few words should be said about the expression of passive. There

are three constructions that are traditionally classified as passive, one
simple and two periphrastic ones—cf. (7):
(7) a. Simple passive: S-verb

Boka/"Boken skal "omsetjas / "oversettes
book.the shall.PRES translate.lNF.PASS
til norsk.
to Norwegian
The book shall be translated into Norwegian'
b. Periphrastic action passive: BLI 'become'
Boka/'Boken blei/"ble "omsett / "oversatt
book.the become.PAST translate.PAST.PCP
til norsk.
to Norwegian
The book was translated into Norwegian'
c. Periphrastic state passive: "VEREIBV/ERE 'be'
Boka/BBoken er "omsett / Oversatt til norsk.
book.the is translate.PAST.PCP to Norwegian
The book is translated into Norwegian'
The so-called s-verbs do not always have a passive meaning. They

may roughly be divided into three types: passive, reciprocal, and de-
ponent, see Table 10.
Table 10. S-verbs

Passive Reciprocal Deponent
'be bought' 'meet' 'think'
N N
Infinitive kj0pas / *kj0pes m0tas/Bm0tes "synas/Bsynes
Present indicative kj0pes m0tes synes
Past indicative (kjtfptes) m0ttes syntes
Past participle m0ttes syntes
Deponent verbs are verbs with no counterpart without the -s suffix.

Notice also that passive s-verbs generally are not much used outside
the infinitive and the present indicative. In Bokmäl, passive s-verbs
are used quite generally in the infinitive and the present indicative,
while the past indicative form is marginal. In Nynorsk (where the s-
verbs may end in -st instead of -s) and in most non-standard dialects
the infinitive is in general use, while the present indicative is
marginal and the past indicative non-existent.
Furthermore, it should be noticed that in Bokmäl (but not in
Nynorsk), the infinitive and the present indicative of the s-verbs are
practically always formally identical. As argued by Enger (1998b),
they should not necessarily be analysed as two different forms of the
verb in Bokmäl.
4. More on Norwegian verbal morphology
4.1. Nynorsk and Bokmäl revisited
Although Nynorsk and Bokmäl are very similar, there are some
morphological differences between them. Generally, Nynorsk has a
slightly more complex morphology than Bokmäl, with more
inflectional classes and a larger variation of forms. In the verbal
morphology, the differences are clearest in the indicative forms, in
particular in the present indicative, as illustrated in Table 11 on the
next page.
4.2. Strong and weak verbs
Norwegian verbal morphology is in many respects similar to that of

other Germanic languages—for instance, it shares the basic
distinction between strong and -weak verb classes. (We use these
classical Germanic terms strong and weak, instead of the terms
irregular and regular, since we find those a bit misleading: using the
term irregular about strong verbs precludes somehow the possibility

that there may be important regularities among these verbs.)
Table 11. Paradigms for some Norwegian verbs

KASTE KJ0PE SKRIVE SOVE
'throw' 'buy' 'write' 'sleep'
Infinitive /käste/ /skri:ue/ /so:ue/
N
Present /kästar/ /90:per/ 7skri:u/ 7s0:u/
B
indicative /käster/ /skri:uer/ 7s0-;uer/
Vsö:uer/
Past /kästa/ /90pte/ /skraeju/ /sö:u/
indicative "/kästet/ Vskre:iV
Past /kästa/ /9<äpt/ 7skri:ui/ 7so:ui/
participle "/kästet/ Vskre:uet/ Vsöiuet/
N
Present Tkästane/ /90:pane/ 7skri:uane/ 7so:uane/
B
participle Vkästene/ /90:pene/ Vskri:uene/ Vsö:uene/
Imperative /käst/ /90:p/ /skriiu/ /sö:iV
In English, the formal distinction between strong and weak verbs is

based on whether they carry a suffix in the past tense or not. In
Norwegian, we may rather base the distinction on whether they carry
a syllabic suffix or not: the verbs carrying a syllabic suffix in the past
indicative are weak, those without a syllabic suffix in the past
indicative are strong. This implies that a few strong verbs may carry
a suffix, as long as the suffix is not syllabic, as illustrated with three
Bokmäl verbs in Table 12:
Table 12
Infinitive Past ind.
TRALLE 'sing, trill' /träle/ /trälet, träla/ —» Weak (larger class)
KALLE 'call' /kale/ /kälte/ —> Weak (smaller class)
FALLE 'fall' /fäle/ /fait/ —» Strong (no change class)
It also implies that verbs with both a vowel change and a syllabic
suffix are counted as weak despite the vowel change, as exemplified
in Table 13, but they are irregular weak verbs, as compared to those
without a vowel change (and there are few such irregular weak verbs,
about 10 in Bokmäl and 20 in Nynorsk).
Table 13
Infinitive Past ind.
*VELGE 'choose' Vuelge/ Vualkte/ —> Weak (smaller class)
"VELJE 7uelje/ 7ualte/
"SP0RRE 'ask' Vsp0re/ /sp«:te/ —> Weak (smaller class)
"SP0RJE "/spdrje/
These examples also serve to illustrate the prototypical structure of

the verb classes. We will return to this point shortly.
For ease of exposition, we will restrict ourselves to Bokmäl in the
rest of this paper. The inflectional patterns are based on the standard
dialect with a South East Norwegian pronunciation, as spoken in
Oslo. The basis for the description of Bokmäl verbs is The Bokmäl
Dictionary (Bokmälsordboka, BO) from 1986. In BO are listed a
total of approximately 4800 verbs, with a distribution of strong and
weak verbs illustrated in Figure 1. (All numbers are approximations.)
THE NORWEGIAN BOKMÄL DICTIONARY

Total number of verbs: 4800
Strong: 200
(4%)
V-class 2700
/ (WL) (56%)
Weak: 4600
(1 5%)
(96%) ./de/ 150 ' '
\
C-class 1900 Bother 80
(WS) (40%) . (1,5%)
^
-/te/ 1750 ^"ere
verbs 1500
V. <32%>
other 250
(5%)
Figure 1. The distribution of weak and strong verbs according to Bokmalsordboka
The terminology V-class and C-class is used to refer to the suffixes

in the two weak classes: The vocalic conjugation or V-class has past
suffixes which are vowel initial, i.e., /-a/ or /-et/, while the
consonantal conjugation or the C-class has past suffixes which are
consonant initial, i.e., /-de/ and /-te/ in the past indicative and /-d/ and
/-t/ in the past participle. This terminology was first introduced in
Endresen (1992). Focussing instead on the number of members in the
conjugational classes, Simonsen (1993, this volume) calls the vocalic
conjugation the larger weak class (WL class) and the consonantic
conjugation the smaller weak class (WS class).
There is no general agreement in Norwegian grammars on what to
call these classes. Faarlund, Lie, and Vannebo (1997: 492) refer to
the vocalic conjugation as the 1st class of weak verbs and the con-
sonantal conjugation as the 2nd class of weak verbs, while more
traditional grammars, like Beito (1986: 302-308); Naes (1972: 181-
187) refer to the vocalic conjugation as the 1st class of weak verbs,
while the consonantal conjugation is split up into the 2nd, 3rd, and
4th classes of weak verbs; this is in accordance with grammars of
Old Norse, like Noreen (1923). Aasen (1864) divides the consonantal
conjugation into the 1st and 2nd classes of weak verbs, while he
refers to the vocalic conjugation as the 3rd class of weak verbs, while
Western (1921) on the other hand refers to the vocalic conjugation as
the 1st class of weak verbs and divides the consonantic conjugation
into the 2nd and 3rd classes of weak verbs.
4.2.1. Strong verbs
As already mentioned, the absence of a syllabic suffix is chosen as

the basic feature singling out strong verbs. This entails that the strong
past indicative form ends in an accented syllable—this is a salient
phonological characteristic, and might be a good basis to form
generalisations over, as indicated in the schema in Figure 2, where
the symbol 'σ indicates an accented syllable.3
'
\\
[[PROCESS, PAST, INDICATIVE] / [...v...]]
Figure 2. Strong verb schema
In addition, there are several other features characterising strong

verbs, to a larger or smaller degree. Most salient and most often cited
is what is often called a vowel change, i.e., a different vowel in the
infinitive and past indicative stems. (These verbs differ as to which
vowel they have in the past participle—the infinitive stem vowel, the
past indicative stem vowel or a third vowel.) Another salient feature
of strong verbs is that they carry the L accent (low tone) in the
present indicative, as opposed to most weak verbs, which carry the
HL accent (falling tone).4 This feature is exemplified in Table 14:
Table 14
Weak Strong
Infinitive Pres. ind. Past ind. Infinitive Pres. Ind. Past ind.
7Y7TE 'look' /titer/ /titet/, SITTE 'sit' /siter/ /sät/
/tita/
LYSE 'shine' /ly:ser/ /ly:ste/ FRYSE 'freeze' /fry:ser/ /fr0:s/,
/fr0js/
Lastly, as indicated in Table 12 above, a few verbs (some with a

vowel change and some without a vowel change) form the past tense
with a non-syllabic suffix.
These four features characterising strong verbs are listed in Table
15, along with the numbers of verbs covered by each.
Table 15. Features characterising strong verbs

1. L accent in the present indicative 151
2. Vowel change 143
3. Past indicative form ending in an accented syllable5 137
4. Past indicative with a non-syllabic suffix 13
88 Rolf Theil Endresen and Hanne Gram Simonsen
The three first features, which are all phonologically salient, are
shared by a lot of the strong verbs, while the last (and possibly less
salient) feature covers only a few. Features 1 and 2 are also shared by
some weak verbs—this is why the number of verbs covered by these
two features exceeds the number of verbs covered by feature 3,
which was chosen as the defining feature for strong verbs. Finally,
there is a small set of verbs sharing the three first features, but each
having an irregularity of its own in the past tense inflection—the past
indicative forms of these verbs may be characterised as more or less
suppletive (e.g., gi 'give': /ji:/ - /g :/, g 'go': /go:/ - /jik/, ligge 'lie
down': /lige/ - /lo:/). Thus, the category of strong verbs has a
prototypical structure, illustrated in Figure 3 below.
(a) (b)
'σ
I l\
[[PROC, PRES, IND] / [...V,...r]] [[PROC, PAST, IND] / [...V...]]
(c) (d)
I
'σ 'σ
l· l\
[[PROC, PAST, IND], [[PROC, PAST, IND] / [...VCt]]
Figure 3
In Figure 3, the square (a) represents the present indicative schema of

the strong verbs, with a L tone on the accented syllable and an -r
suffix. The square (b) represents the most schematic or general past
indicative schema for strong verbs (identical to Figure 2 above),
ending in an accented syllable with an L tone. The square (c) is the
most prototypical instantiation of (b), from which it differs by having
a different vowel (Vb) than the one in the present tense (Va), i.e., so-
called vowel change. Finally, the square (d) represents the schema
for past indicative forms with a non-syllabic suffix.
Verbs likesitte 'sii'-satt.fryse ' freeze'-fr0s/fr0ys, stä 'stand'-sto,

bite 'bite'-bet/beit share the 3 first features, and are typical strong
verbs. Slightly less typical strong verbs are verbs like brenne 'burn'-
brant, sharing the three first features, but in addition having feature
4. Some verbs (so-called No Change strong verbs) like komme
'come'-fcom, I0pe 'run'-/0p share features 1 and 3, but not 2—a few
of these, e.g. falle 'fair-fait, holde 'hold'-holdt also have feature 4.
Table 16. Strong verb subclasses in Bokmäl

Sub- Inf. PT Examples No.
class stem stem of
vowel vowel Vs
1 lil /a/ dritte 'drink' /drike/ /dräk/
lei dette 'fall' /dete/ /dät/
lyl jynê 'sing' /syrje/ /säry 40
2 li'J /ey bite 'bite' /birte/ /be:t/
~/aej/* ~/bajf* 33
3 /yy /0-J ./fyte 'flow' /fly:te/ /f!0:t/
~/fty:te/ ~/ft0jt/* 27
4 /ay /u-J rfra 'pull' /dray /drüy
/ey le 'laugh'
lo-J j/ 'hit' /jloy /jiuy
/ae/ sverge 'swear' /suaSrge/ /suu:r/ 14
5 /aey /ay bcere 'carry' /baSire/ /bä:r/
/ey be 'ask' /bey /bä-7
/iy si 'say' /say
/e/ /egge 'put' /lege/ /läy 9
6 /oy No save 'sleep' /sö:ue/ /sö:u/
/o/ change komme 'come' /körne/ /körn/
/a/ falle 'fall' /fäle/ /fält/
/ey hete 'be called' /he:te/ /he:t/
/0y /0/?e 'run' /10:pe/ /10:p/ 9
The se 'see' /sey /söy
Rest £« 'go' /g*/ /jik/
etc. 5
* With a few exceptions, the choice between a long vowel and a diphthong in these
verbs is sociolinguistically or stylistically conditioned.
Subclasses 1, 2, and 3 all have a relatively high number of members.

The phonologically most coherent classes among these are classes 2
and 3, with one single vowel in the infinitive stem, while class 1 is
slightly less coherent since it has three stem vowels in the infinitive.
As for the classes 4 onwards, they have a clearly lower type
frequency and are much less phonologically coherent—thus the
psychological plausibility of their status as subclasses may be
questionable.
To group the different strong verbs into subclasses, we have
chosen a classification based essentially on the stem vowel in the
past indicative form, i.e., a product-oriented definition. One subclass,
however, is partly source-oriented—the No Change class, which is
characterised as having the same stem vowel in the infinitive and the
past indicative (i.e., verbs lacking feature 2 above). This
classification is not uncontroversial (see for example the discussion
in Enger 1998a);6 however, it has the advantage of being strictly
synchronic, in addition to having a potential psychological
plausibility as providing identifiable patterns. The strong verb
subclasses in Norwegian are listed in Table 16, in an order of
descending type frequency.
4.2.2. Weak verbs
Unlike English, Norwegian has not only one class of weak verbs, but
two. The two main classes are distinguished through the different
suffixes they carry. As indicated in Figure 1 above, the larger of the
weak classes in Bokmäl (the V(owel) class or the larger weak (WL)
class) has approximately 2700 members. This class takes a past
indicative suffix beginning with a vowel (-/et/ or -/a/), where the use
of either of these is sociolinguistically or stylistically determined (cf.
§3.3 above). The smaller one (the C(onsonant) class or the smaller
weak (WS) class) has approximately 1900 members and takes two
suffixes, both beginning with a dental consonant (-/te/ and -/de/). The
use of these two suffixes is lexicalised for each verb.
The WL class is clearly the most phonologically open class. With

a few possible exceptions which we will mention below, stems with
any form can take this suffix. It is also the class which is generally
productive, in the sense that new verbs entering the language
normally fall into this class.
The WS class is phonologically less open. Two subclasses of this
class can possibly be said to be phonologically defined (although the
second subclass needs morphological information for its proper
definition)—in any case each is phonologically very coherent. The
first subclass includes verbs with a stem ending in the derivational
affix -er /e:r/, which are inflected with the -te suffix, as exemplified
in Table 17:
Table 17. Verbs with a stem ending in -er. WS with -te suffix
reparere 'repair' /repareire/ — /repar&te/
sortere 'sort' /sot&re/ — /so\e:\e/
markers 'mark' /marke:re/ — /markeê/
This subgroup of verbs is actually very large, with approximately

1500 members (thus constituting nearly 80% of all the WS verbs,
and 37 % of all verbs in Bokmal!), and since suffixation with -er is a
productive process for verb derivation in Norwegian, the class is
growing. However, generally these verbs have a very low token
frequency. (Aspects of these verbs are treated by Gundersen this
volume.)
The second subclass, exemplified in Table 18, contains all weak
verbs with a monosyllabic infinitive form. These count
approximately 70 verbs, and all take the -de suffix.
Table 18. Weak verbs with a monosyllabic inf. form: WS with -de suffix
sy 'sew' Isy'J — /syde/
na 'reach' /no:/ — /node/
tree 'thread' /treV — /trede/
92 Rolf Theil Endresen and Hanne Gram Simonsen
A third, irregular subclass of WS verbs are verbs with a vowel

change in addition to the syllabic suffix in the past indicative (cf.
Table 13 above).
The rest of the WS verbs, approximately 350 verbs, form a
phonologically relatively open class, but less so than the WL class.
E.g., there are restrictions on the set of stem final consonants for WS
class verbs: verbs with stem final /b {ιχ J il do not occur in this class,
and verbs with stem final consonant clusters are very rare in the WS
class (limited to the two verbs hilse 'greet' - hilste and tenke 'think'
- lenkte). But there are also some more subtle phonological
distinctions governing the distribution of weak verb stems, helping in
determining their class membership. As pointed out by Hagen (1994)
and Bjerkan (2000), there is a strong correlation between the length
of the infinitive stem vowel and verb class. In the WL class,
approximately 80% of the verbs have a short stem vowel, and only
20% have a long stem vowel, while the opposite is the case for the
WS class (these counts exclude the -ere verbs). In addition, in some
cases vowel quality may be a good indicator of class membership: In
the WL class, the short /a/ is the most frequent stem vowel, occurring
in 20% of the verbs, while there is only one single verb with a short
/a/ in the WS class (kalle 'call' - kalte). In the WS class, on the other
hand, /0:/ is the most frequent stem vowel, while there are very few
WL verbs with this vowel. A summary of these phonological
characteristics of the two weak classes is given in Table 19 on the
opposite page.
More research needs to be done for us to be able to characterise
the phonological patterning of the weak classes properly. Many of
the phonological characteristics we have identified so far are only
tendencies, thus not supplying clear-cut class boundaries. However,
these tendencies may have implications for processing and for class
formation during acquisition—questions addressed in the papers by
Bjerkan, by N0klestad, and by Simonsen (all this volume).
Table 19. Phonological characteristics of the two weak classes

WL class WS class
Essentially open less open
Generally productive subclasses, possibly phonologically defined:
• weak verbs with a monosyllabic infinitive
• -ere verbs (with stem ending in -/e:r/)
80% short stem vowel 80% long stem vowel (excluding -ere verbs)
20% long stem vowel 20% short stem vowel
/a/ most frequent stem vowel: /a/ extremely rare as stem vowel:
(20% of verbs) (one verb only)
I0'J very rare as stem vowel I0'J most frequent stem vowel
non-restricted set of stem final restricted set of stem final consonants
consonants
wide range of stem final very few verbs with stem final
consonant clusters consonant clusters (two verbs only)
5. Summary
In this paper, we first gave a short description of the Norwegian

language situation, with the two official language forms, Bokmäl and
Nynorsk, followed by a short summary of South East Norwegian
phonology. Then followed a survey of the main simple and
periphrastic verbal constructions and their use. There are three
simple finite forms (present indicative, past indicative, and
imperative), three simple infinite forms (infinitive, present participle,
and past participle), and various periphrastic forms, expressing, i.a.,
perfect, future, progressive and passive.
Finally, we gave an overview of the system of verb classes. We
focused on the phonological patterning of the classes, taking into
account their frequency characteristics and their prototypical
structure. Norwegian verbs can be divided into strong and weak
classes. Strong verbs have no suffix or a non-syllabic suffix in the
past indicative, while weak verbs have a syllabic suffix. We divided
the strong verbs into six subclasses of varying regularity, plus one
subclass of irregular strong verbs; the classification is based

essentially on the stem vowel in he past indicative, i.e., a product-
oriented definition. A salient feature of the Norwegian verb class
system is the fact that it has not only one, but two main weak classes:
1) The V(owel) class or The larger weak class (WL), and 2) The
C(onsonant) class or The smaller weak class (WS), the first of which
is phonologically more open than the second. This fact plays a
central role in several of the papers in this volume.
Notes
1. In the literature on Norwegian, the word accents are often called Tönerne l and
Tönerne 2, and marked accordingly with a raised number ' or 2 in front of the
accented syllable. In South East Norwegian, Toneme 1 corresponds to the L
accent (low tone), while Toneme 2 corresponds to the HL accent (falling tone).
2 According to Bokmälsordboka, the verb meaning 'castrate' also has the past
indicative form gjeldte /jelte/, but we have ignored forms that to our knowledge
are not current in South East Norwegian dialects. Both Nynorsk and Bokmäl
accept considerable variation.
3. In accordance with Norsk referansegrammatikk The Norwegian Reference
Grammar' (Faarlund, Lie, and Vannebo 1997: 480), approximately 10 very
irregular verbs (modals plus a few other irregular verbs) are counted as
different from from both weak and strong verbs and form a separate group—so-
called preterito-presentic verbs.
4. A few weak verbs have the L accent in the present tense, and there are even a
couple of strong verbs which have the HL accent in the present tense. However,
these discrepancies are due to diachronic movement between inflectional
classes.
5. The total number of strong verbs counted here is 137. This number is lower
than the approximate number of strong verbs cited from Bokmälsordboka to be
approximately 200—cf. the figure in (21). The reason for this is that our verb
classification differs slightly from the one assumed to be underlying the verb
class assignment in Bokmälsordboka; furthermore we have only included non-
compound strong verbs, and we have excluded a few verbs that we do not
consider strong in Bokmäl any longer.
6. Interestingly, until recently, Western (1921) was the only description of this
type in the available literature on Norwegian verb classes. However Faarlund,
Lie, and Vannebo (1997) have adopted a similar solution.
Building blocks or network relations:
Problems of morphological segmentation
Helge Gundersen
Traditionally, morphological analysis is identified with an exhaustive

segmentation into discrete "chunks". This building-block model of
morphology is contrasted with a cognitive-linguistic network
conception of morphology, which may be viewed as its diametrical
opposite. The building-block model induces several unnecessary
problems, and at the same time overlooks certain morphological
phenomena. Three segmentation issues are presented: phonological
residues (cranberry morphs), morphological ambiguity (overlapping
items), and submorphemic relations (morphologically complex items
lacking constituents). The remainder of the paper focuses on
problems of synchronic morphological segmentation in one part of
the Norwegian language, namely in a group of verbs derived by the
suffix-er.1
1. Two conceptions of morphological analysis
1.1. The building-block model
In grammatical theory, it is common to separate grammar from

general skills and knowledge. A morphological analysis will then not
be a study of how the language user handles the words in question.
Instead, it may be seen as a study of how grammar constructs its own
output. Morphology is only seen as a matter of objective composition
independently of the language user. The foundation of this way of
thinking is objectivist philosophy (cf. Lakoff and Johnson 1980: 195-
209; Lakoff 1987: 157-184; Gundersen 1997). Lakoff and Johnson
put it this way (1980: 204):
96 Helge Gundersen
According to the myth of objectivism, objects have properties in and of

themselves and they stand in relationships to one another independently of
any being who understands them. When words and sentences are written
down, they can be readily looked upon as objects. This has been the
premise of objectivist linguistics from its origins in antiquity to the present:
Linguistic expressions are objects that have properties in and of themselves
and stand in fixed relationships to one another independently of any person
who speaks them or understands them. As objects, they have parts—they
are made up of building blocks: words are made up of roots, prefixes,
suffixes, infixes; sentences are made up of words and phrases; discourses
are made up of sentences. Within a language, the parts can stand in various
relationships to one another, depending upon their buildingblock structure
and their inherent properties. The study of the buildingblock structure, the
inherent properties of the parts, and the relationships among them has
traditionally been called grammar.
Langacker (1987a: 452) speaks of a building-block metaphor, "which

sees the meaning of a composite expression as being constructed out
of the meaning of its parts simply by stacking them together in some
appropriate fashion". It seems clear that the notion of a building-
block metaphor may be extended from the semantic compositionality
of expressions (which is Langacker's point) to the morphological
structure in general, cf. Lakoff and Johnson's mention of roots and
affixes in the quote above. In structuralist and generative linguistics
morphology is to a large extent treated as an exhaustive segmentation
into discrete constituents seen as "building blocks". This approach is
depicted by the simple algebra in (1), in which A and B are the
building blocks of the composite structure C.
(1) A+B=C
A complex word is then seen as made up of its constituents, and the

aim is to find these building blocks and the rules that combine them.
When the building-block model is made serious use of in
morphological theory it must be viewed in the light of underlying
objectivist assumptions. Complex words are seen as constructed
independently of the language user, and their morphological
constituents then naturally become the building blocks.
One important manifestation of this way of thinking is the mor-

pheme-based structuralism of Bloomfield and his followers. This tra-
dition is still taken as a point of departure in much of contemporary
morphology. Matthews (1993: 98-110) has examined not only stan-
dard introductory textbooks in linguistics, such as Fromkin and Rod-
man (1988), but also influential generative treatments, such as those
by Halle (1973), Williams (1981), McCarthy (1982), and Lieber
(1989):
[T]he theory of strict lexicalism is the most important new development in

morpheme-based morphology. Of the deeper hypotheses on which it rests,
one was an innovation: that derived words must pass through a lexical filter.
The other can be traced back to the origins of structuralism, half a century
earlier.
This second assumption seems to have been automatic. When I read
Lieber's paper on the percolation of morphosyntactic features (1989), I
pencilled in my copy of the journal: OK but why should the features
originate in morphemes in the first place?' There might have been
substantive reasons, but Lieber does not give them. It was simply taken for
granted from the outset. Sixteen years earlier, Halle had begun his
'Prolegomena' by pointing out that speakers know the words of their
language. They also know' that a word like trans-form-at-ion-al 'is
composed of the morphemes shown and that these five morphemes cannot
be concatenated in most other orders'. Such knowledge must be represented
in a grammar: 'hence', he goes on, 'the assumption has been made quite
generally that a grammar must include a list of morphemes as well as rules
of word formation' (1973: 3). It had been made indeed, first by Bloomfield
and then by those who followed him. (Matthews 1993: 110)
Jensen, in his introduction to generative morphology (1990), bases

much of his exposition on Nida's (1949) principles for the identifi-
cation of morphemes, which was an important contribution to bloom-
fieldian morphology. Before quoting Nida's principles, Jensen him-
self (1990: 21) states a principle not supplied by Nida which, I be-
lieve, appears to give much of the essence of the building-block mo-
del of morphology. Jensen is not necessarily representative of gene-
rative morphology (cf. Marantz 1992), but I mention his principle
here, as I think it will serve well as an introduction to Nida's
principles:
98 Helge Gundersen
(2) A word must be exhaustively divided into morphemes.
Nida's (1949) principles are quoted here without commentary:
Principle 1
Forms which have a common semantic distinctiveness and an identical
phonemic form in all their occurrences constitute a single morpheme.
Principle 2
Forms which have a common semantic distinctiveness but which differ in
phonemic form (i.e. the phonemes or order of the phonemes) may constitute
a morpheme provided the distribution of formal differences is phono-
logically definable.
Principle 3
Forms which have a common semantic distinctiveness but which differ in
phonemic form in such a way that their distribution cannot be phono-
logically defined constitute a single morpheme if the forms are in
complementary distribution in accordance with the following restrictions:
1. Occurrence in the same structural series has precedence over occurrence
in different structural series in the determination of morphemic status.
2. Complementary distribution in different structural series constitutes a
basis for combining possible allomorphs into one morpheme only if there
also occurs in these different structural series a morpheme which belongs to
the same distribution class as the allomorphic series in question and which
itself has only one allomorph or phonologically defined allomorphs.
3. Immediate tactical environments have precedence over non-immediate
tactical environments in determining morphemic status.
4. Contrast in identical distributional environments may be treated as
submorphemic if the difference in meaning of the allomorphs reflects the
distribution of these forms.
Principle 3 is simplified by Nesset (1998b: 74) in this way: "Sets of

morphs realizing the same content belong to the same morpheme if
they are in complementary distribution."
Principle 4
An overt formal difference in a structural series constitutes a morpheme if
in any member of such a series, the overt formal difference and a zero
structural difference are the only significant features for distinguishing a
minimal unit of phonetic-semantic distinctiveness.
Principle 5
Homophonous forms are identifiable as the same or different morphemes
on the basis of the following conditions:
1. Homophonous forms with distinctly different meanings constitute diffe-
rent morphemes.
2. Homophonous forms with related meanings constitute a single mor-
pheme if the meaning classes are paralleled by distributional differences,
but they constitute multiple morphemes if the meaning classes are not
paralleled by distributional differences.
Principle 6
A morpheme is isolatable if it occurs under the following conditions:
1. In isolation
2. In multiple combinations in at least one of which the unit with which it is
combined occurs in isolation or in other combinations.
Jensen is probably correct in stating that his principle seems crucial

to an understanding of Nida's principles. However, while the prin-
ciple (2) says that the segmentation should be exhaustive it does not
per se say that this segmentation is the only way of analysing a word
morphologically. Yet, another consequence of the building-block
model is that analysing a complex word amounts to finding its
constituents. Let us formulate this as another principle:
(3) There is no other way that a word can be morphologically

complex than by having discrete components.
1.2. Cognitive morphology
There are several alternatives to a rigourously implemented building-

block approach, and a spectrum of positions may be envisaged. The
alternative view presented here will be found at the quite opposite
end of this spectrum. According to this view morphological structure
is considered as a set of relationships, or lexical connections, within
the mental lexicon, and morphological elements as linguistic units
activated through the lexical connections when the words are being
used (Bybee 1985, 1988, 1995b). In this mode of thinking, the word
100 Helge Gundersen
as a whole, the composite structure, has its own semantic value, in-
stead of this being solely derivable from the meanings of the morpho-
logical components. (The latter view is what Langacker says is based
on the building-block metaphor, see the previous section.) The mor-
phological components motivate, or highlight, certain aspects of the
composite structure (Langacker 1987a: 453,1997a: 11).
THROW/kast
secondary
activations
THREW/kasta
primary
activation
Ο
/\
A
Figure 1. Kasta 'threw'
Figure 1 is meant to depict this viewpoint with the Norwegian verb

form kasta, the simple past of kaste "throw", as the example. Since
kasta is a relatively frequent verb form, let us assume that the word
as a whole is activated as a unit. Additionally, its components kast
(the stem) and -a (the past tense suffix) are activated. Kasta is not
seen as built up by kast- and -a; its composition is rather seen as a
case of primary and secondary activation.
There are several types of common morphological phenomena

that the building-block model unnecessarily makes into problems, or
overlooks, including the following:
• Phonological2 residues (cranberry morphs)

• Morphological ambiguity (overlapping items)
• Submorphemic relations (an item is morphologically
complex without having constituents)
These problems will be discussed in the following paragraphs.
1.2.1. Phonological"residues"
One type of problem is phonological residues, so-called cranberry

morphs, residual morphs, or unique morphs, such as Tues- in
Tuesday or the Norwegian examples in (4).
(4) durkdreven 'cunning' deüig 'delightful, delicious'

(dreven = 'skilled') (-ig is an adjectival derivational suffix)
uhyre 'monster' rädere 'etch'
(u- is a negative (-er is a verbal derivational suffix,
derivational prefix) see §2; -e is the infinitive suffix)
Phonological residues is the more apt name in the sense that these
items are not really morphs. But "residues" may suggest either that
they are leftovers when we have found the building blocks of the
synchronic analysis, or that they always are historical residues of
what once were meaningful components. The latter is not entirely
correct, since cranberry morphs may also come into being in other
ways. This is at least the case when some other part of the word is the
result of analogical reinterpretation (misleadingly called "folk
etymology"), see Gundersen (1995). An example is when many
Norwegians perceive the otherwise unanalysable noun f0ljetong
'serial' (< French feuilleion) to contain the verb f0lge, or f0lgje,
'follow', thus making tong a "residue". (The variant f0lgje is
102 Helge Gundersen
pronounced in the same way asffilje- inf0ljetong.) The polysyllabic,

opaque, and relatively infrequent word as a whole, f0ljetong,
becomes more cognitively functional by being "anchored" to a more
entrenched item, f0lge. The fact that cranberry morphs may be
created in the process of making words more functional, suggests that
these items are more of a problem for theoretical linguists than for
language users.
The solution offered in a morpheme-based model has been to
extend the definition of morphemes to cover residues that are
combined with recurrent morphological constituents, such as berry in
cranberry. On the basis of the fact that a cranberry is a specific kind
of berry, Bloomfield (1935: 160), in a subordinate clause ("since its
meaning is constant"), takes the meaningfulness of cran- for granted.
Hockett states (1958: 126), for the same reason, that "cran- clearly
carries a meaning". Bauer (1988: 40-41) explains: 'The meaning
associated with a unique morph is determined by subtracting the
meanings associated with the known morphs in the construction from
the meaning of the construction as a whole. So the meaning
associated with cran is precisely what makes cranberries a subset of
all berries." Aronoff (1976: 10) succinctly points out that "there is no
noncircular way to assign meanings" to unique morphs. He makes
use of this observation as one of several arguments for not defining
the morpheme in terms of meaning.
However, neither ascribing meaning to the "residuals" nor making
this a case against a semantic understanding of the minimal
morphological component is a necessity in a morphological theory.
Let us take a look at the Norwegian noun shredder 'tailor', as
depicted in Figure 2, a parallel to Figure 1.
Cognitive grammar enables one to say that the derivational suffix
-er makes explicit some of the meaning of skredder, and in this way
motivating it, while skredd- does not. What makes skredder a subset
of all the things designated by the suffix -er, is already included in
the meaning of the composite structure as a whole, which is primarily
activated. As for the phonological pole of the word skredder, the
suffix -er is certainly dependent on the stem skredd-, so that skredd-
contributes phonological content for -er to "work on" (see §2.3 for
this kind of analysis). As for the semantic pole, however, it is not

necessary to say that skredd- contributes anything, as this is
unproblematic from the standpoint of a network conception of
morphology (see also Bybee 1988: 128). Skredder has lexical con-
nections to the morph -er (specifically, skredder is, in Langackerian
terms, an elaboration of the schema [NOMEN AGENTIS/...er]), but
we are not required to ascribe meaning to skredd-, which remains a
part of skredder as a whole without having connections to other
lexical items.
NOMEN AGENTISAer
secondary
activation
TAILOR/skredder
primary
activation
Figure 2. Skredder 'tailor'

104 Helge Gundersen
The premise for this position is that "the composite semantic

structure has an essential kind of priority vis-ä-vis the component
structures" (Langacker 1997a: 11). The word may secondarily
contain one or more components which support the meaning of the
whole, highlighting certain aspects of it. Among the considerations
underlying this view is a notion of the language user associating
elements with each other (finding similarities and differences) in
accordance with general cognitive principles. In the building-block
model, on the other hand, grammar operates independently of general
cognitive principles, and the parts of a composite expression are
objects stacked on top of one another, with the morphemes making
up the bottom foundation.
1.2.2. Overlapping items
A rigid interpretation of the building-block model excludes, or makes

problematic, the possibility of ambiguous or overlapping analyses, in
which one part of the word belongs to more than one constituent
simultaneously. Well-formed tree structures are desirable in the
building-block model.
An example is the simple past of the Norwegian verb sende 'send'
in some western dialects and in the Nynorsk variety of written
Norwegian; the verb is inflected and pronounced as in (5).
(5) Infinitive: senda /senda/ (-a is the infinitive suffix)

Present: sender /sender/ (-er is the present tense suffix)
Past: sende /sende/ (-de is evidently the past tense suffix,
as in levde, ringde, and many other word-forms)
The speech sound /d/ in the past, and actually also the letter <d>,
may be seen as having a double function: as a part of the stem send,
and as a part of the past tense suffix -de. In a generative analysis, this
example would be handled by a degemination rule which converts an
underlying d + d into one surface d. This analysis involves constructs
not allowed in cognitive grammar, but the point is that cognitive
grammar offers alternatives. For phonetic reasons, only one /d/ is

pronounced, but this does not preclude a morphological analysis in
which sende has connections to both the stem send /send/ and the
suffix -de /de/. The word as a whole is primary to its constituents,
and segmentation into unambiguous components is not a necessary
requirement. Letting the components overlap falls naturally into the
framework without it being a "slackening" of the model in order to
accomodate some special type of occurrence. On the contrary, it
makes it possible to catch connections that would otherwise be
difficult to see.
In the realm of compounding we have a phenomenon that would
be very difficult to handle by rules (cf. Bauer 1983: 232-237, 293-
296), namely telescoped words (blends, clipping compounds,
portmanteau words), such as English slanguage, brunch, motel,
guestimate, infotainment, and bit (< binary digit', presumably, this
example is synchronically telescoped for some speakers only). A new
word is formed by combining the beginning of one word and the end
of another, and the two parts of the new word are overlapping, and/or
one or both parts are "incomplete". Slanguage has overlapping only,
while brunch has clipping only. In motel, both clipping and
overlapping are present, as may be seen in Figure 3.
motor hotel
motel
Figure 3. Motel
The compounding elements in the clipped words would not be

enough to represent the morphs in a morpheme-based model, despite
the fact that the speaker/hearer is dependent on the connections to the
full morphs to apprehend the full impact of the word. The language
user will recognize the "original" words even if their phonological/
orthographical expressions are incomplete. This can happen because
grammar is not something that is autonomously generated, but rather
106 Helge Gundersen
something that the language user does. An experiment by Cutler

(1980), see Prideaux (1984: 90), may be taken as further evidence for
the strategy. On the basis of adjectives like splendid, the subjects
tended to make verbs like splendify, on the analogy of liquid/liquify,
rather than splendidijy. The elimination of the final consonant (-d)
was less of a loss than to move the stress away from the base form
(splendid).
Telescoped words have a flair of something eccentric and funny
about them, seemingly not deserving much interest from the serious
grammarian. However, overlapping items may be posited in non-
compounding morphology as well, such as in sende above
(inflectional morphology) and in some of the derivational examples
in §2. Second, it may be argued, following Bauer (1983: 232), that
blends, clippings, and other unpredictable formations are too
common to be considered out of the ordinary, although they appear
as oddities in the context of generative grammar. Third, according to
the principles of cognitive linguistics a model should accomodate
any kind of occurrence in an equally "organic" way, illuminatingly
describing how it differs from, and is similar to, other kinds of
occurrences.
1.2.3. Submorphemic morphology
Both in Langacker's cognitive grammar and in Bybee's morphology,

the locus of morphology in grammar is the middle part of a
continuum from lexicon to syntax, in which a common denominator
is that sounds are used to symbolize meaning. This continuum of
symbolic units can be handled along several dimensions, such as how
specific (or schematic) the item is, how morphologically (or
syntactically) complex it is, to what degree it is an entrenched unit in
memory, how productive a symbolic pattern is, etc. These
considerations imply a gradient view of morphological structure. In
the building-block model, morphology is seen as segmentation into
separate building blocks. If such a segmentation is unworkable, there
is no morphological relationship at all. We would then have either
fully segmentable words or fully simple words. This view is

expressed rather clearly by Hockett (1954): "A linguistic form is
either simple or composite" (with reference to item and arrangement),
and "[a] linguistic form is either simple or derived" (with reference to
item and process). This stance may be viewed in the context of
similarity and identity (see also Enger 1998a: 140-152 with respect to
inflectional morphology). In the tradition I am criticizing similarity is
expressed by the presence of identical elements and is a matter of
either-or. Cognitive linguists, on the other hand, can draw on the
ability of the language user to observe similarities and differences
between elements, and they can thus make use of degrees of
relatedness.
There are no clear boundaries between different types of
morphological structure, but in Table 1, six groups are selected with
the aim of arguing that words may often have symbolic complexity
without being easily segmentable into discrete components. In
particular, it will be argued that the two "submorphemic" types D
and E are not entirely morphologically simple, despite the fact that
they do not contain traditional morphemic component structures.
Table 1. Some types of word-structure with varying degrees of complexity

A Fully segmentable words (words [un][work][able], [black][board],
analysable into symbolic segments) [bank] [er]
B Partially segmentable words (words Tues[day], possible], butch [er]
with in part symbolic components, in
part purely phonological portions)
C Inflections/derivations with modified sit-sat, begin-began, etc.
stems/bases, instantiating a pattern relief-relieve, strife-strive, etc.
D Words with root-forming morphs glare, gleam, glimmer, glow, etc.
(roots with meaningful portions) sneer, sneeze, sniff, snort, etc.
E Morphologically related wholes east-west, north-south
(semantically and phonologically here-there
similar words) male-female
F Morphologically simple words hair, town, tree
108 Helge Gundersen
The basis for the grouping is to be found in more than one

dimension, and it is not my claim that all words may be put
unambiguously into one place in the table. The words at the bottom,
in F, are morphologically simple, and, roughly speaking, the higher a
word is found in the table, the more easily it is viewed as
morphologically complex. Type F and type B (see §1.2.1) will not be
treated further, nor will type C in any detail (although fully irregular
alternations will be discussed below). It is well-known that this and
other non-agglutinative phenomena are notoriously difficult to handle
convincingly in a morpheme-based grammar (see especially
Matthews 1972). Numerous devices have been launched, including
the rather large apparatus of nonconcatenative morphology (e.g.,
McCarthy 1982), which Matthews (1993: 110) finds hard to see as
"much more than an elaboration, in modern dress, of part of [Zellig]
Harris's account of discontinuous morphemes". It is possible for a
strict building-block approach to analyse quite a few non-
agglutinative phenomena, but at the cost of introducing the extra
machinery in, e.g., nonconcatenative morphology.
Type A may be termed the ideal type, consisting of inflection and
derivation by means of affixes (prefixes and suffixes are certainly the
best representatives), in addition to compounding. The conceptual
structure is symbolized by the phonological structure in a neat way.
However, this type is not that "simple". One reason is that some
words are perceived as more transparent, or analysable, than others
(unworkable is probably the most analysable of the words in the
table). A novel expression will be fully analysable, but when the
word becomes more entrenched, the linguist may be in doubt whether
to segment it. These matters will not be elaborated further here (see,
e.g., Langacker 1987a: 457-466). However, this gradient ana-
lysability should be incorporated in a grammatical model in a
consistent manner, with independently motivated constructs.
As regards affixes, I shall hold that they may show submorphemic
characteristics; that is, they may display parallel phonological and se-
mantic connections below the level of the morph. In the Nynorsk va-
riety, many nouns (many feminines and some masculines) end in -er
and -ene in the plural, while many others (a large minority of
feminines and most masculines) end in -ar and -ane. When ending
in -enel-ane they are at the same time definite, while the word-forms
in -erl-ar are not inflected for definiteness. The lexemes vegg 'wall'
and hest 'horse' (both masculine) thus go like this in the plural:
(6) vegger 'walls' hestar 'horses'

veggene 'the walls' hestane 'the horses'
As Endresen (1982: 105-106) points out, the -e in -er may easily be

perceived as the same as the first -e in -ene, and the -a in -ar as the
same as the one in -ane. The endings -enel-ane have been viewed in
two ways: The Norwegian Reference Grammar (Faarlund, Lie, and
Vannebo 1997: 161, 193) sees them as a sequence of the plural -er/
-ar (in a form without the -r) and a definite suffix -ne. The
segmentation is then vegg-e-ne, hest-a-ne. The other view (cf.
Vangsnes 1999) is to see -ene and -ane as morphs with two
functions; "plural" and "definite". (The exact motivations for the two
analyses may vary between scholars.) It is tempting to suggest that
both views have their advantages; the analyses vegg-e-ne and hest-a-
ne appear counter-intuitive and exaggerate the segmentation (the
missing -r is possibly also problematic), while they still catch a con-
nection between -er and -ene, and -ar and -ane, respectively. A per-
missible analysis in cognitive morphology is to regard -er and -ene
(and -ar and -ane) as two morphs (suffixes) indeed, but to ascribe a
secondary morphological connection between them based on
semantic commonality (plural) and phonological similarity (both
begin with an -e, or an -a). This move is allowed in a model which
deals with network relations and which does not equate analysis with
segmentation.
Turning to type D, we have a situation where one portion
(traditionally called a phonestheme) of the root is meaningful, but,
putatively, still not a discrete component that takes part in the
hierarchical organization of the word. A number of linguists, such as
Bloomfield (1935: 245) and Rhodes and Lawler (1981), have seen a
gl- in English pertaining to the field of light, occurring initially in a
number of words, such as glare, glimmer, and glow. English words
110 Helge Gundersen
like squash, squeeze, and squelch have in common both the squ- and
the meaning "compressed". Many of those doing research in the
field, sometimes with the aid of experimental evidence (e.g. Abelin
1999), consider phonesthemes fairly widespread. For example,
Rhodes and Lawler (1981: 328) believe that more than 50 per cent of
English monosyllabic words are analysable, and Abelin found that 36
of the 37 word-initial consonant clusters in Swedish may function as
phonesthemes.
The phonestheme, such as g/-, may be viewed as "halfway"
separated from the rest of the word, simultaneously being a
meaningful unit and a part of the root. It is perhaps relevant that the
remaining parts of the words (-are, -immer, etc.) can be very
dissimilar from each other phonologically, so that it is the
phonestheme alone that symbolizes the shared semantics in the group
of words, in the case of gl- something concerning light. In this way,
these units are similar to affixes. As Rhodes (1994: 286-287) points
out, phonesthemes are often allotted a special status below the
morphological level, but Rhodes and Lawler (1981: 326) and Rhodes
(1994: 288-290) argue for a case of derivational morphology, with
the word being either wholly or partially analysable. We may follow
Bloomfield (1935) and call them "root-forming morphs", thereby
indicating that they are minimal symbolic (morphological) units and,
at the same time, part of the root. (The latter point implies that they
may not have all the characteristics of standard morphs, and that the
words in which they are found may not be as easily segmentable as in
"type A".) Because the building-block model identifies morphology
with exhaustive segmentation into morphs, as conventionally
understood, the whole phenomenon then disappears from
morphology. In a cognitive-linguistic network conception, any
instance of phonological and semantic connections that are
recognized as running in parallel in words, may be treated as a case
of morphology. This does not mean that cognitive morphology
cannot make use of morphs; Langacker does, and defines the morph
("morpheme" in his terminology) as an "expression of minimal
symbolic complexity, i.e. one not analyzable into smaller symbolic
components" (1991: 550). But it is not a decisive question whether
we can single out morphs in particular cases. Bybee (1995b: 233)

characterizes morphemes as epiphenomenal or "emergent". (See also
Langacker 1987a: 400-401; Bybee 1988: 128-129, and Davis 1992
for discussions of submorphemic units in relation to morphological
segmentation.)
In type E we have a different situation than in D. Certain words
may have a morphological relationship to one another that cannot be
ascribed to any specific part of the whole. It is the words as wholes
that are morphologically related. There are phonological and
semantic connections running in parallel, but the phonological
connections are in the form of a similarity relation, not an identity
relation. Examples may be English east and west, north and south,
which are most likely perceived by the language user as similar, both
semantically and phonologically. (The pairs are discussed by
Matthews 1991: 18, while Plank 1981: 187-189 discusses the
German counterparts.) From the standpoint of cognitive linguistics, it
is natural to draw attention to the fact that all the words are
monosyllabic (prosodic factors may be included in a morphological
relation), that the two members of each pair are similar (not identical)
in their CV-structures, and, notably, that both end in -stl-th. It is not
difficult to concur with the opinion that it is very doubtful to see -st
and -th as suffixes, but it is still possible to regard the words as
morphologically related. The Norwegian words her /hae:r/ 'here' and
der /dar.r/ 'there' illustrate the point even more clearly, since der has
been changed from dar, in my opinion largely because of analogical
influence from her (Gundersen 1995: 68, 132), while no sane linguist
(including me) would even consider positing the part -er as a morph.
As shown in Figure 4, these two words start with one consonant and
continue with /se:r/. A broken line indicates shared features, while the
solid lines stand for identity relations. The present model allows us to
consider the relation between her and der morphological. The
alternative in a strict building-block approach would be to ignore the
contribution of the phonological similarity altogether, and to treat her
and der as fully on a par with purely semantically related words like
floor and ceiling.
112 Helge Gundersen
her
ι
der
Figure 4. Her 'here' and der 'there'
The relationship between the two words these and those as wholes
looks like a clear case of type E: The two words are monosyllabic,
and both the th- and the -s- appear in the same positions. The vowel
is long in these, and a diphtong in those, making up a similarity
relation (just like the initial consonants in her and der). The word this
is also morphologically related to these words, especially these. As
for the th- by itself, this portion of the words perhaps rather points to
type D, but bordering on type E; the this tied up to a very limited
set of words with schematic meanings (also including they, that, and
others), but they may nevertheless form the basis of a schema
(pattern). In this way th- would be similar to a phonestheme, thus a
root-forming morph (see also Bloomfield 1935: 244). A clearer
candi-date may be the wh- that appears initially in most English
interro-gatives (what, where, etc.). The English demonstratives
display in this way the gradual transition existing between words
with root-forming morphs (type D) and singular relations between
whole words (type E).
Pairs, or small groups, of words with close-knit semantic relations
frequently resemble each other phonologically, and occurrences like
her and der, east and west are examples of that phenomenon. Over
time, semantically related words may also achieve a phonological
similarity originally not present or become more similar than to begin
with, as has been noted in studies of analogical change (see, e.g.,
Osthoff 1878 and Winter 1969 on, mainly, numerals). Some other
types of words than those noted in the previous paragraphs are perso-
nal pronouns, auxiliary verbs, numerals, kinship terms, names of
months and the days of the week, basic colour terms, basic anto-
nymies like "up-down", "left-right", "heavy-light", "male-female",
and also idiosyncratic pairs of words like Norwegian ridder 'knight'
(from Low German, cf. English ride) and rytter 'rider' (originally
probably from Latin rutarius, ruptuarius 'highwayman', according to
Bokmalsordboka). We often have a singular relation between two or
three words, but at other times a morph-like portion of the words
seems to emerge.
The above will fall outside of what has traditionally been con-
sidered morphology. This cannot be said so easily about der-ivational
pairs like English pope and papal, which show alternations of the
base that do not follow a pattern. However, this phenomenon can
fruitfully be treated in line with the above discussion. (This may also
be said about the suggested analysis of the Norwegian suffixes -ene
and -one above, although these units are affixes, not roots.) While
-al is a morphological component (type A above), the alternation
between -o- and -a- is more open to debate. The alternation is not
regular, but pap still appears to be derived from pope. In a structura-
list or generative framework, the linguist is forced either to list the
two as unrelated forms in the lexicon, or to make use of some special
device, as in the form of Jackendoffs (1975) "redundancy rule" or
some later, equivalent development. Since the alternation is unique, a
"rule" like this will, fairly obviously, not tell us anything more than
listing the two forms in the lexicon as unrelated. (This actually
follows from Jackendoffs own formulations, cf. Carstairs-McCarthy
1992: 42, and it seems that devices like morphological redundancy
rules are intended to apply to occurrences of type C in Table 1.)
Rules will not work satisfactorily, and the phonological similarities
between the alternating forms are too apparent to be dismissed. If the
fundamental assumptions are revised, however, we can simply
assume that the items pope and pap- are related by shared semantics
and similar phonology, and take this to be a morphological statement.
The solution will be based on theoretical concepts needed
independently in the model. (See further treatment in §2.3.) This step
cannot fully be taken if we entertain the view that a grammatical
analysis largely consists of listing building blocks and stating how
these are combined. In fact, I think it would be fair to say that it is
necessary to adopt a view of morphology as a network in the mental
lexicon.
114 Helge Gundersen
Now, even when all the principles behind the discussion in the
previous paragraphs are accepted, it will still be possible to argue that
the term "morphology" is best reserved for "parallel phonological
and semantic connections, if they represent a pattern found in mul-
tiple sets of items", which is what Bybee (1995b: 232) holds, as a
cor-rection to statements in her earlier work. "If there were only one
pair of items with such connections this would not be regarded as
consti-tuting a morphological relation." The proposed connections
be-tween, say, Norwegian her and der, or ridder and rytter, would
then not belong to morphology, and this is most likely in accordance
with standard conceptions. As mentioned at the beginning of §1.2.3,
there is a continuum between lexicon and morphology along several
di-mensions, and we may see what belongs to what as partly a matter
of terminology. Nevertheless, the transition from a relation between
two singular items to a set of items conforming to a pattern is gradual
with room for considerable indeterminacy. In particular, morpholo-
gists are sometimes confronted with a doubt as to whether an inflec-
tional or derivational alternation instantiates a pattern or is a fully ir-
regular similarity relation (such as between pope and pap- in papal).
Thus, it seems profitable to be able to sort the study of parallel
phonological and semantic connections in singular pairs of items
under the heading "morphology", as long as we have to do with
phonologically different words. It is in any event desirable to make
explicit the continuum from complex to simple items. When the
cases in question are excluded from both morphology (because they
do not represent a pattern) and lexical semantics (because there are
differences in the phonological, not only semantic, structure of the
items), they may easily fall through the cracks without receiving
attention.
To sum up, there is no clear boundary between what ordinarily is
considered morphology on the one hand and submorphemic (yet
morphological) occurrences on the other. This is not a problem in a
cognitive-linguistic network approach, because segmentation is not
required. Morphology is under all circumstances seen as a network in
the mental lexicon, and a morphological viewpoint may easily be
adopted with respect to various "difficult" cases, including sub-

morphemic relations.
2. The Norwegian verbs in -ere
In §1.2 I argued that cognitive morphology does not suffer from the
shortcomings imposed by a model that invites to exhaustive (§1.2.1),
unambiguous (§1.2.2) segmentation into separable constituents
(§1.2.3). In this section, I shall exemplify my arguments further, but
this time from a different angle. Instead of going through types of
phenomena, and exemplifying them, I shall concentrate on a limited
area in the Norwegian language, and show how various segmentation
problems will crop up there. My intention is not to propose solutions
to all of these problems. My aim is rather to argue that a cognitive-
linguistic network conception of morphology partly opens up fields
of study necessarily ignored in a building-block framework, and
partly suggests ways out of segmentation problems that otherwise
seem inextricable.
The area in question is the synchronic derivational morphology of
a group of Norwegian verbs with stems ending in —er (-ere in the
infinitive). These make up about 1500 of the 4800 verbs in
Bokmalsordboka (cf. Bjerkan and Simonsen 1996: 193; see also
Endresen and Simonsen this volume). The verbal derivational
element in these words is the suffix -er, which also exists in
Swedish, Danish, Dutch, and (in the incarnation of -ier) in German.
A number of ere-verbs were borrowed as wholes into Norwegian
from Low German during the Hanseatic period (1200-1550), and
became productive (Leira 1992: 37-38, 147-148), although zero-
derivation of verbs occurs more frequently. As is often the case with
loans, irregularities arise in the course of the morphological
adaptation into the receiving language. (7) shows some well-behaved
introductory examples of the verbs, with the derivational bases to the
left. I cite the verbs in the infinitive, with the infinitive suffix -e
included at the end.
116 Helge Gundersen
(7) respekt 'respect' respektere 'respect'

steril 'sterile' sterilisere 'sterilize'
jazz 'jazz' jazziflsere 'jazzify'
-iser (as in sterilisere), and the longer -(i)fiser (as in jazziflsere) are
very often taken as variants of -er in Norwegian (cf. Leira 1992,
Johannessen 1993, Faarlund, Lie, and Vannebo 1997: 121; see also
Söderbergh 1967, Thorell 1981 for Swedish), -iser is the most
productive of these (Leira 1992: 38). The -er, including the part -er
in the longer variants -iser and -ifiser, is always stressed. The base
for -er-derivation as well as zero-derivation of verbs in Norwegian is
ordinarily a noun or an adjective. Both types of derivation (with -er
and zero) appear as ways of making verbs in general, not restricted to
certain verb meanings. (However, there is a strong tendency to prefer
the one over the other on phonological, and probably also semantic,
grounds, and the same may be said for the choice between the
variants of-er.)
There are a number of morphological problems on the boundary
between -er and what comes before it. One type of problem is where
we have one or more speech sounds before -er that are difficult to
account for, in that it is not clear which morphological components
they are parts of, if any at all. Verbs with three frequent terminations
will be used as examples: -tere, -isere and -ulere.
2.1. The termination -tere
The verb handtere 'handle' illustrates the problem nicely. Handtere,

or handtere (depending on the variety of Norwegian), stems from
French (it is related to English haunt), and is borrowed into
Scandinavian from Low German hanteren, which, according to
Lübben (1993), meant 'visit frequently', 'go hither and thither',
'trade' (the latter meaning originating in the traffic of pedlars). It is
associated with hand /hon/ or hand /han/ 'hand' in Norwegian, as
shown by the spelling with -d, and the fact that it may be pronounced
with an /o/. Here we may imagine the following possible seg-

mentations in modern Norwegian:
(8) a. händt-ere /hont-e:re/

b. händ-tere /hon-te:re/
c. both händt-ere /hont-e:re/ and händ-tere /hon-te:re/
d. hand-t-ere /hon-t-e:re/
In (8a), händt /hont/ 'hand' can be seen as a variant of hand /hon/

'hand'. This segmentation would most likely be considered the least
controversial, since it introduces neither any unorthodox suffix, nor
any residuals. However, the relationship between hand and händt
remains a bit problematic from a theoretical point of view, as these
lexical items are clearly related, but as singular items, not via any
rule. It is perhaps relevant that the pronunciation /hont/ for hand can
be found, although not in all dialects, in a few other loans from Low
German: händverk 'craft', händverker 'craftsman', and händlanger
'helper, assistant'.
In (8b), -ter is seen as the suffix. This solution would be
considered very original, since no one has ever posited a suffix -ter
in the Norwegian grammar. But there are in fact quite a few verb
stems ending in -ter, about 230 out of the 1500 ere-verbs in
Bokmalsordboka, and one might think that this massive number
could be an additional source of reanalysis. Instead of legitimasjon
'identification card' one can surprisingly often hear people say
legimitasjon, and, correspondingly, legimitere seg 'prove one's
identity' instead of legitimere seg (Gundersen 1995: 76). However,
one may convincingly argue that this is a case of purely phonological
analogy.
In (8c), we have an ambiguous structure; that is: (8a) and (8b) are
both valid, the speech sound I\J playing an ambiguous role.
In (8d), the /t/ is not part of the morphs hand and -er. This
solution retains hand as the base and -er as the suffix, but gives us
the -/- as something that, in a building-block framework, would be
considered a residue. It solves the problem of (8a), but evokes
another theoretically difficult point instead.
118 Helge Gundersen
In theory, all four options should be considered equally permiss-

ible and unproblematic, and it is an empirical question which one to
choose.
2.2. The termination -isere
Verbs ending in -isere include those with the variant suffixes -iser
and -ifiser, like the two last examples in (7), sterilisere and
jazzifisere.
-is- is a phonological string that may also appear in other
environments, such as in (9):
(9) organisere Organize' (approx. 300 -isere)

organisasjon Organization' (23 -isasjon)
Excluding compounds, 23 nouns in Bokmälsordboka end in -isasjon.

If the derivational suffix in organisere is -iser, how is the form
organisasjon to be accounted for? The morphological status of -is-
is therefore somewhat disputed. Some have regarded -is- as a purely
formal "concatenator" with the function of linking morphemes to one
another (cf. van Stek 1983: 51). This solution seems to be at odds
with the principles of cognitive grammar, in which grammatical
functions are described in semantic terms. One possible solution
within this framework, however (there may be other possibilities,
too), is to regard -iser and -ifiser indeed as variants of -er, but still
to retain a connection between the -is- appearing here and the -is-
appearing before -asjon in nouns like organisasjon, regardless of
whether —is- in the latter case belongs to a nominal suffix -isasjon,
or to the derivational base before -asjon. While speakers do seem to
analyse words into components (although these components do not
have the status of building blocks), this is only one facet of the
description, and morphological relations need not always follow the
"componentiality" of the words.
The string -is- may be seen as problematic in two more ways.
First, some of the verbs with -iser correspond to adjectives with the
derivational suffix -isk, to nouns with the derivational suffix -z, and
to nouns without any of these suffixes:
(10) psykologisere 'psychologize'

psykologisk 'psychological'
psykologi 'psychology'
psykolog 'psychologist'
The verb psykologisere corresponds to psykologisk, to psykologi, and

to psykolog. As depicted in Figure 5, we have a difficulty deciding on
whether we have psykologisk + -er 'make psychological' (without
the /k/), psykologi + -iser 'make into psychology' or psykolog + -
iser 'reason like a psychologist'. One possibility allowed by the
network model is that all (or two) of these analyses are valid at the
same time, i.e. that we have a case of simultaneous multiple analyses.
Another possibility is that each of the analyses is correct at different
times in different contexts.
p s y k o l o g i s e r e 'makepsychological'
//\\
-isk -er
p s y k o l o g i s e r e 'make into psychology'
-ι -iser
p s y k o l o g i s e r e 'reason like a psychologist'
-iser
Figure 5. Interpretations of psykologisere

120 Helge Gundersen
Another problem is that -er seemingly is a part of -iser, while -ifiser

may be seen as an extension of -iser. Figure 6 illustrates the formal
similarities. If these items are seen as three more or less equivalent
suffixes, it is hard to believe that their phonological similarity is
strictly coincidental, unrelated to their semantic similarity from a
synchronic point of view. One hypothesis might be that there is a
morphological relationship between them that is below the
morphemic level. If this is so, they are three different morphs, but
each of them carries phonological similarities to the other two that
are connected to their semantic similarities.
-er
-iser
-ifiser
Figure 6. The suffixes -er, -iser, and -ifiser
2.3. The termination -ulere
Approximately 120 verbs in Bokmälsordboka end in -lere; 44 of

those end in -ulere. The sequence -ul- that appears in front of some
derivational suffixes, never appears at the end of independent words.
27 of almost 500 nouns with the suffix -asjon end in -ulasjon in
Bokmälsordboka, while 6 of the 71 adjectives ending in -ativ also
end in -ulativ. Some of the verbs in -ulere correspond to other
derivations, like spekulasjon and spekulativ, while others do not.
(11) spekulere 'speculate' (44 -ulere)

spekulasjon 'speculation' (27 -ulasjon)
spekulativ 'speculative' (6 -ulativ)
In the literature (cf. references in the introduction to §2), -ul- is not

treated as part of a suffix, but my question is: do language users in
some way connect -ul- to the suffix coming after it? One interesting
formation is spankulere 'strut', 'stroll', which, historically, is created
from the shorter (now obsolete) verb spanke 'strut', in accordance
with verbs ending in -ulere, such as spekulere 'speculate'. On the
other hand, if we say that -ul belongs to -er, how shall we then
handle the fact that -ul appears before other derivational suffixes,
such as before -asjon and -ativ in (11)? This type of problem was
discussed in section 2.2 with examples among verbs in -isere
(organisere, etc.).
If we do not see -uler as a suffix, we shall still have to deal with
some morphological problems concerning all those derivational bases
ending in -ul. Are they really—synchronically speaking—
coincidental, unrelated phonological elements, or do they perhaps
take part in a network of submorphemic relations within
morphology?
Furthermore, a handful of the verbs ending in -ulere bear both
phonological and semantic similarities to nouns ending in -el. Four
of them are common, and are shown in (12).
(12) fabel 'fable'

fabulere 'give one's imagination a free rein'
regel 'rule'
regulere 'regulate'
sirkel 'circle'
sirkulere 'circulate'
tittel 'title'
titulere 'address' (i.e. 'address somebody as...')
These correspondences seem to create a type of problem that we find

in many other ere-verbs, such as the relationship between the noun
kalkyle 'calculation, estimate' and the verb kalkulere 'calculate', and
between the noun applaus 'applause' and the verb applaudere
'applaud'. The first of the segmentations in §2.1 considered handt-
122 Helge Gundersen
in kandiere 'handle' a variant of hand 'hand', and this is another

instance of the phenomenon. Similar to kandiere are verbs like
flambere 'flambe, flame (a dish)', cf. the noun flamme 'flame', and
griljere 'grill', cf. the noun grill 'grill'. In §1.2.3 the equivalent
example was pope and papal. If we posit that the semantic and
phonological connections are perceived as running in parallel in these
pairs of items, we may view that as morphology. This type of
morphology, however, cannot be captured in a general statement,
such as by means of rules.
In the case of the verbs in (12), there is still a certain possibility
that speakers have extracted a schema (pattern, "rule"), so that we
should have a case of what I labelled type C in Table 1, equivalent to
belief-believe, calf-calve. If so, this case would be accounted for in
cognitive grammar with analytic tools conforming to general
cognitive principles, such as Schemas and comparison (in a similar,
but not quite identical, way as in Figures 8 and 9 below). In a
building-block approach, I assume that special devices (considered
artifactual in cognitive grammar) like the tiers of nonconcatenative
morphology might come into play. For unique correspondences like
those in the other examples above (kalkulere, etc.), such devices
would not be feasible in any event. Unique alternations concern
submorphemic phenomena of what I labelled type E in Table 1. If we
choose to see the correspondences in (12) as unique, sirkul in
sirkulere is neither the noun sirkel in some unclear sort of disguise,
nor completely unrelated to sirkel. Instead it may be viewed as a
unique item that is morphologically related to sirkel below the level
of constituency. A simple diagram in line with Bybee's morphology
may look like Figure 7.
'circle' s i r keI
ι
I
I
'circle' s i r ku l e re
Figure 7. Sirkel and sirkulere

There are lexical connections between the semantics of the

expressions and between their phonological structures, and these run
in parallel. The broken line indicates similarity (both phonemes are
vocals), the others phonemic identity. Likewise, connections could
have been drawn from -er to other verbs with the same derivation. In
Langacker's more comprehensive cognitive grammar, the notation
will be more complex, taking more factors into consideration. There
are several ways to go about it, but some central points will be noted
here. For the sake of simplicity, I shall restrict myself to the stem
sirkuler.
The composition of the phonological pole of the expression (only
the relevant aspects) is shown in Figure 8. For the sake of
convenience, orthographical representation is employed.
•
sirkel •
hΨ sirkul ^ '4ffffl> er
<%%$%
Figure 8. The phonological pole of sirkuler
Note that sirkuler is an entrenched unit in grammar in its own right

(it is not a novel expression), but that does not prevent it from having
component structures, specifically the derivational suffix and the
base. More important, perhaps, than this constituency are the
horizontal relations, i.e., the dependency. Within the base sirkul is
124 Helge Gundersen
compared (by the speaker/hearer) to sirkel. These two units share

some phonological content, at the same time as they deviate from
each other. Sirkul only partly instantiates sirkel (hence the broken
categorization arrow). The suffix -er, or ...er, is seen as having some
schematic phonological content, indicated by the three dots. (This
move is allowed in a model that regards phonological structure as
conceptual.) Sirkul elaborates, and fully instantiates, this so-called
elaboration site (cross-hatched).
The central points of the composition of the semantic pole is
shown in Figure 9 below, but will not be explained in detail (see
Langacker this volume for an explanation of trajector [tr] and
landmark [1m]). The box at the top shows the relevant aspects of the
composite expression ('circulate'), in which one and the same entity
(hence the dotted lines) is moving around in something akin to a
circular path.
We may note that sirkul would have been a purely phonological
"residue" if it were not for the similarity relation to the word sirkel.
This relation gives sirkul some lexical content, shown as a circle at
the bottom left. (It is likely that the "circle"-aspect of the meaning of
sirkuler will be more weakly activated than if the word sirkel had
been present directly in sirkuler. This difference is not shown in this
diagram.) This meaning ('circle') integrates with the meaning of -er
to form the meaning of the composite expression. (However, there is
unquestionably more to the meaning of sirkuler than what is
activated through the connections to its component structures; these
merely motivate aspects of the meaning of the whole.) If we assume
that -er is a general verbal suffix, its meaning may roughly be termed
'process', which is shorthand for the meaning of verbs in cognitive
grammar (see Langacker this volume). The composite expression,
sirkuler, inherits this processual meaning from -er, in a manner that
will not be explicated here. However, another way of presenting this
relation, and the equivalent relation at the phonological pole, is
shown in Figure 9, with the semantics in capitals:
[[PROCESS]/[...er]] -> [[SIRKULER]/[sirkuler]]
tr
Im
Figure 9. The semantic pole of the stem sirkuler
Sirkuler is here categorized by the more schematic unit -er. A

derivational suffix may be seen as a schematic word (or stem), and
the more specific word, in this case the stem sirkuler, fully
instantiates, or elaborates, this schematic word. It does that
semantically by having a more specific meaning that is compatible
with the meaning of the suffix, and phonologically by having a more
specific phonology (notationally by filling in the three dots).
126 Helge Gundersen
3. Summary and conclusions
This paper has shown that cognitive grammar, or the cognitive-

linguistic framework, has something to contribute to morphological
analysis, including the seemingly problematic cases.
I have contrasted two very different conceptions: In the building-
block approach, segmentation into discrete components is a defining
characteristic of morphology. These components, and the means to
assemble them, are conceived of as being independent of the
language user. The cognitivist approach sees grammar as a reservoir
of linguistic units structured in a mental network and conforming to
general cognitive principles. More specifically, I have made use of a
strictly symbolic, bipolar view of morphology, in which the
discipline becomes the study of the relation between sounds and
meaning in (complex) words. Humans employ speech sounds to
symbolize meaning, and morphological structure constitutes an often
quite complex web of parallel phonological and semantic
connections. In this picture, the segmentation of words into separate
components is merely one dimension, important as it may be. I do
believe that it is psychologically valid to say that the English word
falsehood may be divided into two components: false and hood. But
firstly, as argued in §1.1 (residuals) and §1.2 (overlapping items), the
status of such components is not that of building blocks. Rather, they
are secondary to the word as a whole, and can only be singled out on
the basis of relations to other items in the web that makes up
morphology. Secondly, as discussed in §1.3, a language may contain
symbolic structures that are below the level of the morph (as this is
commonly conceived), and cognitive morphology allows us to
describe these structures in a coherent manner.
Although the "component status" of portions of words—to what
degree the language user processes these portions as separate
morphological components—is an important aspect of morphology,
there is more to the morphological structure in languages like English
and Norwegian than the building-block model implies. One example
is the so-called residual "morphemes", which are a normal part of
word structure, as opposed to what the awkward evoking of
cranberries suggests. Another example is the meaningful sub-

morphemic relations by which many words are knit together, also in
conjunction with inflection and derivation. It should indeed be
possible to account much more precisely for those relations, given
that the requisite fundamental assumptions are in place. Fully
analysable words may well represent the morphological ideal type,
but linguists should strive for a description covering all facets of
symbolic structures in words. When a type of structure is covered, it
should be laid down on the linguistic map in an illuminating way,
without resort to extra theoretical machinery. I hope to have
demonstrated that models like cognitive grammar offer a framework
that has room for an undertaking of that type.
Notes
1. I wish to thank Rolf Theil Endresen and Hanne Gram Simonsen for their
support and for their invaluable help in the improvement of this paper. The
paper has also benefitted greatly from the input of two anonymous referees.
2. Admittedly, this paper falls into the ubiquitous "phono-centric" view in
linguistics to the extent that phonology receives most of the attention among
the means to express meaning. However, there is nothing in the framework
presented that precludes an inclusion of orthography and paralinguistic
features, although I am not sure that the relationship between speaking, writing,
and non-verbal elements is a straightforward and obvious one.
Past tense acquisition in Norwegian:
Experimental evidence
Hanne Gram Simonsen
This paper reports an experimental study of Norwegian children's

acquisition of past tense (PT) forms. Using a method similar to that
ofBybee and Slobin (1982), past tense forms were elicited from three
groups of children, at 4, 6, and 8 years, with approximately 30
children in each age group. Each child was tested individually on 60
verbs, half of them strong and half of them weak. The verbs were
carefully chosen according to type frequency (number of verbs
belonging to each inflectional class), token frequency (frequency of
use of each verb), and phonological properties of the classes.
The results indicate that input factors, such as phonological
properties of the verbs and their type and token frequencies, play an
important role in the acquisition of Norwegian PT forms, both
regarding correct performance and regarding the errors children
make, and for all verb classes. Children generalise over all the main
inflectional patterns, but to a varying degree, and the error patterns
change over time. The differences between verb classes seem to be of
a gradual nature, and not a question of either/or, thus lending
support to models assuming that the mechanisms involved are not
dichotomous, but rather of a homogeneous nature.1
1. Introduction
In this article the results from an experiment concerning past tense

acquisition in Norwegian are presented and discussed in the light of
different models for acquisition. The experiment was constructed by
Hrafnhildur Ragnarsdottir, Kim Plunkett and the present author to
test and compare past tense acquisition in Icelandic and Norwegian
children, and a full description of the whole experiment with its
130 Hanne Gram Simonsen
results are presented in the main report (Ragnarsdottir, Simonsen,

and Plunkett 1999). However, since many of the articles in the
present volume are closely connected with and refer directly to this
experiment (Bjerkan; Matcovich; N0klestad; all this volume), the
Norwegian data from the experiment are presented again here. Also
presented are some additional results for Norwegian which were not
included in the above mentioned comparison with Icelandic.
The past tense, with its combination of regular and irregular
patterns, provides a particularly instructive window into how
children gradually master the morphology of their language, and into
how morphological patterns are stored and processed. Thus, the
question of how children acquire the past tense of verbs has been
discussed within different research paradigms and on the basis on
different types of data over the last 50 years, with an overwhelming
majority of studies based on English-speaking children. For example,
there are naturalistic studies based on longitudinal data bases,
(Brown and Bellugi 1964; Brown 1973; Marcus et al. 1992),
experimental studies based on nonsense forms (Berko 1958; Derwing
and Baker 1986) or real forms (Bybee and Slobin 1982; Marchman
1997), simulations of past tense learning in connectionist nets
(Rumelhart and McClelland 1986; Plunkett and Marchman 1991,
1993, 1996), and PET imaging studies addressing the same questions
in adult processing (Jaeger et al.1996)—to mention a few. Although
the bias towards English has changed somewhat with more cross-
linguistic work during the last 10-15 years, the perspective has
remained relatively English-centered. This seems also to have
influenced some of the models put forward to account for the
acquisition process.
1.1. Models for past tense acquisition and processing
There are currently two main models which aim to explain the
acquisition and processing of past tense, which may be called the
dual mechanism account and the single mechanism account.
Past tense acquisition in Norwegian 131
The dual mechanism account is related to a generative or

symbolic framework (cf. Pinker and Prince 1988; Marcus et al. 1992,
1995; Prasada and Pinker 1993). Here, the basic assumption is that
there are minimally two different mechanisms at work: on the one
hand a rule-based mechanism which operates on an all-or-none basis,
taking care of the regular (weak) inflections including over-
generalisations of them; and on the other hand a lexical retrieval
mechanism which is sensitive to input factors like frequency and
phonological factors, and which is responsible for all irregular
(strong) forms. The retrieval of an irregular form from the lexicon
blocks the application of the general rule. The later developments of
this model allow for a second split in the lexical retrieval mechanism,
where the most suppletive forms are listed separately, but where an
associative memory component takes care of all subregularities and
similarities among the irregular verbs. However, the basic,
dichotomous split is still there—between the "rule-governed"
regulars and the "exceptional" irregulars, treated by two qualitatively
different mechanisms.
The single mechanism account is related to a connectionist
framework (Rumelhart and McClelland 1986; Plunkett and
Marchman 1991, 1993, 1996), and an analogical (Skousen 1989,
1992; Derwing and Skousen 1989) or cognitive linguistics approach
(Langacker 1987a; Bybee 1988, 1995a). This account assumes that
there is only one mechanism at work, a mechanism which is sensitive
to input factors like type and token frequencies of verb forms, and to
their phonological properties. All past tense forms (with the
connection to their base forms) are memorised to a certain extent.
But given a certain amount of phonologically similar forms,
generalisations are formed across these forms.
Thus, the dual mechanism account assumes that there is one
default rule which is applied to weak verbs, while the strong verbs
are handled by a qualitatively different, more rote-like mechanism. In
the processing of a strong verb, if the mechanism does not manage to
retrieve a strong form from the lexicon, there is nothing to block the
general rule from applying, and we get errors of overgeneralisation of
the default weak inflection, like goed and silled instead of the correct
went and sat. Such overgeneralisations of weak verb inflections

should not be influenced by input factors—nor should the correct
acquisition of weak verbs. On the other hand, since the strong verbs
are handled by a different mechanism, overgeneralisations of strong
inflections are expected to be more rare, and to apply only where
phonological similarity and frequency factors allow for them.
The single mechanism account, on the other hand, holds that the
same mechanism regulates both strong and weak verbs. This implies
that overgeneralisation errors take place in essentially the same way
both for weak and strong inflectional patterns, once there are enough
phonologically similar forms to generalise over. It also implies that
input factors will be important for both acquisition and over-
generalisations of all kinds of verbs.
1.2. Norwegian verb morphology as compared to English
An outline of Norwegian verb morphology is given in Endresen and

Simonsen (this volume). Here, only a reminder will be given of a few
interesting points where Norwegian verb inflection differs from that
of English. The variety of Norwegian described in this paper is the
dialect spoken in and around the capital of Oslo, a variety close to the
Bokmäl standard. To a large extent it also covers other dialects
spoken in Eastern Norway.
One interesting difference between Norwegian and English is that
the former has two distinctive word accents or tones. Word tone is
usually lexically contrastive, but there are also morphological tone
effects: in verbal inflection, word tone in the present tense is a
reliable cue as to whether the verb is strong or weak. With a few
exceptions, strong verbs have a Low Word Tone in the present tense,
while weak verbs have a High-Low Word Tone in the present tense
(while the infinitive has a High-Low Word Tone in both weak and
strong verbs). For example, the strong verb sitte /site/ 'sit'with the
past tense form satt /sät/ 'sat' has sitter /siter/ in the present tense,
while the weak verb titte /tite/ 'look' with tittet /titet/ 'looked' in the
past tense has titter /titer/ in the present tense.
The basic split between weak and strong verbs found in most
Germanic languages is similar in Norwegian and English. However,
as illustrated in Figure 1, Norwegian has a slightly more complex
inflection than English, with 6 word forms in the weak paradigm as
compared to 4 in English.
And in addition there is a difference in the number of weak
inflectional classes. English has only one weak inflectional class,
while Norwegian has two, one of which (represented with the verb
kaste 'throw') is more productive than the other (represented with the
verb rope 'call'). Thus, there is a less clearcut division into two
different groups of verbs ("regular" vs. "irregular") in Norwegian
than in English.
ENGLISH One single weak inflectional class

stem (= inf./pres./imp.) dance call
-ing-form dancing calling
-5-form (3.p. sg. pres.) dances calls
-erf-form (past/past part.) danced called
NORWEGIAN Two weak inflectional classes

stem (= imperative) kast( 'throw') rop('call')
infinitive kaste rope
present kaster roper
past kästet ropte
past participle kästet ropt
pres. participle kastende ropende
Figure 1. Weak verb paradigms of English and Norwegian
This way, the testing of verbal inflection in Norwegian as compared

to English may provide a new angle to the understanding of past
tense acquisition, and possibly shed new light on the suitability of the
different models mentioned above.
2. The experiment
To test the possible effect of input factors on acquisition, parallel

experiments were designed for Icelandic and Norwegian, eliciting
past tense forms of cognate verbs in these languages. A full
description of this study including the details of the statistical
analyses is found in the main report (Ragnarsdottir, Simonsen and
Plunkett 1999). As mentioned above only the Norwegian data from
the experiment will be reported and discussed here.
2.1, Elicitation task
The elicitation task follows the pattern used in Berko (1958), Bybee
and Slobin (1982), where the child is shown a picture of someone
performing an action. The experimenter says: "Here is a boy who
knows how to VERB. He is VERBing. He did the same thing yesterday.
What did he do yesterday?" And the child fills in (presumably with
the past tense form of the verb): "He VERBed (or VARB)." Since the
word tone in the present tense is a reliable indicator as to whether the
verb has a strong or a weak past tense inflection, approximately half
of the children were given the infinitive as a cue (e.g., "Her ser du en
gutt LIGGE i sengen" = Here you see a boy LIE in the bed), while the
other half were given the present tense as a cue (e.g., "Her ser du en
gutt som LIGGER i sengen" = Here you see a boy who LIES in the
bed). (The actual phrasing of the rest of the cue in Norwegian
corresponds closely to the English version.)
2.2. Subjects
Three groups of children were tested, with approximately 30 children

in each group, at 4, 6, and 8 years of age, half of them boys, half of
them girls. The children were recruited from preschools and schools
in middle class areas in Oslo, and chosen from families where both
parents spoke an East Norwegian dialect.
Table 1. Subjects
Subjects n Mean Age SD
4-year-olds 28 4; 4. 4 0;3.5
6-year-olds 29 6; 3. 15 0; 2. 10
8-year-olds 39 8; 3. 0 0;1.27
2.3. Verbs
Each subject was tested individually, on 60 verbs, 27 strong and 33

weak, presented in a single random order. (To control for fatigue
during testing, half of the children were presented with the last thirty
verbs first. Before testing, the children were presented with a few
warm-up verbs to check that they understood the test and that they
managed past tense inflection.)
The verbs in the test were carefully chosen, taking into account
input properties assumed to influence acquisition—both frequency
factors and phonological factors. Concerning frequency factors, both
type frequency (i.e., the number of verbs belonging to each verb
class) and token frequency (the frequency of use of each verb) were
taken into consideration. The three main verb classes in Norwegian
vary in type frequency, as illustrated in Figure 2 (cf. also Endresen
and Simonsen, this volume, section 4.2).
Type frequency
three main classes
Strong \\feak
Smaller Larger
0*8) (WL)
LowTyfi· MdTyft HghTyfi-

(4%) (40%) (56%)
Figure 2. Relative type frequency of the main verb classes in Norwegian
Verbs from all the three main classes were included in the test. For
the weak verbs, half were taken from the larger weak class (WL) and
half from the smaller weak class (WS). For the strong verbs,
subclasses and their relative type frequency were also taken into
account (cf. Endresen and Simonsen this volume, section 4.2.1): two
groups of verbs were chosen from two of the subclasses with a
relatively higher type frequency among the strong verbs (the drikke-
drakk class with 40 members, [SI], and \h& flyte-fl0t class with 27
members, [S2]). The third group of strong verbs was chosen as a
mixture of different subclasses with few members each (less than 9;
[S3]).
For each group of verbs, items were included to represent both
high token frequency and low token frequency in an approximately
equal amount, excepting items from the third subgroup of strong
verbs (S3), which all had a relatively high token frequency. (Good
token frequency measures are hard to find for Norwegian, in
particular for spoken language. See Ragnarsdottir, Simonsen, and
Plunkett 1999 for a description and discussion of the token frequency
counts used.)
As for phonological factors, both phonological coherence/
openness within each inflectional class and phonological similarity/
difference between different inflectional classes were taken into
account. Phonological coherence/openness is a measure of the degree
of phonological homogeneity of the members within each verb class
and as such an indicator of how predictive the phonological
properties of the infinitive or present tense in each class are of the
past tense form. The weak verb classes in Norwegian differ with
respect to phonological coherence/openness, in that the WL class is
more open than the WS class (cf. Endresen and Simonsen this
volume, section 4.2.2). The strong verbs also differ in this respect, as
exemplified by the number of different vowels in their infinitive
stems. The flyte-fl0t subclass is phonologically very coherent with
one single infinitive stem vowel (/y:/) for all verbs, while the drikke-
drakk subclass is less coherent with three different infinitive stem
vowels (III, /e/, or /y/).
Measures of phonological similarity/difference may for instance

be developed on the basis of whether the verbs from different classes
have infinitive stems which rhyme or not. In the test, this factor was
built in for the weak verbs through splitting each of the main weak
classes into two subgroups—one containing verbs rhyming with
strong verbs, and one containing verbs not rhyming with any strong
verbs.
The test contained the seven groups of verbs listed in Table 2.
Table 2. Verb classes included in the test

• 3 groups of strong verbs (9 verbs in each group):
1 SI (HiTyFr; LoPhoC) Nldlyl ~ /a/ sitte-satt
2 S2 (HiTyFr; HiPhoC) /y:/ ~ /0:/(/0y/) fryse-fr0(y)s
3 S3 (LoTyFr; LoPhoC) Strong verbs
of different, I0pe-I0p, le-lo
small classes
• 4 groups of weak verbs (8 verbs in groups 4-6; 9 verbs in group 7):
4 WLR (Larger weak class -/et/or-/a/ titte-tittet/titta
rhyming with strong)
5 WSR (Smaller weak class -/te/, -lad lyse-lyste, gre-gredde
rhyming with strong)
6 WLN (Larger weak class -/et/ or -/a/ kaste-kastet/kasta
not rhyming with strong)
7 WSN (Smaller weak class -/te/, -/de/ kj0re-kj0rte
not rhyming with strong)
Results were calculated on the basis of these seven verb groups, and
they will occasionally be referred to in the presentation below.
However, since only a few of the differences between these
subgroups turned out to be significant, a less fine-grained division
will most often be sufficient, i.e., a distinction into the three main
verb classes: S (strong verbs), WS (the smaller weak class), WL (the
larger weak class).
3. Results
3.1. Correct performance
There was a clear developmental progression in the overall correct

performance, as shown in Table 3. The mean percentage of correct
answers increased and the standard deviation decreased substantially
with age. (Of course, age is a notoriously poor indicator of
development, but we have tried to eliminate the worst features of
overlap by using 2-year gaps.) All differences in performance
between age groups were significant.2
Table 3. Correct answers for all verbs by age group

4-year-olds 6-year-olds 8-year-olds
Mean 51% 72% 90%
SD 19% 12% 8%
However, the differences in performance were not the same across

the three main verb types, as shown in Table 4:
Table 4. Correct answers by verb type and age group

WL class WS class S class
Age Mean SD Mean SD Mean SD
4 years 85 % 20 % 47 % 27 % 33 % 24 %
6 years 94 % 8% 71% 18% 60 % 22 %
8 years 96 % 5% 91 % 10 % 86 % 14 %
The children scored highest on the verbs from the larger weak class
(WL), followed by the smaller weak class (WS), and lowest on the
strong verbs (S). The differences in performance were significant
between all age groups for all main verb types—except for the WL
verbs, which reached a ceiling at age 6 (i.e., the difference between
ages 6 and 8 was not significant).
For the WS verbs, performance increased significantly from 47%
at age 4 to 71% at age 6, but it was still significantly behind the
performance for the WL verbs at this age, clustering with the strong
verbs. It was not until age 8 that the children performed equally well
on the verbs from the two weak classes.
Performance for the strong verbs lagged significantly behind that
for the WL verbs in all age groups. At age 4, it was significantly
behind that for both weak classes; at age 6, it clustered with that for
the WS class significantly behind that for the WL class, and at age 8
it was still significantly behind that for the WL class verbs. As for
the subgroups of strong verbs, the performance for subgroup S2 was
the lowest throughout the period tested, and significantly lower than
for most other groups.
The effect of getting the infinitive or the present tense as a cue
seemed to be important for the younger children, but not for the 8-
year-olds, as illustrated in Table 5—the children who got the present
tense as a cue (where the word tone provides information about
whether the verb is weak or strong) performed better than those who
got the infinitive (where the word tone does not provide such
information):
Table 5. Correct answers by cue and age: Percentage correct with either infinitive
or present tense as a cue.
Infinitive Present tense
Age Mean SD Mean SD
4 years 47% 19% 56% 18%
6 years 69% 12% 78% 8%
8 years 90% 9% 90% 6%
Statistical analyses revealed a significant main effect for cue, but

when each age group was considered separately, this effect was
significant only for the 6-year-olds (for the 4-year-olds, there was
probably too much variation for the difference in performance to be
significant).3
As mentioned earlier, the weak verbs were split into two groups,
rhyming and non-rhyming. There were no significant differences in
correct performance between rhyming and non-rhyming groups of
weak verbs. However, for the 8-year-olds the rhyme effect was close
to being significant, and interestingly, in a replication of this

experiment performed on adult Norwegian subjects by Kirsten Meyer
Bjerkan, there was a significant rhyme effect, with consistently and
significantly better performance on the non-rhyming weak verbs
(Simonsen and Bjerkan 1998).4
On the other hand, looking at effects of token frequency on correct
performance, this performance was significantly better on high token
frequency verbs than on low token frequency verbs, for all the
children at all ages. As for the different verb classes, token frequency
was significant for strong verbs and for WL class verbs, but not for
WS class verbs. However, the token frequency effect seems to
diminish with age: for adult Norwegian subjects no main effect of
token frequency was found (Simonsen and Bjerkan 1998).
3.2. Errors
The errors made by the children were classified into 6 main types:
1. Generalisation of the weak large pattern (GEN>WL)

2. Generalisation of another weak pattern (GEN>WS)
3. Generalisation of a strong pattern (GEN>S)
4. No change of stem (NO CHANGE)
5. No answer (NO ANSWER)
6. Non-past form (NON PAST)
As was the case with correct performance, the types of errors made
by the children also changed with age, as illustrated in Table 6. (Note
that the percentages of errors in this table add up to 100% of the
errors, not of the total performance. The general error level in each
age group is indicated in the column "average number of errors",
referring to the average number of incorrect forms out of the 60 verbs
contained in the test.)
By far the most frequent error types were the types 1-3,
overgeneralisations of one of the main inflectional patterns. One
exception was found in the 4-year-olds, who made quite a few NON
PAST errors, where infinitive or present tense responses were given

instead of a past tense form. These errors occurred in all verb classes
and can be interpreted as a general error strategy of "repeating the
cue"; this error type more or less disappears at age 6 and will not be
discussed further here.
Table 6. Error types: Percentages of all errors across verb groups

E R R Ο R
Average #1 #2 #3 #4 #5 #6
Age number GEN GEN GEN NO NO NON-
of errors >WL >WS >S CHA- ANS- PAST
(max 60) NGE WER FORM
4 years 29.4 44% 12% 2% 1% 4% | 38% |
6 years 16.7 52% 27% 9% 2% 2% 7%
8 years 6 37% 44% 13% 2% 4%
Focussing on the overgeneralisation errors, we see that at ages 4 and

6, GEN>WL was clearly the dominant overgeneralisation error type.
However, GEN>WS errors increased steadily, and already at age 6
they accounted for more than 1/4 of the errors. At age 8, this
tendency continued, and as the percentage for GEN>WL decreased,
that for GEN>WS increased, so that the error type GEN>WS
surpassed GEN>WL in the 8-year-olds. And again, interestingly, in
the replication of this experiment made on adult subjects, GEN>WS
was found to be the dominant error type, accounting for 45% of the
errors as compared to 10% of GEN>WL and 20% GEN>S
(Simonsen and Bjerkan 1998). So-called "irregularisation" errors,
i.e., the overgeneralisation of certain strong verb patterns (GEN>S),
were a much less frequent error type, but they did occur, increasing
slowly with age. The developmental pattern of overgeneralisation
errors, here including the adults, is illustrated in Figure 3.
As for the distribution of error types across verb classes,
GEN>WL errors were found in all verb types, with one exception:
verbs with a monosyllabic infinitive showed no such errors. Among
the strong verbs, the subgroup SI in which all verbs have a short
infinitive stem vowel, overwhelmingly got GEN>WL errors, while

subgroup S2, in which all verbs have a long infinitive stem vowel (as
well as the verbs from subgroup S3 with a long infinitive stem
vowel) showed a more balanced distribution between GEN>WL and
GEN>WS errors, with an increasing amount of GEN>WS errors with
age. GEN>WS were also an important source of errors in WL verbs.
GEN>S errors occurred both in weak and strong verbs, but in a
relatively small set of verbs, and only in verbs which rhyme with
strong ones—thus there was an effect of rhyme here after all. In the
strong verbs, such errors typically occurred either in verbs which
have an irregularity within their own class (yielding, for example,
lyve 'tell a lie' -I0yv instead of Iyve-I0y), or they were inflected
according to another strong pattern with a higher type frequency than
their own (for example, ligge 'lie down' -lagg instead of ligge-la.)
•GEN>WL
•GEN>WS
-GEN>S
0%
Age 4 Age 6 Age8 Adults
Figure 3. Development of overgeneralisation errors: Percentage of total errors
4. Discussion
It seems evident from these results that input factors play an

important role for the acquisition of past tense in Norwegian.
4.1, Type frequency
Type frequency seems to play a major role in correct performance.

All the three main inflectional classes were acquired in an order
reflecting their type frequency: WL>WS>S. Thus, the larger the
number of verbs in an inflectional class, the earlier it was acquired.
However, for the different subgroups of strong verbs tested, we did
not find any positive effect of type frequency on correct performance.
Subgroup S2, which has a relatively high type frequency among the
strong verbs, was acquired later than subgroup S3, which has a lower
type frequency. It is possible that the differences in type frequency
between these subclasses are not large enough to have an effect. In
addition, token frequency may play a role here: overall, the S2 verbs
have a lower token frequency than the S3 verbs.
Type frequency is clearly also relevant for overgeneralisations.
The overgeneralisation errors are relatively well synchronised with
correct performance: all the three main classes form bases for
overgeneralisations, and to a large extent their relative importance as
sources of overgeneralisations reflect their type frequency. But since
the WS class becomes increasingly more important as a source of
overgeneralisation with age, other factors (such as phonological
factors) must also contribute to the pattern of performance.
4.2. Phonological factors
In fact, to a large extent the phonological properties of the verb

classes seem to line up with the pattern of type frequency in
Norwegian. For instance, if we put the three main inflectional classes
on a scale of segmentability and salience of their past tense forms,
the ranking will match a type frequency ranking.
The larger weak class is inflected with a syllabic suffix, it has one
single past tense suffix, starting with a vowel, and there is no change
in the stem when the suffix is added. (Examples: haste 'throw' -
kast-et; tine 'peep' - titt-et, see also Table 2 above.) The smaller
weak class is also inflected through suffixation, but there are two
different suffixes to choose between; the suffixes start with a

consonant, possibly making them less easily segmentable from the
stem; and often assimilations and other smaller stem changes like
vowel shortening also occur. (Examples: lyse /ly:se/ 'light' - lys-te
/ly:ste/; leve /le:ve/ 'live' - lev-de /levde/; kj0re /90:re/ 'drive' - kj0r-
te /90:te/; see also Table 2 above.) The strong verbs are found at the
lowest end of the segmentability scale, being inflected without a
syllabic suffix, but most often with a change in the stem vowel.
(Examples: sine 'sit'-sa/ί; sove 'sleep'-sov.) These characteristics
are summed up in Table 6 below.
Table 6. Type frequency and salience/segmentability

WL class WS class Strong verbs
Type High Mid Low
frequency •56% •40% •4%
Salience/ High Mid Low
segmenta- •suffix: •suffix: •no syllabic suffix
bility syllabic, syllabic,
vowel initial consonant initial
•stem: •stem: • stem:
no change often consonant vowel change
assimilation and
vowel shortening
As already mentioned, the verb classes also differ by way of degree

of phonological openness and coherence. The WL class is clearly the
most phonologically open, while the WS class has subclasses with a
high phonological coherence, as well as a more restricted set of
phonological patterns involving prosody (vowel length), vowel
quality, and consonants. Since these patterns are not absolute, but
only tendencies, they may take more time (and more verbs) to
discover than more clearcut class patterns. This way, we seem to
have evidence of a critical mass effect: before the child is able to
generalise over these subtle patterns, a substantial number of verbs
have to be acquired which illustrate them. This may help explain the
developmental profile of the weak verbs: the higher segmentability
and the higher type frequency of the WL class, combined with its
phonological openness, makes this class the easiest one to acquire

and generalise from. However, given more time and more verb
forms, the higher phonological coherence of the WS class makes it
possible to establish phonological patterns for children to generalise
over - both correctly and incorrectly.
As for the strong verbs, phonological coherence does not seem to
have a positive effect on correct performance: in fact, the subclass
with high coherence (and high type frequency), S2 (the fryse-fr0s
class), was the one acquired latest of all. As indicated above, token
frequency may play a role here, since this class has a lower global
token frequency than the other strong classes. However, phonological
coherence does seem to have an effect on the overgeneralisation
errors made: the strong subclasses with the highest phonological
coherence served as the most important bases of GEN>S errors.
Effects of phonological similarity were found only in the older
children. The rhyme effect—i.e., a poorer performance on weak
verbs rhyming with strong verbs than on those not rhyming with
strong verbs—did not even reach significance in the 8-year-olds, but
was significant in the adults. This indicates that the strong verbs and
their phonological patterning need to be mastered for this effect to
emerge. The error patterns corroborate this assumption: for the strong
verbs, all the verbs generalised to strong classes rhyme with strong
verbs, showing a similarity effect, but again, such errors do not occur
before the strong verbs are mastered well enough to be generalised
over. Similarity effects were also found in the weak verbs:
overgeneralisation errors of verbs with a monosyllabic infinitive
were never of the GEN>WL type. Furthermore the distribution
between GEN>WL and GEN>WS, in particular in the older children,
was to a large extent predicted by the predominant vowel length in
the weak classes.
The information provided in the cue by the word tone concerning
verb class membership, helping to distinguish between strong and
weak verbs, turned out to be important for the younger children but
diminished with age, having no effect for the 8-year-olds. This cue is
salient, but can only help the child distinguish between a strong and a
weak verb; it does not indicate which of the two weak inflections or
which one among the strong inflections is needed. Thus, to inflect the
verbs correctly, the child needs to discover other, more detailed
patterns for identifying each of the verb classes, and at that point the
tone cue loses some of its importance. Interestingly, in a
connectionist network simulation of past tense acquisition in
Norwegian, a similar pattern was found: word tone information was
most important for correct performance early in the training period
(N0klestad 1996). And in a similar experiment to the present one for
Italian children, a parallel cue effect was found to be significant only
for the 4-year-old children (Matcovich 1998).
4.3. Token frequency
Finally, the results showed a clear token frequency effect on correct

performance in all age groups, and notably this effect was found not
only in the strong verbs, but in the WL class as well—the class
which would otherwise be the closest candidate to functioning as a
default class. (I have no explanation for the lack of token frequency
effect in the WS class). This indicates a similarity in the processing
of verbs irrespective of their status as weak or strong. As opposed to
the phonological similarity effects, the token frequency effect seems
to diminish with age, with the adults showing no such effect. A token
frequency pattern with a similar developmental profile was found for
Italian children (Matcovich 1998, see also Matcovich this volume).
Again it seems possible that in the course of development, when
enough verbs are acquired for patterns to be established, the strength
of the patterns for the individual verbs may be overridden by the
strength of the pattern representative of the class as a whole.
5. Conclusions
In sum, the results support the position that input factors such as
phonological and frequency factors play an important role in the
acquisition of past tense morphology in Norwegian, both in terms of
correct performance and the types of errors made. Importantly, these

factors also seem to play a role for all verb classes, not only for the
strong verbs. All the three main verb classes are used as bases for
generalisation: the weak classes are used more often, but there is no
single class which is used as a default across all age groups. In the
course of development, children gradually use more classes
productively. The more verbs and verb classes the child acquires, the
larger the basis for generalisation, and the relative importance of each
class in this respect changes during development. There is little
evidence of clear-cut differences between the classes of an either-or
type, but rather the differences seem to be of a gradual (and dynamic)
nature.
How do these facts fit with the two different models mentioned in
the introduction? For the dual mechanism model, the fact that there
are three main classes (WL, WS, and S) and not only two (W and S)
already presents a problem: where should the third (WS) class be
placed—in the rule mechanism, or in the lexical retrieval
mechanism? At the outset, the children seem to treat verbs from this
class in the same way as the strong verbs, indicating that the WS
verbs are retrieved from the lexicon. However, in the course of
development children gradually come to master these verbs better,
and also start overgeneralising increasingly more to this pattern—and
at the adult level the WS class becomes the clearly dominant source
of overgeneralisation errors. This shift complicates matters for the
dual mechanism model: although an analogical component in the
lexical retrieval mechanism might be posited to allow for
generalisation effects to occur, the fact that such effects become
predominant seems to contradict or at least strongly stretch the idea
of the lexical retrieval mechanism as a "minor" mechanism.
The larger weak (WL) class is the most obvious candidate for the
default, rule-governed class in Norwegian, and at the beginning of
the development it does seem to function as a default for the
children, being clearly the class they master best, as well as the
dominant source of overgeneralisation errors. However, according to
the dual mechanism model neither frequency factors nor phono-
logical factors should influence the default class, yet both the correct
performance and the overgeneralisation errors in our Norwegian data

seem to contradict this claim.
As for type frequency and phonological openness, these factors to
a large extent coincide in the different verb classes, with the WL
class as the one which is both phonologically most open and the class
with the highest type frequency. The way the Norwegian verb classes
are structured, then, it is difficult to single out the effects of each of
these factors and to separate those again from the possible effects of
the WL class as a default. Since the WL class is so open as to allow
nearly any phonological pattern, phonological similarity effects in
the errors are also hard to identify. However, the few phonological
restrictions found in this class are in fact mirrored in the errors.
Verbs with monosyllabic infinitives are not found in this class, and
accordingly there are no errors of overgeneralisation of such verbs to
this class. In addition, the distribution of vowel length in the two
weak classes is to a large extent mirrored in the distribution of
overgeneralisation errors to these two classes, in particular in the
older children.
Furthermore, the results show token frequency to have a
significant effect on performance already from the beginning of the
development registered in the Norwegian children, not only for the
strong verbs, but also for the verbs from the larger weak class—
contradicting the claims made by the dual mechanism model.
The results show that the relative importance of the different verb
classes as bases for generalisation changes during development,
indicating that the differences between the classes are gradual rather
than of a dichotomous nature. This seems poorly explained through a
model with a strict partition into two qualitatively different
mechanisms. Furthermore, the manner in which the relative role of
the different input factors also changes in the course of development
makes it worth while turning to the other model, the single
mechanism model.
This model sees all inflectional patterns built up in essentially the
same way, through usage of the different verb forms. Generalisations
are formed over verb forms with perceived similarities, and the larger
the type frequency, the stronger the schema. Thus input factors
should play a similar role for all verb classes. The developmental
pattern found here, with type frequency as an important predictor of
correct performance, a relatively well synchronised development
between correct performance and overgeneralisation patterns, and
token frequency playing a role for both strong and weak verbs,5 are
all in accordance with such a model.
The same goes for the evidence of changing importance of
different input factors during development: While token frequency,
salience, and the effect of word tone play an important role earlier in
development but become less important with age, other factors like
the effects of phonological patterning of the different classes and
rhyme effects do not show up at the beginning, but become in-
creasingly important with age.
These results can all be reasonably explained in a model based on
increased entrenchment of verb forms and Schemas formed across
them through usage. When enough verbs from a verb class have been
acquired to form a pattern or schema, the importance of the token
frequency of each verb may become less important than the strength
of the schema. The information carried by the word tone is salient,
but can only partly help the child in identifying class membership;
more subtle phonological information is needed to assign the verbs to
specific strong or weak inflectional patterns. However, since many of
the phonological patterns found are indeed not absolute, but only ten-
dencies, a substantial amount of verbs have to be acquired for these
patterns to be identified, and well enough entrenched for the child to
start generalising over them. This may explain the late but increasing
importance of phonological patterning during development. Thus,
this evidence of a dynamic interaction between different input factors
is consistent with an exemplar- or usage-based, single mechanism
model of acquisition, storage, and processing.
To what extent these results for Norwegian find a parallel in other
languages, is of course an empirical question. Ragnarsdottir,
Simonsen, and Plunkett (1999) show that a similar acquisition
pattern applies to Icelandic—another Scandinavian language.
Experimental studies of Italian, still another language with less of a
dichotomous distinction into regular and irregular patterns than
English, have shown that irregular and regular verbs have the same
priming patterns, and that phonological factors and frequency factors
seem to play a similar role for regular and irregular verbs—these
results have been found both for adults (Orsolini and Marslen-Wilson
1997) and for children (Matcovich 1998, this volume). This might
indicate that English, with its clearcut division which has served as
the starting point for the dual mechanism model, is more of an
extreme case, and as such less suited as a universal template.
However, even for English, recent experimental studies have yielded
parallel results to the ones found here. Marchman (1997), using the
same elicitation task as the present one for past tense forms in
English-speaking children, found that both regular and different
irregular patterns were used productively, and that their use was
influenced by sets of item-based properties, notably frequency and
phonological factors, which all contributed through competing and
converging pressures to the productivity patterns found.
Notes
1. I am indebted to Kim Plunkett and Hrafnhildur Ragnarsdottir for their

contribution to the research behind this paper. I also wish to thank two
anonymous reviewers for helpful comments.
2. Statistical significance was tested through ANOVAs and Tukey post hoc tests
for multiple comparisons, with a significance level of p<0.05 in all cases.
Details of the statistical analyses are found in Ragnarsdottir, Simonsen, and
Plunkett (1999).
3. A two-way ANOVA was performed using Age (4,6 and 8 years) and Cue
(infinitive, present tense) as factors.There was a main effect of Cue (F(l,90) =
5.424, p=0.022), but no interaction between Age and Cue (F(2,90) = 1.484,
p=0.232). For each age group, a one way ANOVA revealed that Cue was
significant only for the 6-year-olds (F(l,27) = 5.5192, p=0.0264), but not for
the 4-year-olds (F(l,26) = 1.5435, p=0.2252) nor for the 8-year-olds (F(l,37) =
0.0010, p=0.9748). I am grateful to Paola Matcovich for performing this
statistical analysis.
4. The test was administered in a slightly different way to the adults, with a time
constraint to induce errors. The subjects were presented orally with the
infinitive forms of the verbs, and were asked to produce the past tense form
immediately, without time to ponder. Although the test conditions were not
identical, it is interesting to compare the adult and child data for similarities
and differences in processing patterns.
5. The fact that token frequency is not significant for the WS class remains
unexplained for this model, too.
Individual variation in past tense inflection:
Experimental data from Norwegian SLI children
Kirsten Meyer Bjerkan
6- and 8-year-old children with specific language impairment (SLI)

are tested for past tense inflection of verbs in Norwegian. The focus
in this article is on two of the 6-year-old children, who both have
fairly low correct scores, but who use very different strategies in
their response to the test; one of them tends to repeat the present
tense input form, whereas the other always uses the past tense, but
not always the correct inflection. The difference between these
children is discussed with relation to the formation of Schemas, and
to what degree these Schemas are entrenched.
I. Introduction
Specific language impairment is an impairment which affects only

language in otherwise normal children. To be diagnosed as
specifically language impaired, a child must have normal non-verbal
intelligence, normal hearing abilities, normal physical development
and no severe emotional disturbances that could cause language
problems (PICA 1962). Approximately 7% of all pre-school children
are affected (Leonard 1998; Tomblin et al. 1997). They may have
problems with all aspects of language, but morphology is often
particularly difficult for these children (Leonard 1998).
Past tense, with its combination of regular and irregular (or weak
and strong) inflection, provides an instructive insight into the
morphological competence of SLI children. In this article, experi-
mental test data from SLI children and normally developing children
will be presented, and two of the SLI children will be selected for
further discussion. The analysis is framed within Cognitive
Grammar, using Schemas (Langacker 1987a).
154 Kirsten Meyer Bjerkan
2. Method
2.1. Subjects
For this study, eight SLI children were selected. They all
demonstrated a score of at least one standard deviation below the
mean for their age group on standardised language tests, and showed
cognitive functioning within normal limits. Four of them were 6
years old, and the other four were 8.
Three groups of normally developing (ND) children, studied by
Hanne Gram Simonsen, serve as normal control data. These children
are in the age groups 4, 6 and 8, and there are approximately 30
subjects in each group. (Simonsen this volume; Ragnarsdottir,
Simonsen, and Plunkett 1999).
2.2. Procedure
The data were collected by means of a past tense test, constructed by

Ragnarsdottir, Simonsen, and Plunkett (1999). The test consisted of
60 verbs, half of them weak and half of them strong. The formal
distinction between strong and weak verbs is that weak verbs carry a
syllabic suffix in the past tense, whereas strong verbs do not.
Norwegian has two weak classes, here called the larger weak class
(WL) and the smaller weak class (WS). The larger one has more
members and is generally more productive than the smaller class.
These two classes differ in that they carry different suffixes. The WL
class takes a suffix beginning with a vowel, -/et/ or -/a/, where the
variation between the suffixes is sociolinguistically or stylistically
determined. The other weak class, WS, takes a suffix starting in a
consonant, -/te/ or -/de/. The use of these two suffixes is lexicalised
to each verb and is to some extent phonologically predictable. The
two weak classes were represented by approximately the same
number of verbs in the test.
The verbs in the test were carefully chosen to take into account
properties assumed to influence acquisition, such as frequency
Past tense inflection in S LI children 155
factors and phonological properties of verbs and verb classes. Type

frequency (the number of verbs in a verb class) vary among
Norwegian verbs, in that the larger weak class has the highest type
frequency, the smaller weak class has a medium type frequency, and
the strong class has a fairly low type frequency. Between the
different strong subclasses, there are also differences in type
frequency, and this was mirrored in the test; one subclass with
relatively high type frequency and one with low type frequency were
included. Token frequency (the frequency of use of each verb) was
varied so that items with both high and low token frequency were
included in all the verb classes.
As for phonological factors, the phonological coherence of each
verb class (that is, the degree of phonological similarity between the
verbs within each class) and the phonological similarity/difference
between inflectional classes (that is, whether verbs from different
classes have stems which rhyme or not) were taken into account. The
phonological coherence differs between the strong subclasses, and
there is also a difference between the weak classes, in that the smaller
weak class is more coherent than the larger one. Regarding rhyming,
in each of the weak classes, half of the verbs rhymed with strong
verbs, and the other half did not.
See Simonsen (this volume) for a thorough presentation of the
past tense test; for a presentation of the verb classes in Norwegian,
see Endresen and Simonsen (this volume).
The elicitation task was the same as the one used by Bybee and
Slobin (1982). The child is shown a picture of someone performing
an action. The experimenter says "here you see a girl who is
VERBing. She did exactly the same thing yesterday. What did she do
then?", and the child does not need to say more than the past tense
form of the verb.
2.3. Results
2.3.l. Normally developing children
The total number of correct responses shows a clear developmental

effect, and the difference between each of the three age groups is
statistically significant. The performance is, however, not the same
across all the verb classes. The 4- and 6-year-olds have a relatively
similar profile, whereas the 8-year-olds are different. The difference
lies in the performance on the smaller weak (WS) class. The 4- and
6-year-olds perform significantly better on the larger weak (WL)
class than on the two other classes. There is no significant difference
between the WS and the strong classes. For the 8-year-olds, on the
other hand, the performance on the WS class clusters with that of the
WL class, whereas there still is a significant difference between the
WL class and the strong class. The correct performance for each age
group is shown in Table 1 below.
Table 1. Correct answers, ND children

AGE Total WL class WS class S class
ND4 51% 85% 47% 33%
ND6 72% 94% 71% 60%
ND8 90% 96% 91% 86%
Both the number and the types of errors change with age. The main
error types are overgeneralisation errors (that is, productive use of
one verb class, Gen>WL is the productive use of the larger weak
class, etc), and Non-Past errors (that is, to give the infinitive or pre-
sent tense instead of the past tense form). The 4-year-olds generally
either overgeneralised to the larger weak class, or they gave a Non-
Past response. The 6-year-olds almost never used a Non-Past form.
Their most frequent error type was Gen>WL, but they also had a
number of Gen>WS. The 8-year-olds overgeneralised slightly more
to the WS than to the WL, and both these weak classes were used
productively to a large extent. There was also a slight increase in the
Gen>S (productive use of strong patterns) with age. Cf. Table 2.
Past tense inflection in SL1 children 157
Table 2. Error types, ND children

AGE GEN GEN GEN NO NO NON-
>WL >WS >S CHA- ANS- PAST
NGE WER FORM
ND4 44% 12% 2% 1% 4% 38%
ND6 52% 27% 9% 2% 2% 7%
ND8 37% 44% 13% 2% 4%
2.3.2. SLI children
The SLI children generally had a poorer performance than the ND

children, and they often showed particular problems with one of the
verb classes, either WS or S. Table 3 shows the correct performance
of all the SLI children. Since they are too few to constitute a
representative group of the SLI population, they are presented
individually.
The variation is large, both with respect to the total correct scores
and the scores on the separate verb classes. Generally, the SLI
children have a lower correct score than the ND children of the same
chronological age.
Table 3. Correct answers, SLI children

Subject Age Total WL class WS class S class
Ida 6;1;4 53% 88% 35% 44%
Alex 6;6;20 40% 63% 53% 19%
Tim 6;7;21 63% 75% 71% 52%
Rita 6;9;14 72% 88% 94% 48%
Henry 78% 88% 82% 70%
Michael 8;2;0 67% 94% 29% 74%
Albert 8;2;7 68% 88% 88% 44%
Karen 8;8;28 90% 100% 82% 89%
The types of errors they make, shown in Table 4, are to a large extent
synchronised with the correct performance, in that they over-
generalise the most to the class on which they have the highest
correct score. All of them are able to use verb patterns productively,
but most of them also have a number of Non-Past responses, as was

the case for the ND 4-year-olds.
Table 4. Error types: numbers across all verb groups, SLI children
Subject Total no. GEN GEN GEN NO NO NON- WRONG
of errors >WL >WS >S CHA- ANS- PAST VERB
(max 60) NGE WER FORM
Ida 28 8 1 2 3 14
Alex 36 16 8 2 1 9
Tim 22 5 6 1 6 4
Rita 17 6 9 2
Henry 13 5 2 6
Michael 20 8 2 4 1 5
Albert 19 5 7 1 5 1
Karen 6 2 3 1
3. Presentation of two SLI children, Ida and Alex
In this article, I will focus on two of the six year old SLI children, Ida
and Alex. I have chosen these two because they differ the most in
their approach to this task.
3.1. Child 1, Ida
Ida (age 6;1;4) is the youngest of the 6-year-olds. Her overall score is
53% correct answers, which is below the mean for the ND 6-year-
olds (72%), but approximately the same as the ND 4-year-olds
(51%).
Considering the different verb classes separately she performs
fairly well on the WL class, but not on the WS and S classes (Table
3). When compared to ND children of her own age, she is late in
acquiring all verb classes. As for the ND 4-year-olds, she is better
than them on the strong verbs and at about the same level on the WL
class, but is behind their mean on the WS class (although she is
within one standard deviation below). Her particular problem is the
Past tense inflection in SLI children 159
WS class. It seems like she has noticed the regularities of the large
weak class, but not of the small one. This is confirmed by her errors:
She does not have one single overgeneralisation to the WS class.
As is shown in Table 4, her most frequent error type is the use of a
Non-Past verb form. Half of her errors are of this type. This strategy
is used both for strong and weak verbs, but most of the weak verbs
are from the WS class. All the answers are in the present tense, not
the infinitive; she only repeats the input form.
The fairly high number of GEN>WL shows that this pattern is
acquired and that she is able to use it productively. Both strong verbs
and verbs from the WS class are overgeneralised to this class. She is,
however, not totally unable to see any other patterns, she has one
GEN>S; a verb from the small weak class, spise 'eat' is inflected as
/spasis/ according to one of the most frequent strong patterns. We
can, however, not draw any conclusions on the basis of one
occurrence.
Ida is the only SLI child who makes No Change-errors, that is the
use of the verb stem as the past tense form. This error occurred quite
frequently in Bybee and Slobin's (1982) English data, but is rare
among all the Norwegian children, which may be due to
morphological differences between Norwegian and English. Ida
inflects two verbs according to the No Change pattern. Both of them
are strong, and both belong to the same strong class: /fry:s/ instead of
fr0s 'froze' and /ly:v/ instead of I0y 'lied'.
The high number of Non-Past errors (the use of the present tense)
shows that she does not have an obligatory past tense marking on all
verbs, although the grammatical concept of past tense is mastered to
a large extent.
3.2. Child 2, Alex
Alex (age 6;6;20) has the lowest percent of correct answers of all the
SLI children, with a total score of only 24 of the 60 verbs correct, or
40%. As is shown in Table 3, his main problem is the strong verbs.
Like all the other children, he performs best on the WL class. He is
well below the mean for ND 6-year-olds on all the verb classes, and
also below the ND 4-year-olds on the strong verbs and the WL class.
On the WS class his performance is at about the same level as the
ND 4-year-olds with 53% vs their 48%.
What is interesting about Alex, is that even though his score for
correct answers is fairly low, he does mark past tense obligatorily.
He makes only one Non-Past form error, and in this respect he is the
best in his SLI age group, and also better than three of the four SLI 8-
year-olds. This one error is for the verb reparere 'repair' which
belongs to a subclass of the WS class which has very low frequency
in language spoken to children.
The majority of his errors are overgeneralisations to the WL class.
He has a very productive past tense pattern which says that past tense
ends in -a, and this pattern is applied both to strong verbs and to
verbs from the WS class (also in his spontaneous speech, not just in
the test).
He overgeneralises to the WS class, showing that he has acquired
this pattern, too, at least partly. Mainly strong verbs are over-
generalised to this class. Of the eight overgeneralisations, there are
six strong and two weak verbs.
Another frequently occurring error is the use of a wrong verb. The
verbs he uses are always synonymous to the target verbs, and they
are always in the past tense. For Alex, morphological marking of past
tense is obligatory.
4. Discussion
The only similarity between Ida and Alex is that both perform best
on the large weak class. For Ida, the main difficulty is the smaller
weak class, and for Alex it is the strong verbs. They are both able to
overgeneralise to the larger weak class, but only Alex uses the
smaller weak pattern productively as well. Ida's default answer is to
repeat the input present tense; Alex always uses the past tense.
We will now turn to a discussion of the data related to the
concepts of symbols and Schemas (see Langacker 1987a). A
linguistic symbol consists of a phonological and a semantic pole, and

on the basis of all the symbols we hear, we create Schemas. Possible
Schemas for weak past tense inflections in Norwegian (adapted from
Endresen, Simonsen and Sveen 1996: 242) are:
WL: [/2...a/ "~, past tense"]

WS: [/2...te/ "~, past tense"] [A..de/ "~, past tense"]
The schema is to be read as follows: The slashes indicate the

phonological transcription of the verb. The upper case 2 shows the
word tone (Norwegian being a pitch-accent language with two word
tones). The three dots serve as place holders for the stem of a verb,
whereas /a/, /te/ and /de/ are the past tense suffixes in the weak verb
classes. The quotation marks indicate the verb's meaning. The tilde is
an abbreviation of its lexical meaning, whereas its grammatical
meaning (here, past tense) is specified.
Once such a schema is created, it can be instantiated to new verbs.
Some instantiations are conventional in a language community,
whereas others are idiosyncratic or unconventional, occurring only in
one or a few individuals. Examples of conventional and uncon-
ventional uses of weak past tense are:
WL: [/2...a/ "~, past tense"] -> [/2hopa/ "jump, past tense"]
(conventional)
WL: [/2...a/ "~, past tense"] -> (/2drika/ "drink, past tense")
(unconventional)
WS: [/2...te/ "~, past tense"] -» [/2lekte/ "play, past tense"]
(conventional)
WS: [/2...de/ "~, past tense"] -> (/2lede/ "laugh, past tense")
(unconventional)
The conventional past tense forms for 'drink' and 'laugh' are drakk
and lo, respectively.
In order to use a pattern productively, one must not only have
created a schema, but the schema must also be entrenched to a certain
extent. A symbol is entrenched by encounter and by use, and in order
for it to be entrenched, one must pay attention to it. A lack of

attention may, according to some researchers, be an explanation for
the problems of the SLI children. But as we have seen in the data
presented here, both children are able to create Schemas and use them
productively.
SLI children are often hypothesised to have particular problems in
discovering patterns and regularities, and are thought to learn all
inflected forms of words by rote. E.g., Gopnik (1994) and van der
Lely and Ullman (in press) found indications of this in their data.
Their subjects performed better on strong than on weak verbs. They
claim that SLI children's main problem is the formation of rules.
This hypothesis is contradicted by my data. Ida does perform
better on the strong verbs than on the WS class, but she is much
better on the WL class, and she is able to use this pattern
productively. Alex's performance is an even clearer counterexample.
He does not master the strong verbs at all, he has a low score for
correct performance and no overgeneralisation to the strong class.
The weak patterns, on the other hand, are used productively to a large
extent.
For Alex, these Schemas for weak inflection have high degrees of
entrenchment, particularly the WL class, which is his default pattern.
On this basis, we would expect his correct performance on the WL
class to be high, but it is not. This is mainly due to his use of wrong
verbs, which is probably not caused by lack of knowledge about the
inflection of the target verb, but rather by lack of concentration or
attention; he pays more attention to the semantic than to the
phonological pole of the symbol (see Kirchner and Klatzky 1985).
He even sometimes inflects the new verb according to the schema of
the target: e.g. for the target kästet 'threw' he answers hiva, a strong
verb inflected according to the WL schema.
The concept of morphological rules is problematic in accounting
both for the individual children's performance and for the difference
between them. A rule is expected to be either present or absent, and
not only partly acquired. Schemas, on the other hand, can have
different degrees of entrenchment. The schema for the WL class is
well entrenched in both children, but more in Alex than in Ida. Alex
also has a fairly well entrenched schema for the WS class, whereas
Ida has not. She may have created a schema, but that is impossible to
know as long as she has a low score and does not overgeneralise. If it
exists, it is not well entrenched.
Similar Schemas may, of course, be created for the different strong
subclasses. The only indication we find in these data of a schema for
strong inflection is Ida's one overgeneralisation, and that is not
enough to make any claims in either direction. The two other SLI 6-
year-olds I have not presented here did not overgeneralise strong
patterns, either, but the SLI 8-year-olds did, showing that also SLI
children are able to create different Schemas, and extend them to new
verbs.
5. Conclusion
SLI children, even though they do not perform as well as their

normally developing peers, are able to create Schemas based on verbs
they hear, and they are able to extend these Schemas to other verbs.
There is no evidence for an impaired ability to form generalisations.
Of course, we cannot generalise from two cases to a whole
population of SLI children. The variation is vast. And because of this
variation, cognitive grammar is a well suited framework, with its
Schemas which are often conventional, but which can be extended
unconventionally, and the notion of entrenchment which is always
individual.
A connectionist model of past tense acquisition
in Norwegian
Anders N0klestad
In cognitive linguistics theories and in connectionist models, both

regular (or weak) and irregular (or strong) inflectional morphology
are handled by the same kind of mechanism, a mechanism that is
sensitive to the phonological form and the type and token frequencies
of items in the linguistic input. In this paper 1 present a connectionist
model of past tense acquisition in Norwegian and compare the
model's behaviour to results from past tense elicitation experiments
with Norwegian speakers. These comparisons, as well as
investigations of the model's treatment of novel verbs, show that the
model reflects many aspects of past tense inflection among
Norwegian speakers, thus lending support to the single mechanism
view on the acquisition and processing of inflectional morphology.1
1. Introduction
Over the past 40 years, a considerable amount of research has been

carried out on the acquisition and processing of inflectional
morphology, and in particular on the acquisition of the English past
tense. English verbs can be relatively easily divided into two main
classes based on past tense inflection: regular, or weak, verbs like
jump, whose past tense is formed by adding the suffix -ed to the
stem, and irregular, or strong, verbs, which form their past tense
either by a bare stem (such as hit), by some modification of the stem
(sing-sang; bring-brought) or by suppletivism (go-went). The weak
inflection can be applied to any novel or nonsense verb, even to
verbs that violate the phonotactics of English, while generalisation of
strong verb classes is much more restricted (Prasada and Pinker
1993). Furthermore, when children make overgeneralisation errors,
166 Anders N0klestad
i.e., when they apply the wrong inflectional pattern to an existing

English verb, they almost invariably overgeneralise, or overapply,
the weak pattern (Marcus et al. 1992).
This has lead many researchers to believe that strong and weak
inflections are carried out by qualitatively different mechanisms: a
rote learning or associative memory for storing the past tense of
strong verbs, and a suffixation rule that applies to any verb which
does not have a past tense form stored in this memory, or whose past
tense form the speaker is unable to access (see e.g. Marcus et al.
1992, 1995; Pinker and Prince 1988; Prasada and Pinker 1993).
According to such models, only the storage of strong forms should
be affected by input factors like type frequency (the number of verbs
in the class), token frequency (the frequency of occurrence of a
particular verb form) and the phonological form of the verb. Weak
inflection should be insensitive to such factors, since it is carried out
by a rule applying to all words that are marked as verbs.
An alternative account is offered by models such as Bybee's
(1985, 1988, 1995a) Network Model and Langacker's (1987a, 1990)
Cognitive Grammar. According to these models, strong and weak
inflection are carried out by the same kind of mechanism, one that
acquires all inflectional patterns by generalising over the linguistic
input. Consequently, inflection of both kinds of verbs should be
affected by frequency and phonological factors. Single mechanism
accounts for English have also been explored on computers through
the use of connectionist models (e.g. Plunkett and Marchman 1991,
1993; MacWhinney and Leinbach 1991; Rumelhart and McClelland
1986). These are general processing mechanisms which do not
contain any built-in linguistic knowledge, but which are able to learn
regularities like verbal inflections by generalising over the linguistic
input and by being sensitive to the kind of input factors mentioned
above.
Most previous studies of past tense acquisition have focused on
English. However, the English verbal system, with its clear
dichotomy between a very large group of highly regular verbs and a
rather small group of fairly irregular verbs, is not representative of
the verbal systems found in the world's languages. Hence, if our aim
Norwegian connectionism 167
is to discover the mechanisms involved in the acquisition and

processing of inflectional morphology in any language, the English
data need to be complemented by studies of languages with different
past tense systems. Here I will present some of the results from a
series of connectionist simulations of past tense acquisition in
Norwegian (see N0klestad 1996 for a more comprehensive account).
Since Norwegian does not exhibit such a clear regular/irregular verb
dichotomy as is found in English, studies of past tense acquisition in
this language might yield valuable contributions to the debate over
the kinds of mechanisms involved in morphological acquisition and
processing.
The simulation results will be compared to data from past tense
elicitation experiments with 4-, 6- and 8-year-olds reported by
Ragnarsdottir, Simonsen, and Plunkett (1996, 1999), see also
Simonsen (this volume), and similar experiments with adults
described by Bjerkan and Simonsen (1995, 1996). In these
experiments, the subjects were given sentences which contained
infinitive or present tense forms of Norwegian verbs, and they were
encouraged to give replies that contained corresponding past tense
forms. Note that the adults were asked to produce the past tense form
of the verb as quickly as possible, so that unlike the children, the
adults were required to respond under time pressure. By comparing
the connectionist modelling results with the results from these
experiments, we will be able to see to what extent the model supports
a hypothesis that the experimental data seem to suggest, i.e., that
input factors are important for the acquisition of both strong and
weak inflection in Norwegian.
2. The Norwegian past tense system
Norwegian weak verbs can be divided into two main classes,

characterised by different past tense suffixes. Following
Ragnarsdottir, Simonsen, and Plunkett (1999), cf. also Simonsen
(this volume), I will refer to these classes as the larger weak, or WL,
class and the smaller weak, or WS, class, respectively. As the names
168 A nders N0klestad
imply, the WL class is considerably larger than the WS class, and it

is also far more productive, in the sense that the large majority of
novel verbs are treated by Norwegian speakers as belonging to this
class.2 This is a commonly held view among Norwegian linguists
(see e.g. Ragnarsdottir, Simonsen, and Plunkett 1999), and it is
supported by application of Baayen's (1991, 1992) measures of
productivity on the past tense forms found in the bokmäl part of The
Oslo Corpus of Tagged Norwegian Texts (OCTNT) constructed at
the Text Laboratory, University of Oslo. This corpus contains
material from a wide range of text types written in the bokmäl
language standard, which is very close to the spoken language
variety used by the subjects in Ragnarsdottir, Simonsen, and
Plunkett's and Bjerkan and Simonsen's experiments. Only the 11
million words of newspaper, magazine and fiction material in the
corpus were used for the present purposes, since these were the
categories considered to be most similar to spoken language.
The Norwegian past tense system is similar to the English one in
that the majority of Norwegian verbs form their past tense by
suffixation. However, while all English weak verbs take the -ed
suffix, there are a number of different suffixes used by the group of
suffixing verbs in Norwegian. Following Simonsen (1993), I
consider Norwegian weak verbs to be those which take some syllabic
suffix in the past tense, such as hoppe - hoppet 'jump -jumped' and
spise - spiste 'eat - ate'. All other verbs, including those with
asyllabic suffixes, like binde - bandt 'bind - bound', are considered
to be strong. Thus, although weak verbs tend to show far less
modification of the stem than strong verbs, the weak/strong
distinction used here for Norwegian is based on the syllabicity of the
past tense suffix, and not on the amount of stem change or on the
degree of phonological conditioning (i.e., the criteria on which the
regular/irregular distinction has traditionally been based). There are a
number of classes of strong verbs, which either form their past tense
by a bare stem or by one of several possible vowel alternations.
Baayen's concept of global productivity for a morphological class
is defined along two dimensions: a) extent of use, corresponding to
the type frequency of the class, and b) degree of productivity,
obtained by counting tokens from the class in a corpus (the token

frequency of the class) and dividing the number of hapax legomena
(single occurrences) by the token frequency of the class.
Table 1. Type frequencies, token frequencies, number of hapax legomena and

degree of productivity of the main verb groups in Norwegian as measured
on the OCTNT corpus, including and excluding auxiliary verbs
Non- Non- Degree of Degree of
All aux. AU Hapax
aux. productivity productivity
types types tokens legomena
tokens (all verbs) (non-aux.)
Strong 519 514 254947 115973
113 0.0004 0.0009
verbs (14.4%) (14.3%) (61.0%) (48.7%)
WS 1374 1371 120911 80332
337 0.0028 0.0041
class (38.1%) (38.1%) (28.9%) (33.7%)
WL 1716 1716 42058 42058
407 0.0097 0.0097
class (47.5%) (47.7%) (10.1%) (17.6%)
3609 3601 417916 238363
Total 857
(100%) (100%) (100%) (100%) — —
Table 1 shows the type frequencies, the token frequencies, the

number of hapax legomena, and the degree of productivity of the two
weak classes and the group of strong verbs in the OCTNT corpus,
including and excluding auxiliary verbs. Examples of verbs from the
different verb groups are shown in Table 2.
Table 2. Examples of verbs from the different verb groups

Infinitive Present tense Past tense
Larger weak class verbs
hoppe 'jump' /2hope/ hopper /2hoper/ hoppet /2hopet/
Smaller weak class verbs
spille 'play' /2spile/ spiller /2spi|er/ spilte /2spijte/
levere 'deliver' /|e'ue:re/ leverer /|e'ue:rer/ leverte /le'ue:(;e/
skje 'happen' /'Je:/ jJt/er/'Jeir/ skjedde /2Jede/
kreve 'demand' /2kre:ue/ krever /2kre:uer/ krevde /2kreude/
Strong verbs
synge 'sing' /2syrje/ synger /'syrjer/ sang /'sarj/
ligge 'lie' /2Uge/ ligger /'Ijger/ la /l\p'J
I0pe 'run' I0per /'j0:per/ I0p /'lp:p/
Since both type frequency (extent of use) and degree of

productivity are highest for the WL class, Baayen's measures rank
this class highest in global productivity, followed by the WS class
and with the group of strong verbs at the bottom of the productivity
scale.
Verbs in the WL class have a past tense suffix beginning with a
vowel. This suffix is either -et or -a, the choice between them being
sociolinguistically or stylistically conditioned. Verbs in the WS class
have past tense suffixes beginning with a consonant. The majority of
these verbs take the suffix -te, while the rest get a -de suffix in the
past tense.
Related to the difference in productivity between the weak classes
is the fact that they also differ with respect to the phonological
variation of their members. Almost any verb can be inflected as a
WL class verb. The WS class, on the other hand, puts more
restrictions on the phonological form of the verbs. About 80% of the
verbs from this class have a stem ending in -er, and there are also
other restrictions on verb stems in this class (Hagen 1994; Bjerkan
2000; Ragnarsdottir, Simonsen, and Plunkett 1999): Verbs that take
the -te suffix have stems ending in unvoiced stops, nasals, liquids or
/s/, while verbs that take the -de suffix have stems ending in /g/, /υ/,
/j/ or a long, stressed vowel. Stem-final consonant clusters are very
rare in this class. Finally, it should be noted that the WS class is more
irregular than the WL class, in that the verbs in this class more often
exhibit some modification of the stem—usually a shortening of the
root vowel. Thus, we see that in Norwegian there is no clear
dichotomy between regular and irregular verbs, which is why I (like
Ragnarsdottir, Simonsen, and Plunkett, and Bjerkan and Simonsen)
prefer to use the strong/weak classification described above rather
than a classification based on regularity.
3. Connectionist models
We will now take a closer look at connectionist models; consider the

simple illustration of a model in Figure 1 on the next page.
Such models consist of a number of processing units which are

organised into a network. Each unit has a certain activity level, which
may be taken to represent, for instance, the value of a phonological
feature. This means that we can represent a phonological present
tense form by a pattern of activity across a set, or layer, of units,
such as the input layer in Figure 1. Activity will then spread from the
input layer along the connections and make another pattern of
activity emerge on the output layer. This pattern can then be
interpreted as the network's proposal for a past tense form.
output layer
connection
hidden layer
processing unit
input layer
Figure 1. A simple connectionist model
The flow of activity from the input layer to the output layer is
determined by the strength, or weight, of each connection. By
modifying the connection weights the network can gradually learn to
produce the correct past tense form for each given present tense
form. Thus, the network's knowledge about verb inflections will
emerge gradually in the form of a specific set of connection weights.
If the network has to learn to inflect a large number of verbs, this
knowledge will have the form of generalisations over the inflections
of the individual verbs in the input, generalisations which are similar
to the Schemas found in Bybee's model and the schematic units in
Langacker's Cognitive Grammar.
Of course, such a model will necessarily be extremely simplified
compared to the mechanisms found in the minds of real speakers,
and the data it is exposed to will be very limited compared to a real
linguistic environment. Still, if the behaviour of such models is

compared to data on humans' performance on past tense inflection,
they can be valuable tools for investigating the role of input factors
in the acquisition and processing of inflectional morphology.
4. The Norwegian past tense model
A schematic representation of the specific model that was used in the

Norwegian past tense simulations is shown in Figure 2a.
a)
101 output units
60 hidden units
100 input units
b) 100 input units
oo ooo ο oo oo oo
phoneme phoneme suffix stress word tone
stem
Figure 2. Schematic representations of a) the whole Norwegian past tense network,
and b) the input layer of processing units
The model's processing units were organised into three layers.

Present tense forms were represented on the input layer, as illustrated
in Figure 2b. The activation values of some of the units represented
Norwegian connectionistn 173
particular suffixes, while each of the others encoded a phonologically

distinctive feature of Norwegian word forms, i.e., either a feature of
one of the stem phonemes or the presence or absence of a stress
pattern or word tone. The distinctive feature analysis of the stem
phonemes was based on Endresen (1991).
From the input layer, activation spread to the output layer, and the
output was interpreted as a past tense form, again by interpreting the
unit values as the values of phonemic distinctive features or as suffix,
stress or tone markers. The middle layer of units and the bias unit
were not used for explicit phonological coding; they were only there
to increase the computational power of the model.
5. Training the model
Network training was performed with the tlearn simulator using the
back-propagation learning algorithm. The model was trained on
phonological representations of 1709 Norwegian verbs taken from a
frequency list based on 30 novels from the 1950s, '60s and 70s.3
These verbs were the ones that were listed with a token frequency of
2 or more; the rest were used for testing the network's generalisation
to novel verbs, i.e., verbs it had not been trained on (cf. section 5.3).
The model was given the present tense form of the verb as input and
learned to produce the correct past tense form.
Here I will focus on a set of simulations where the model was first
trained to a hundred percent correct performance on the 20 most
frequent verbs, and then the size of the vocabulary, or training set,
was increased gradually, reflecting the vocabulary growth of a child
acquiring the Norwegian past tense (see N0klestad 1996 for
descriptions of simulations with other training regimes). At a
vocabulary size of 100, the vocabulary growth was accelerated, in an
attempt to incorporate the so-called vocabulary spun. The verbs
were introduced according to decreasing token frequency. Thus, the
training set contained only very high-frequency verbs at first and was
gradually extended with verbs of decreasing frequency. Since high-
frequency verbs are unequally distributed between the different
classes, this had the effect that the proportions of both types and
tokens from each of the classes varied throughout the training period.
Figure 3 illustrates the changes in relative type frequency.
Strong vette
WS class vats
WL class verbs
100 200 300 400 500 600 703 ΘΟΟ 900 1000 1100 1200 1300 1400 1500 1ΘΟΟ 1700 1800
Vocabdarysize
Figure 3. Proportions of types from the different verb groups as a function of
vocabulary size
At the point where the vocabulary size started to increase, the

training set contained 60% strong verbs and 40% verbs from the WS
class, but no verbs from the WL class. The proportion of strong verbs
then decreased rapidly, while the proportion of WS class verbs
increased equally fast. During the rest of the training period the
proportion of verbs from these two classes decreased slowly, and the
WL class gradually grew to become the one with the highest type
frequency. At the end of training, the relative ordering of the type
frequencies was the same as that found in the OCTNT corpus,
although the WL class was larger and the WS class smaller than in
the corpus (56.6% and 30.7% respectively). The proportion of strong
verbs was about the same (12.6%). The low initial type frequency of
the WL class had important consequences for the model's behaviour,
see section 5.2.
Figure 4 shows how the proportions of tokens from each of the
verb groups changed during training. The relative token frequencies
are very similar to the relative type frequencies shown in Figure 3,
contrary to what one would expect from the numbers in Table 1. The
reason for this similarity is that the token frequencies used in the
simulations were highly reduced compared to the actual frequencies
found in the frequency list. The reduced frequencies were obtained
by dividing each frequency in the list by a large constant and
rounding upwards to the closest integer, leaving only a very few
verbs with token frequencies above one. There were several reasons
for this frequency reduction.
Strong verbs
WSdass verbs
\M_dass verts
1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1700 1ΘΟΟ

Vocabulary size
Figure 4. Proportions of tokens from the different verb groups as a function of
vocabulary size
Firstly, the most frequent verbs typically have short duration and
receive relatively low stress in connected speech, and thus are
phonetically less salient than many lower-frequency verbs. This is
especially relevant for Norwegian present and past tense forms, since
most strong verbs—which tend to have high frequency—are mono-
syllabic, while weak verbs are generally polysyllabic.
Secondly, although the current model only learns mappings from
present tense forms to past tense forms, this does in no way indicate
a belief in a similar modularity within the mental linguistic system of
a speaker. A natural extension of the single-mechanism view of past
tense morphology is the hypothesis that the mechanism responsible
for acquiring verb form associations in the human mind is also
involved in the processing of other linguistic—and possibly non-
linguistic—input. Such additional input, if included in the simu-
lations, would function as "noise" relative to the past tense

mappings, and might prevent the network from moving too far into
connection weight configurations that are only compatible with verbs
of very high frequency.
Thirdly, a relatively large proportion of the verbs in the frequency
list had to be removed from the training set because they did not fit
into the input representation or because they do not occur in the East
Norwegian dialect that was being modelled. Since most of these
verbs had low frequencies, a number of the remaining verbs would
have been far too high-frequent relative to the size of the training set
if the original token frequencies had been kept unchanged.
6. Analysis of network performance
While the model was being trained, its correct performance was
measured, both on the whole training set and separately on the two
weak classes and on the strong verbs. The errors made by the model
were also analysed.
6.1. Correct performance
Figure 5 shows how the overall percentage of correct past tense

forms varied throughout the training period.
After the initial period of error-free performance, the percentage
of correct forms decreases dramatically as the size of the training set
starts to increase. This is to be expected, as the network then begins
to receive a continuous influx of new verbs. The model quickly
recovers, however, and it increases its performance until the training
set reaches a size of about one hundred verbs. This is where
acceleration of the vocabulary growth begins, which is likely to be
one reason for the sudden drop in correct performance at this point.
However, it may also be related to the fact that around this
vocabulary size the proportion of strong verb types in the vocabulary
has dropped to about 30%, and the proportion of WS class verbs has
stopped growing. At the same time there is an increasing proportion

of verbs from the WL class, which is a class that the model has not
yet learned how to handle, due to the very low type frequency of this
class earlier in training. But again the model recovers, and it
gradually improves its performance until it reaches a 94% correct
performance at the end of training.
100-n
0 100 200 300 400 500 600 700 800 900 100011001200130014001500160017001800
Vocabulary size
Figure 5. The network's correct performance on the entire training set
These results should be compared to the data on correct performance

from Ragnarsdottir, Simonsen, and Plunkett's and Bjerkan and
Simonsen's psycholinguistic experiments, shown in Figure 6.
The 4-year-olds in the experiment got 51% of the past tense forms
right. For the 6-year-olds, the score was 72%, and for the 8-year-olds
it was 90%. The adults scored only a little bit better than the 8-year-
olds, giving 94% correct responses.
A problem with comparing this developmental profile to the data
from the simulations is to decide which stages of the network's
performance to compare to which age groups. This decision might
have been facilitated by estimates of typical type frequencies in the
vocabularies of Norwegian children at different ages. However, since
no such estimates are available, we will have to make a rough
approximation to an appropriate starting point for comparison.
Since the youngest children were 4 years old, they should be well
past the point where the vocabulary spurt begins, so we should
compare the experimental data to the performance of the model after
it has made its second recovery. Now, there do not seem to be any
convincing a priori reasons for choosing a more specific starting
point for comparison. However, other results, which will be
presented below, provide the best match with the experimental data
if comparisons are made from about a vocabulary size of 350, so I
will choose that as the starting point for all comparisons.
Age 4 Age 6 Age 8 Adults
Figure 6. The human subjects' correct performance on the entire set of verbs.
Adapted from Bjerkan and Simonsen (1996)
At the 350 verb mark, then, the network has already reached 87%
correct performance, which is much better than the 4-year-olds in the
experiments, and consequently, its subsequent performance increase
is considerably smaller than the increase with age that was found in
the experiments. Nevertheless, the network's performance curve
shows a continuously decreasing slope, and in this respect it reflects
the tendency found in the psycholinguistic experiments, where the
degree of improvement decreased with age.
Figure 7 shows the network's separate performance on strong
verbs and on verbs from the larger and smaller weak classes, and
corresponding data from the psycholinguistic experiments are found

in Figure 8.
-Strong verbs
•WS class verbs
•WL class verbs
1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 Θ 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1700 1800

Vocabulary size
Figure 7. The network's correct performance on the different verb groups
Strong classes
WS class
WL class
Age 4 Age 6 Age 8 Adults
Figure 8. The human subjects' correct performance on the different verb groups.
Adapted from Bjerkan and Simonsen (1996)
If we continue to use a vocabulary size of 350 as our first point of

comparison, we see that from this point onwards the relationship
between the model's hit rates corresponds to what was found for both
the 4-, 6- and 8-year-olds, with the highest score on the WL class,
followed by the WS class, and with the worst performance on the
strong verbs. However, there is a larger discrepancy with the results
obtained for the adults in the experiments, since the network's
performance on the WS class never supersedes that on the WL class,
and since the slightly better score on strong verbs than on WL class
verbs that was found for the adults is not reflected in the model's
performance.
6.2. Errors
I will now consider the kinds of errors that the model made during
training. It should be pointed out that in the psycholinguistic experi-
ments most of the errors were overgeneralisations. By overgenerali-
sation of a verb class I mean cases where the inflectional pattern of
this class is applied to a verb belonging to a different class, e.g. when
WL inflection is applied (overgeneralised) to a strong verb.
12 CX/ergeneralisation of the strong dasses

Overgeneralisation of the WSdass
5210
S • Q/ergeneralisation of the WL dass
Φ ο
•δ 8
jje
«ο
Ο 4
-§-· η
"ο
2
Ο 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1703 1ΘΟΟ

Vocabulary size
Figure 9. Percentages of the network's total number of errors that were over-
generalisations of the different verb groups
In contrast to what was found in the psycholinguistic experiments,

overgeneralisations never constituted more than 15% of the errors
made by the model. The main part of the errors consisted of an in-
correct change of stress, tone or a stem phoneme, which are types of
errors that human subjects are unlikely to make. Here I will never-
theless concentrate on the overgeneralisation errors, since these were
the most common errors in the psycholinguistic experiments.
Figure 9 shows the model's overgeneralisation rates on the
different verb groups as proportions of the total number of errors. In
other words, it shows how many of the errors resulted from
inappropriate application of WL inflection, WS inflection or some
strong inflection, respectively. Corresponding data from the psycho-
linguistic experiments are shown in Figure 10.
Overgeneralisation of the strong classes

Overgeneralisation of the W S class
Overgeneralisation of the WL class
Age 4 Age 6 Age8 Adults
Figure 10. Percentages of the human subjects' total number of errors that were
overgeneralisations of the different verb groups. Adapted from Bjerkan
and Simonsen (1996)
If we once again concentrate on what happens to the network after

the 350 verb mark, we see that overgeneralisation of the strong
classes and of the WL class stay relatively constant at a fairly low
level throughout training. The low proportion of strong class over-
generalisation is consistent with the experimental results, although
the model does not reflect the small and statistically insignificant
increase of such errors found in the experiments. As for the rate of
WL overgeneralisation, it matches the results for the adults, who
made very few errors of this kind, but it does not reflect the high
proportion of such errors among the children. This is probably due to
the low type frequency of this class early in training; as was shown
in Figure 3, the WL class does not actually become the largest class
until the vocabulary reaches about 550 verbs.
Why, then, did the children in the experiments overgeneralise this
class to such a large extent? One reason, suggested by Ragnarsdottir,
Simonsen, and Plunkett (1997, 1999), might be that verbs from the
WL class are easier to segment than verbs from the WS class. This is
because the suffix of the WL class verbs begins with a vowel, which
probably makes it more salient than the suffixes of the WS class,
since these suffixes begin with a consonant that is also often merged
with the final consonant of the stem (cf. the past tense of levere in
Table 2). The phonological representation used with the
connectionist model, on the other hand, did not reflect differences in
salience, nor did it incorporate any merging of consonants. Hence it
did not reflect this difference in ease of segmentability, which may
be at least part of the reason for the difference in the rate of WL
overgeneralisation between the network and the children.
I will now turn to the most striking feature in Figure 9, which is
the development of overgeneralisation of the WS class. It shows a
steady increase throughout the training period, so that at the end of
training it constitutes two thirds of the overgeneralisation errors. This
development is remarkable, since it corresponds to a clear, but
unexpected, tendency found in the psycholinguistic experiments.
Figure 10 shows that among the human subjects, overgeneralisation
of this class increased dramatically with age. Considering the low
productivity of this class compared to that of the WL class, this was a
surprising experimental result, and it is therefore interesting to see
that the model shows a similar increase in the proportion of WS
overgeneralisation.
In order to explain the low WS overgeneralisation with their
youngest subjects, Ragnarsdottir, Simonsen, and Plunkett (1999)
again refer to the lower phonological salience of the WS suffixes and
to the fact that many WS verbs show stem-suffix assimilation in the
past tense, making them harder to segment than the WL verbs. The
increase in overgeneralisation with age is attributed to the high token

frequency of the WS class leading to more robust long-term learning
and increasing levels of generalisation. While these factors might
well be part of the explanation, they cannot explain the similar
tendency found in the simulation results. As mentioned above, the
model's way of representing word forms did not reflect differences in
ease of segmentability. Moreover, although the WS class did have a
relatively high token frequency throughout the training period,
during the period when its overgeneralisation rate increased the
proportion of WS tokens was lower than that of the WL class and
decreasing (Figure 4).
Ragnarsdottir, Simonsen, and Plunkett do, however, point to the
differences in degree of phonological conditioning of the two weak
classes (cf. Section 1) as an explanation for the slower rate of WS
acquisition. They propose that it takes a longer period of time and
exposure to a larger number of verbs for the language learners to
figure out the phonological restrictions on the WS class than it takes
to learn the relatively open schema of the WL class. As noted by
Simonsen and Bjerkan (1998) and Simonsen (this volume), such a
proposal could also be used for explaining the increase in WS
overgeneralisation in the experiments, and this is an explanation that
also applies to the connectionist simulations. If it takes some time to
learn the more restricted WS schema, overgeneralisation of this class
should emerge only gradually, but its greater degree of phonological
explicitness could eventually lead to a greater force of attraction on
verbs from other classes that match the schema than is the case for
the relatively unrestricted WL schema.
The model's increase in WS overgeneralisation is also relevant for
the much debated issue of prerequisites for generalisation in
connectionist models (cf. e.g. Plunkett and Marchman 1991, 1993;
Marcus et al. 1995). The debate has focussed on the acquisition of
default inflectional patterns like English past tense suffixation and
German plural formation with the -s suffix, and default inflection in
connectionist models has been claimed to depend on superior token
frequency of the default class (Marcus et al. 1995) and/or on a
situation where the verbs forms from the class that are presented to
the model cover the entire "phonological space" of the language

(Hare, Elman and Daugherty 1995). In Norwegian the most "default-
like" class, i.e., the one with the fewest phonological restrictions
(Bybee's 1995a definition of "default"), is the WL class, and this
class also has the highest type and token frequencies throughout most
of the training period. Hence it comes as no surprise that this is the
class that the model most often generalises to novel verbs, as we will
see in the next section. What the increase in WS overgeneralisation
shows is that a non-default class, i.e., one with a number of
phonological restrictions on the verb stems, can become the
predominant source of overgeneralisations even when it has lower
type and token frequencies than another class in the language.
6.3. Generalisation to novel verbs
Finally, the network was tested on 200 novel verbs, i.e., verbs that
were unknown to the model since they were not included in the
training set. The network's treatment of these verbs is shown in
Figure 11. (Note that the maximum value on the y axis is 40%; a
considerable amount of the network's output did not qualify as a
valid past tense form of any inflectional class).
40 ·_
•Generalisation of the
35 strong classes
• Generalisation of the WS
class
I 25 •Generalisation of the WL
class
20··
1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1700 1800

Size of training vocabulary
Figure 11. Network performance on novel verbs

The subjects in the psycholinguistic experiments were not tested on

verbs they did not know, so the model's performance cannot be
evaluated against experimental results in this case. However, it is
worth noting that the network gradually develops generalisation rates
for the different verb groups that correspond to the relative verb class
productivity reported in section 1, with the WL class as the most
productive one, followed by the WS class, and with the lowest
productivity for the strong classes. This result is particularly
interesting when we compare it to the overgeneralisation rates that
were given in Figures 9 and 10, where it was shown that among
errors of overgeneralisation to known verbs, it is the smaller weak
class that predominates in the later stages of development, in the
simulation as well as for the human subjects.
7. Summary and conclusion
In this paper we have seen that it is possible for a connectionist

network to learn to produce correct Norwegian past tense forms
when it is given the corresponding present tense form, and that it
does so without being given explicit information about inflectional
patterns in Norwegian. Furthermore, the model's overall proportion
of correct past tense forms shows a developmental profile that is
similar to the one found in psycholinguistic experiments. Its relative
performance on the different verb groups reflects the performance of
the children, though not the adults, in the experiments, with the best
performance on WL verbs, followed by WS verbs and with the worst
performance on strong verbs. With respect to error types, the model
makes far less overgeneralisation errors than the human subjects do.
Nevertheless, in accordance with the experimental results it
overgeneralises all of the verb groups to some extent. The model also
reflects the fact that there are few overgeneralisations of strong verb
inflection regardless of age, that the overgeneralisation rate for the
WS class increases with age, and that the WL class is rarely
overgeneralised by adults. What it does not reflect, however, is the
high rate of WL overgeneralisation among the youngest children.
Finally, most of the novel verbs for which the model produces a
valid past tense form are inflected as WL verbs, fewer as WS verbs
and only a very few as strong verbs, which shows that the model has
developed an appropriate relationship between the degrees of
productivity for the different verb groups.
These results indicate that a single mechanism that generalises
over its input is sufficient for explaining the relationship between the
productivity of the verb classes as well as the human subjects'
tendency, increasing with age, to overgeneralise the WS class. What
the model does not explain is the high WL overgeneralisation rate
among the youngest children. This could therefore be taken as an
indication that a symbolic rule is needed for WL inflection in the
early stages of past tense acquisition. Furthermore, if one finds the
reduction of token frequencies used in the simulations unacceptable,
it could be argued that such a rule is needed for WL processing later
on as well, since pilot simulations showed that use of higher token
frequencies prevented the model from showing such a high
productivity for the WL class.
However, the idea of symbolic WL inflection is seriously
weakened by Ragnarsdottir, Simonsen, and Plunkett's (1999) finding
that all of their subjects do significantly better on WL verbs with
high token frequencies than on those with low token frequencies. In a
dual mechanism model of the kind proposed by Marcus et al. (1992,
1995), Pinker and Prince (1998) and Prasada and Pinker (1993), high
token frequencies are only needed in order to block application of the
default rule. The rule itself, being the default, applies to any verb that
is not associated with another inflectional pattern in memory, and is
thus not dependent on the verb having a strong mental representation
caused by a high token frequency. Ragnarsdottir, Simonsen, and
Plunkett's results, then, seem to rule out the possibility that WL
inflection is a symbolic default. A better explanation might be
offered by the hypothesis outlined in section 6.2, which implies that
the high rate of early WL overgeneralisation is due to phonological
factors which make the WL class easier to segment than the WS
class. Although connectionist models with more realistic input rep-
resentations are needed in order to explore this hypothesis, the

present simulation results are at least compatible with it.
Interestingly, the only class that did not show token frequency
effects in Ragnarsdottir, Simonsen, and Plunkett's experiments was
the WS class. However, if WS inflection is carried out by a symbolic
rule, it cannot be a default rule of the kind discussed above, due to
the phonological restrictions on the verb stems in this class.
Furthermore, the present simulations show that both the limited
productivity and the overgeneralisation rate of this class can be
obtained without a symbolic rule.
In sum, when combined with results from psycholinguistic
experiments, the results from the connectionist simulations presented
here do indeed seem to lend support to the idea that type frequency,
token frequency and phonological form are important for the
acquisition of both strong and weak past tense inflection in
Norwegian.
Notes
1. I want to thank Hanne Gram Simonsen and Kim Plunkett for their guidance
during the work on this project, and two anonymous reviewers for valuable
comments on an earlier version of this paper. I also thank the Oceania group
and the Research Committee at the Department of Linguistics, University of
Oslo, for grants which enabled me to visit the Department of Experimental
Psychology at Oxford University, where much of this work was carried out.
2. Exceptions are monosyllabic verbs and verbs with stems ending in -er, which
are always inflected according to the pattern of the smaller weak class.
3. This list was provided by the Text Laboratory at the University of Oslo and the
Norwegian Universities Documentation Project. The novels on which the list is
based are now part of the OCTNT corpus, but the corpus itself was not yet
available at the time when the simulations were run.
Regular inflection in the mental lexicon:
Evidence from Italian
Paola F. Matcovich
This paper addresses the competing predictions made by the dual

mechanism model and the single mechanism model, concerning the
base for productive language use. It reports some of the results from
two psycholinguistic experiments, where past participle forms
(existing verbs, as well as nonce forms) were elicited from Italian
children and adults. The first experiment (adapted from
Ragnarsdottir, Simonsen, and Plunkett 1999) was conducted on 130
children aged 4, 6 and 8 and consisted of 61 verbs from different
regular and irregular inflectional classes, varying in type and token
frequency and in phonological predictability of inflection. The
second experiment elicited the past participle forms of 40 nonce
items from 60 adult speakers.
In my interpretation, the obtained results provide support for the
single mechanism model, in that regular and irregular inflection are
shown not to differ qualitatively either with regard to the effect of
token frequency or with respect to phonological factors.
1. Introduction
A much debated issue in the linguistic and psychological literature

has been the question of how lexical and morphological features are
stored and accessed in the brain. More specifically, the issue
addressed has been whether regular and irregular forms should be
treated as qualitatively different from each other, as claimed by the
dual mechanism model (Marcus et al. 1992, 1995; Prasada and
Pinker 1993), or whether they should be treated as the same kind of
entities, as advanced by single mechanism models: connectionist
models, Bybee and Langacker's network models (Bybee 1985, 1988,
190 Paola F. Matcovich
1995a; Langacker 1987a, 1990). Note that while several other

theories address the issue of storage and processing,1 the findings in
this paper are exclusively interpreted relative to the single/dual
mechanism controversy, this having been hotly debated in the last
decade, and hence the discussion refers to these only.
The core of the dual mechanism model in its most recent version
is the claim that a rule-based symbolic component takes care of the
representation and generation of regular forms, while an associative
rote-memory system deals with irregular forms. The single
mechanism approach maintains that all forms, regardless of whether
they are regular or irregular, are processed by one and the same
mechanism. By virtue of this fundamental difference between the
two approaches, the two models make quite different predictions
concerning the role played by factors such as the phonological
properties of regular forms, as well as their type and token
frequency, in establishing and maintaining lexical representation.
Whereas in the dual mechanism model neither token frequency
effects nor the phonological properties of word forms are predicted
to affect regular inflection (this being rule governed), in the single
mechanism model such factors are thought of as playing a pivotal
role both within regular and irregular inflection.
In this paper I will report some of the results concerning these
competing predictions from two psycholinguistic experiments I have
conducted, eliciting past participle forms (existing verbs, as well as
nonce forms) from Italian children and adults.
2. Past participle inflection in Italian
Before moving on to the experiments a very brief description of past

participle inflection in Italian is in order.
Most textbooks of Italian divide verbs into three basic conjugation
classes, identified by the thematic vowel in the verb stem:
Conjugation I with the thematic vowel -a (as in parlare 'to talk'),
Conjugation II with -e (as in credere 'to believe'), and Conjugation
III with -i (as in dormire 'to sleep'). These are exemplified in Figure
Italian regular inflection 191
1, Both historically and currently Conjugation I is the most numerous

and productive inflectional class, with over 3000 members.
Conjugation II and III are much smaller classes, with approximately
400 and 500 members respectively. Conjugation II is no longer
productive; historically it is the most unstable class, with verbs
migrating to the other classes. Conjugation III is still partially
productive, hosting some deadjectival verbs, especially those derived
from adjectives expressing colours, such as the existing verb
ingiallire 'to yellow' or the neologism ingrigire 'to become grey'.
While Conjugation I differentiates itself from the other two classes in
terms of type frequency, Conjugation II differs from Conjugation I
and III in terms of distribution of regularity, with irregular forms
being mostly Conjugation II verbs.
Conjugation I Conjugation II Conjugation III

Infinitive /par'lare/ /'credere/ /dor'mire/
Past participle /par'ato/ /cre'duto/ /dor'mito/
Figure 1. The past participle of regular verbs
Regular Conjugation I and III verbs form their past participle with
the suffix -to attached to the stem ending in the thematic vowel, with
stress falling on the thematic vowel. The same applies to regular
Conjugation II verbs, with the only difference that the vowel -« is
used instead of the thematic vowel -e. Examples are given in Figure
1, where the verbs parlare 'to speak', credere 'to believe' and
dormire 'to sleep' are used.
Conjugation II Conjugation III

Infinitive piangere, prendere morire, apparire
Past participle pianto, preso morto, apparso
Figure 2. The past participle of irregular verbs with the suffixes -(t)to and -(s)so
In the case of irregular verbs, (with the exception of twelve

Conjugation II verbs ending in -sistere in the infinitive and forming
a little subclass of their own), the past participle is formed with the
suffixes -(t)to or -(s)so attached to a stressed stem which has
undergone the loss of the thematic vowel. This is illustrated in Figure

2 and applies both to Conjugation II irregular verbs such as
piangere-pianto 'to cry-cried' and prendere-preso 'to take-taken',
and to Conjugation III irregular verbs, such as morire-morto 'to die-
died' and apparire-apparso 'to appear-appeared'.
Further, the past participle form of nearly all irregular -ere verbs
involves a consonant elision or modification of the verb stem, and in
some cases also a vowel change. This elision or modification of one
(or two) stem final consonants is often governed by phonological
factors such as phonotactic and assimilation processes, though not
always.
Finally, the phonological properties of a given irregular verb often
offer a cue as to what subclass the verb belongs to, and hence are
only partially predictive of its past participle suffix.
Italian past participle inflection, then, opens for two alternative
interpretations of the dual mechanism model according to how the
different types of past participle formation are to be distributed
between the two mechanisms. Under the first interpretation, (which
for simplicity reasons I will refer to as the "three-rule inter-
pretation"), the rule component which deals with regular inflection is
responsible for attaching the -ato, -uto, and -ito suffixes to regular -
are, -ere and -ire verbs respectively. Under the second
interpretation, (which I will refer to as the "one-rule interpretation"),
the symbolic component consists of a single rule (just like in
English), concatenating the -ato suffix to regular Conjugation I
verbs. Both these interpretations have been advanced in the
literature.2
3. The experiments
3.1. Experiment!
The first experiment, which I will refer to as experiment 1, consisted

of an oral sentence completion task eliciting past participle forms of
Italian verbs. The task resembled the pattern designed by Bybee and
Slobin (1982) and reelaborated by Ragnarsdottir, Simonsen, and

Plunkett (1999) for Norwegian and Icelandic children. The
experimenter presented the child with the picture of a person
engaged in a particular activity and described the situation using a
sentence which contained the present tense of the verb depicted in
the picture. She formed a question referring to the past, in order to
elicit the past tense form of the same verb.3
Three groups of children aged 4, 6 and 8 were tested, with 40
children in the 4- and 8-year-old groups and 50 in the group of the 6-
year-olds, for a total of 130 children (as indicated in Figure 3). All
subjects used the dialect spoken in the Trieste region, and so did the
parents of the children in the experiments.4
Age Males Females Tot. # of children

4-year-olds 23 17 40
Figure 3. Distribution of subjects in experiment 1
Each group of children included approximately the same number of

males and females.
Each child was tested on 61 verbs; 15 were regular -are verbs, 14
were regular -ere verbs and the remaining 32 were irregular -ere
verbs belonging to different subclasses. Conjugation III verbs were
excluded since all the variables we were interested in manipulating
(briefly described in the next paragraph) could be tested on
Conjugation I and II verbs alone.
The parameters taken into account in choosing the verbs and verb
classes to include in the test were essentially the factors on which the
models briefly described earlier seemed to make competing
predictions: the inflectional classes had to differ in terms of type
frequency and phonological coherence, whereas the single verbs had
to vary in token frequency and predictability of inflection, in order to
test the role of these factors in acquisition of past participle
inflection.
Here, however, we will be concerned only with the token

frequency parameter. A few methodological details are in order. In
order to investigate the role of token frequency in acquisition, each
verb class was further subdivided into two groups (a and b),
containing verbs with high and low token frequency respectively and
approximately the same number of verbs.
Token frequency values for each of the verbs included in the test
were drawn from De Mauro et al. (1993) Lessico di frequenza
dell 'italiano parlato ("Frequency dictionary of spoken Italian"). The
corpus in this dictionary consists of recordings of spoken Italian (half
a million word tokens, over 15 thousand lemmas), taken from five
different contexts of use.
The values High and Low were assigned to each verb according to
its token frequency (LoToFrVerb < 50 < HiToFrVerb, where ToFr is
the sum total of the five contexts for each verb as in De Mauro et al.
1993). Extremely high token frequency verbs were excluded, as they
were not thought of as being interesting test items due to their
disproportionate frequency of use, as well as very low token
frequency verbs, since the meaning of the verbs that were ultimately
included in the test had to be familiar to the children.
Table 1. Correct PP answers for all verbs by verb class and token frequency
(experiment 1)
Regular -are Regular-ere Irregular-ere Irregular-«re
verbs (p.p. -οίο) verbs (p.p. -uto) verbs (p.p. -so) verbs (p.p. -to)
HiToFr 99% 96% 84% 82%
LoToFr 87% 83% 50% 58%
H-L +12% +13% +34% +24%
The results from experiment 1 showed that token frequency had a

strong effect on children's ability to inflect regular (as well as
irregular) verbs in the task (F,(l,124) = 783.59, ρ < .0001; F2(l,50) =
13.02, ρ = .001).5 This effect was found to be significant (Tukey-b
post hoc analyses with ρ < 0.05) in both regular inflectional classes
that were tested (results are given in Table 1) and for all age groups
(Table 2).
Table 2. Correct PP answers for all verbs by age group and token frequency
(experiment 1)
HiToFr LoToFr
Age 4 84% 51%
Age 6 92% 69%
Age 8 97% 83%
Before commenting upon these results, let us turn to experiment 2.
3.2. Experiment 2
Experiment 2 consisted of an oral nonce task eliciting the past

participle forms of 40 nonce items from 60 adult speakers of Italian.
The aim of the task was to test for effects of phonological similarity
within Conjugation I regular inflection. The reason why Conjugation
I was chosen, is the fact that if effects of phonological similarity do
matter and thus are found to be statistically significant within this
class, then, neither the "three-rule interpretation" of the dual
mechanism model (according to which the rule component is
responsible for -ato, -uto and -ito suffixation) nor the "one-rule
interpretation" (according to which the rule component is responsible
only for -ato suffixation) can be said to draw support from the data.
In designing such a test a major problem was that, due to the high
type frequency and very open phonological specification of the class
of -are verbs, both a nonce verb modelled on existing -are verbs and
a nonce verb devised as having no similarity to any existing -are
verb were likely to be inflected in -ato. This is very similar to the
situation we have in English, with the default past tense suffix -ed.
A way of solving this problem, then, was to model the nonce
items simultaneously on one or several existing -are verbs and on a
group of -ere verbs. This was possible thanks to the wide range of
patterns found within Conjugation I verbs. If the input provided to
the subjects consists of a form that obliterates all differences between
the three conjugations (for instance the first or the second person
singular of the present indicative), then the verb would be susceptible
to be inflected in -ato, but also to take the suffix which is

characteristic for the group of -ere verbs that it has been modelled
on.
For example, if the input form were to be a nonce verb like
appogli (2nd person singular present indicative), designed as similar
to both regular Conjugation I verbs like spogli 'you undress', sfogli
'you leaf and irregular Conjugation II verbs such as togli 'you take
out', raccogli 'you pick up', the subjects have another plausible
alternative to inflecting the verb as appogliato (as in spogliato and
sfogliato), namely appolto (as in tolto and raccoho). Similarly, the
nonce item affargi, which does not rhyme with any existing -are
verb, but which is similar to -ere verbs like spargi 'you spread', is
susceptible to being inflected both as affargiato (due to Conjugation
I's very high type frequency and extremely open phonological
specifications) and as affarso (cf. sparso). This is illustrated in
Figure 4.
Nonce item Rhyming -are verbs Rhyming -«re verbs

spogli (inf. spogliare, p.p. spogliato) togli (inf. togliere, p.p. tolto)
s
appogli f°gH (inf. sfogliare, p.p. sfogliato) raccogli (inf. raccogliere, p.p.
raccolto)
appogliato appolto
spargi (inf. spargiere, p.p. sparso)
affargi
affargiato affarso
Figure 4. Example of two nonce items (experiment 2)
Further, two conditions were created: in Condition 1 nonce verbs

were devised that were similar to existing regular Conjugation I
verbs, whereas Condition 2 comprised novel verbs which did not
rhyme with any existing -are verb. Moreover, both Condition 1 and
Condition 2 verbs, had in turn to be similar to existing irregular
Conjugation II verbs. This is summarised in Figure 5.6
Test cond. Description

Each nonce item was modelled simultaneously on a group of
existing -are verbs and on a family of existing -ere verbs.
Ex.: Condition 1 item appogli is similar to -are verbs like spogli
'you undress' (as well as to -ere verbs like togli 'you take out').
Each nonce item was devised so as not to rhyme with any existing
-are verb, and at the same time was modelled on the same group of
existing -ere verbs as its pair counterpart.
Ex.: Condition 2 item affargi does not rhyme with any existing
-are verb (but is similar to -ere verbs like spargi 'you spread').
Figure 5. Description of Conditions 1 and 2 (experiment 2)
In order to control for possible token frequency effects, the nonce

items were devised in pairs, with one item from each of the two
conditions outlined above, so that each pair was modelled on one and
the same group of irregular -ere verbs.
To achieve this, the test exploited a certain property of Italian verb
inflection, namely the fact that a number of Conjugation II verbs
present a systematic alternation in the verb stem. The results are laid
out in greater detail in Matcovich (1998).
For our purposes it is enough to remember that in each nonce pair
(ex. appogli/appolgo), both items are modelled on the same group of
-ere verbs, namely that of verbs such as togli-tolgo ('you take out-I
take out'). However, the Condition 1 item appogli rhymes also with
existing -are verbs such as spogli 'you undress', while the Condition
2 item appolgo does not rhyme with any existing -are verb. This is
illustrated in Figure 6.
Nonce pair Rhyming -are verbs Rhyming -ere verbs

appogli spogli (inf. spogliare, p.p. spogliato) togli (inf. iogliere, p.p. tolto)
sfogli (inf. sfogliare, p.p. sfogliato) raccogli (inf. raccogliere,
p.p. raccolto)
appolgo no verb ending in -olgare tolgo (inf. togliere, p.p. tolto)
raccolgo (inf. raccogliere,
p.p. raccolto)
Figure 6. Example of a nonce pair (Experiment 2)
Finally, two versions of the test were created (A and B), each
containing one item from all the twenty pairs, so that each test
included ten Condition 1 and ten Condition 2 items. The two tests
were mirror images of each other, in that for every Condition 1 item
included in test A the Condition 2 item of the same pair was assigned
to test B and vice versa. Attention was paid to have roughly the same
number of items in the first person singular as in the second singular
in both tests.
4. Results and discussion
Condition 1 items (rhyming with existing Conjugation I verbs) were

found to be inflected regularly (-ato suffix) significantly more often
than Condition 2 items (which did not rhyme with any existing -are
verb) (F(l,36) = 36.916, p < .0001).7 This difference in regular
response rate between Condition 1 and 2 remained significant also
for (1) all the four alternations employed in the experiment,8 (2)
seven of the eight rhymes,9 (3) over three fourths of the nonce pairs.
Table 3. Percentage of regular Conjugation I (p.p. -ato) responses as a function of

test condition.
Test Condition Gen>ato Gen>irr'°
1 59% 41%
2 31% 69%
As briefly mentioned earlier, in the dual mechanism model token

frequency and phonological similarity effects are predicted to affect
children's correct performance only on irregular verbs. This is due to
the fact that regular forms are believed to be computed by rule, and
rules are insensitive both to token frequency and phonological
similarity effects. Conversely, in all single mechanism models, token
frequency and phonological similarity effects are expected to be
significant for both irregular and regular verbs, since no structural
distinction between the two types of inflections is postulated by this
approach.
Now, the fact that both token frequency (experiment 1) and

phonological similarity effects (experiment 2) were found to be
statistically significant for regular inflection, seems to be perfectly
compatible with the single mechanism account, in that regular and
irregular inflection are shown not to differ qualitatively either with
regard to the effect of token frequency or with respect to
phonological factors. Conversely, these results do not seem to lend
any support to the dual mechanism model.
Finally, note that this applies not only to the "three-rule-inter-
pretation" of the dual mechanism model, but also to the "one-rule-
interpretation", since Conjugation I (which is handled by the rule
component under both interpretations) showed statistically
significant effects of both token frequency and phonological
similarity.
Notes
1. Cf. for instance the Analogy Model of Royal Skousen, (Skousen 1989;
Derwing and Skousen 1994), and psycholinguistic models such as the
Augmented Addressed Morphology Model developed by Burani and
Caramazza (1987), Caramazza, Laudanna, and Romani (1988), Burani and
Laudanna (1992) and the Morphological Race Model formulated by
Frauenfelder and Schreuder (1992).
2. For the "three-rule interpretation" see for instance Orsolini and Marslen-Wilson
(1997). For the "one-rule interpretation" see Say (1997).
3. The elicited form was passato prossimo, which is a compound tense made up
by an auxiliary verb and the past participle of the main verb.
4. The past participle inflection in the variety of Italian spoken in the Trieste area
corresponds (with some entirely predictable phonological mappings) to the
standard norm.
5. Correct response proportions were entered into a multivariate analysis of
variance on subjects and items, with Age (4, 6 and 8 years), Token Frequency
(High and Low), Input Form (first and third person singular of the present
indicative) and Verb Class (regular -are verbs p.p. -ato, regular -ere verbs p.p.
-uto, irregular -ere verbs p.p. -to and irregular -ere verbs p.p. -so) as factors.
The results reported here concern the main effect of Token Frequency.
6. A third condition consisting of rhymes between nonce items and Conjugation I

verbs only was not included since it was not believed to result in added
significance, all items of such a condition being extremely likely to be inflected
in -ato.
7. For each nonce verb the number of regular (p.p. -ato) responses was calculated.
These results were entered into a multivariate analysis of variance with Test
Condition (1 and 2) and Test Version (A and B) as factors. The results reported
here concern the main effect of Test Condition.
8. The four alternations in question were /lg-X/, /k-tj7, /g-dy and /rjg-ji/. Paired
samples t-test revealed that regular Conjugation I responses (p.p. -ato) differed
significantly between the two conditions for all alternations (df = 59; /g-dy: t =
9.17, p < .0001, /k-tj/: t = 3.77, p < .0001, /lg-/C/: t = 6.12, p < .0001, /rjg-ji/: t
= 2.45, p < .017).
9. The eight rhymes employed in the test were the following:
Rhyme 1: -ar-g/dj-are, as in the nonce pair affargo-qffargi
Rhyme 2: -an-g/dj-are, as in the nonce pair miango-miangi
Rhyme 3: -un-g/dj-are, as in the nonce pair pillungo-pillungi
Rhyme 4: -in-k/tf-are, as in the nonce pair sfinco—sfinci
Rhyme 5: -or-k/ff-are, as in the nonce pair allorco-allorci
Rhyme 6: -e-lg/rf-are, as in the nonce pair sfelgo-sfegli
Rhyme 7: -o-lg/a-are, as in the nonce pair infolgo-infogli
Rhyme 8: -e-rjg/ji-are, as in the nonce pair sfengo—sfegni.
Paired samples t-test revealed that regular Conjugation I responses (p.p. -ato)
differed significantly between the two conditions for seven of the eight rhymes:
df = 59, rhyme 1: t = 5.42, p < .0001; rhyme 2: t = 3.16, p = .002; rhyme 3: t =
5.66, p < .0001; rhyme 5: t = 7.29, p < .0001; rhyme 6: t = 5.60, p < .0001;
rhyme 7: t = 3.15, p = .003; rhyme 8: t = 2.45, p = .017.
10. Those are non-Conjugation-I-responses, that is past participle responses other
than -ato. Of these, approximately 70% are irregular Conjugation II responses
(p.p. in -so οτ-to), 18% are regular Conjugation II responses (p.p. in -uto), and
12% are regular Conjugation III responses (p.p. in -ito).
Section III
Constructional aspects of the verb
Malefactive by means of GIVE
Öle Torf inn Fagerli
The present study is concerned with two common ways of expressing

benefactive and malefactive in some selected languages. Serializing
languages—which include among others the Kwa and Gur groups of
West-Africa, languages of South-East Asia, and some pidgin and
Creole languages—very often use a word identical to the verb GIVE
juxtaposed to a host verb in such circumstances, while typical
extensional languages, like for instance West-Atlantic and Bantu
languages, use benefactive or applicative affixes to modify the verbal
phrase. Most of what can be expressed by verbal extensions in some
languages can be directly transferred to serializing languages by
means of serial GIVE, and vice versa. Focus is directed to the fact
that a malefactive event may be linguistically marked by an item
corresponding to the verb GIVE, or by an extension which is typically
used to mark benefactive events. The apparent paradox is accounted
for on the basis of the semantics of the host verb (and its comple-
ment's), if any), together with semantic bleaching of the original be-
nefactive marker. Malefactive GIVE is sometimes merely an affection
marker, the malefactive content being carried by the host verb.
1. Introduction
The present study is concerned with two common ways of expressing

benefactive and malefactive in some selected languages. By the term
malefactive we understand the linguistic coding of an event
describing that something is done to the detriment of somebody or
some situation is bad or hurtful for somebody. Adversative is another
term applied to cover the same meaning. Malefactive coding is often
linked to benefactive in a formal way, by the fact that the two often
share morphophonological patterns. Serializing languages, which
204 Öle T. Fagerli
include among others the Kwa and Gur groups of West-Africa;

languages of South-East Asia; and some pidgin and creole languages,
very often use a word identical to the verb GIVE1 juxtaposed to a host
verb in such circumstances, while typical extensional languages, like
for instance Atlantic and Bantu languages, use benefactive or
applicative affixes to extend the verbal phrase. Most of what can be
expressed by benefactive extensions in some languages can be
directly transferred to serializing languages by means of serial GIVE,
and vice versa. Focus in this paper is directed to the fact that a
malefactive event may be linguistically marked by an item cor-
responding to the verb GIVE, or by a verbal extension which is
otherwise used to mark benefactive events. The apparent paradox is
accounted for on the basis of the semantics of the host verb (and its
complement, if any), together with extension2 of sense of the original
benefactive marker. Malefactive GIVE is sometimes merely an
affectedness marker, the malefactive content being carried by the
host verb or its complement. The presentation and analysis of data is
accompanied by an attempt at approaching the phenomenon within
the framework of Cognitive Grammar. Recent works by Cognitive
Linguists, most importantly Newman (1996, 1997), confirm the
validity of the results of the present study. At the same time, it is
believed that this presentation (slightly) enriches the already well-
elaborated typological description of the GIVE verb.
2. The prototypical GIVE scene
The meaning of GIVE is complex. I hold the literal, concrete sense of

GIVE as basic and as a point of departure for related meanings. This
basic predicate can be viewed schematically as in Figure 1.
There are three participants, a GIVER, a THING, and a RECIPIENT.
The GIVER hands the THING over to the RECIPIENT. In more detail,
the transfer typically involves physical contact with the THING, first
in the hands of the GIVER, then in the hands of the RECIPIENT.
Consequently change of control or possession is often associated
with the act. In the terms of Langacker (1991: 292), giving involves
an energy flow, initiated by the benevolent GIVER (the "energy

source") and ending up with the RECIPIENT (the "energy sink") who
normally accepts the THING. This is referred to as the force-dynamics
of the verb. The energy consists of mental activity on the part of both
the GIVER and the RECIPIENT, and physical movement of all three
participants.
TIME
Figure 1. A basic representation of the GIVE scene
The scenario sketched above may be viewed from different angles,

some parts of it may be focused while other parts may be omitted in
appropriate contexts. Some parts, the THING participant for instance,
may be extended metaphorically to denote abstract entities. Or the
movement to a recipient may be focused/extracted and used to
describe movement along a path to a GOAL, leaving other subparts of
the meaning of GIVE aside. A rather complex verbal meaning like
GIVE capturing such a basic human experience has the potential of
being extended to other semantic fields than just the literal GIVE
scene. There are indeed various ways of both extracting subparts of
the literal meaning and of extending the meaning figuratively in
order to apply the verb to "new" contexts. Cognitive Grammar
emphasizes the importance of discovering relatedness of meanings
and of establishing and explaining (wherever possible) polysemous
patterns. In the sections to follow, GIVE morphemes in complex
predicates of different kinds are used to express among other things
benefactive and malefactive.
206 le T. Fagerli
3. Benefactive by means of verbal extensions
In Fula, an Atlantic language spoken from Senegal to Sudan, we find

a range of verbal suffixes, extending or modifying the meaning of the
verb in various ways, and being linearly ordered between the verbal
radical and its obligatory inflectional marker. I will concentrate on
Fula data in the following two sections, in particular the Aadamaawa
Fula dialect (Cameroon and Nigeria). For an overview and details
about verbal extensions in Fula, see Arnott (1970), Fagerli (1994),
and Breedveld (1996).
Benefactive, malefactive, and allative is marked by the same
verbal extension -AN-.3 Benefactive here covers both the prototypical
case in which we find a GIVER and a RECIPIENT of a concrete THING,
and cases in which there is a benefactor and a beneficiary of an
action, often referred to as a dative event. The difference between the
two types is mainly one of metaphorical extension of the prototypical
act of giving into "giving" of an action. These two main types of
benefactive, the recipient-of-thing type and the beneficiary-of-action
type, are found in the sentences below. In (1) and (2), it is possible to
interpret the RECIPIENT as either recipient of thing or beneficiary of
action. In (3) and (4) we find examples of the latter:
(1) Mi winnd-an-i Yaaya pataakeewol. (Fula)

I write-BEN-PERF Yaaya letter
(i) Ί wrote a letter to Yaaya.'
(ii) Ί wrote a letter for Yaaya.'
(2) Abbo te6-an-i Didi maygoro. (Fula)

Abbo pick-BEN-PERF Didi mango
(i) 'Abbo picked a mango to Didi.'
(ii) 'Abbo picked a mango for Didi.'
F.W. Taylor (1932) reports the same for e.g. sorr-an- 'to sell-BEN',
which may mean 'to sell to' or 'to sell for'.
(3) Mi huuw-an-i mo. (Fula)

I work-BEN-PERF him
Ί worked for him.'
(4) Yaya wacf-an-i yam jahaarjgal rjgal. (Fula)

Yaya make-BEN-PERF me journey this
'Yaya traveled on behalf of me.'
A more subtle classification of benefactive events in Fula is available

in Fagerli (1995).
4. Malefactive by means of verbal extensions
Constructions with the -AN- extension in Fula may have malefactive

interpretations in suitable circumstances. Three characteristic types
of malefactive events are easily observable. First, in an act of giving
the transferred element may be regarded as negative for the
RECIPIENT. While picking a ripe mango or a flower to somebody is
generally seen as benefactive to the RECIPIENT, picking up a rotten
mango or garbage to give to somebody is generally malefactive to
the RECIPIENT, cf (5) below. Second, the host verb of the extension
may denote an action or state that is generally regarded as bad or
negative, like LIE for instance, cf. (6). Third, the host verb of -AN-
may have ablative characteristics, i.e indicate motion away from the
patient, cf (7) and (8). In the second and third cases, the patient
participant is affected by a negatively loaded action. Thus we may
speak of two maleficiary-of-action types of malefactive events, along
with the noted maleficiary RECIPIENT-of-thing type.
(5) Abbo ted-an-i Didi maygoro nyolnde. (Fula)

Abbo pick-MAL-PERF Didi mango rotten
'Abbo picked a rotten mango to Didi.'
Note however that this sentence also may be interpreted in the

beneficiary-of-action way, meaning that Abbo picked a rotten mango
208 Öle T. Fagerli
on Didi's behalf. (6) below is ambiguous. There are two possible

maleficiary-of-action readings of the sentence:
(6) Buuba few-an-i Jekariaw haala mbeewa. (Fula)

Buuba lie-MAL-PERF Jekariaw affair goat
(i) 'Buuba lied to Jekariaw concerning the goat.'
(ii) 'Buuba lied about Jekariaw concerning the goat.'
If the situational background is that Buuba has stolen a goat, the first
interpretation (6i) is presumably that Buuba lies when he tells
Jekariaw that he, Buuba, is not the thief. In the second interpretation
(6ii) Buuba blames Jekariaw for the theft. In both cases Jekariaw is a
maleficiary of the action, though in the second case he may not be
aware of being a RECIPIENT, if at all he is. The semantics of the
sentence is a bit tricky. The RECIPIENT of the LYING message might
be somebody else, while the maleficiary no doubt must be the
indirect object referent (Jekariaw). A potential third reading of the
sentence, meaning that Buuba lied for Jekariaw, is rejected by
language informants, and is at best a highly marked interpretation.
In (7) and (8) the movement described by the verbs hoo'- 'TAKE'
andja6-t- 'SNATCH/TAKE' explain the malefactive reading.
(7) Aaman hoo'-an-i deerdiiko mbeewa. (Fula)

Aaman TAKE-MAL-PERF brother-his goat
'Aaman took a goat from his brother.'
(8) O ja6-t-an-ilaw-i yam defte. (Fula)

he get-REV-MAL-CEL-PERF me books
'He quickly snatched my books from me.'
(REV = REVERSIVE; CEL = CELERATIVE)
The THING (transferred element) moves from the TAKEE to the

TAKER, in other words in the opposite direction of the THING in an
act of GIVING. The verbal extension only indicates to whom the
action is directed, in other words it functions as a kind of
affectedness marker in these sentences. Senses of concrete recipiency
and benevolence, typically associated with GIVE, are suppressed. The

effect of the adversative act is transferred to the RECIPIENT (the
affected maleficiary). Thus there is movement along two lines, one
THING (the goat) moves from TAKEE to TAKER, while the other
THING (the adversative act) metaphorically moves from TAKER to
TAKEE. The subject of the sentence is both TAKER and GIVER. The
force-dynamics of complex predicates like these is far more compli-
cated than in "simple" GIVE predicates. But it seems clear that the
benefactive morpheme functions as a carrier of energy flow all the
way to the energy sink, the negatively affected RECIPIENT.
We get an adversative reading with other extended verbs as well,
where the host verb follows the TAKE pattern, e.g. wujj-an- 'steal
from' (rather than 'steal for'). It would be reasonable however, in the
case of polysemous structures in the individual languages, to bear in
mind the domain of human interest associated with the ablative verbs
in question. BUY has got the same characteristics as TAKE with
respect to direction of movement, but there is normally no adversa-
tive effect on the person with whom one trades. This would be the
case in Aadamaawa Fula, where the extended verb stem sood-an-
preferrably is understood as 'buy for', hence a benefactive reading.
For a special case of BUY in serializing languages, see example (24).
5. Allative
In addition to benefactive and malefactive, the -AN- extension also in-

dicates movement towards a GOAL. The allative function is
sometimes referred to as goal.
(9) Pallaandi dogg-an-i legged. (Fula)

lizard run-ALLATIVE-PERF tree
'The lizard ran towards the tree.'
(10) Nagge umm-an-ake mo. (Fula; Arnott 1970)

cow rise-ALLATIVE-PERF him
The cow started up at him.'
210 Öle T. Fagerli
Noye (1989) uses the term Destinative as a cover term for bene-
factive, malefactive, and allative. Though it is not commonly named
so in the linguistic literature, it seems like a suitable notation on
semantic grounds: There is movement of some element, concrete or
abstract, to a certain destination in all three cases. The polysemy of
the extension is better accounted for by the term destinative than for
instance applicative, which is a rather meaningless term. Applicative
is commonly used to denote a common verbal extension in Bantu
languages. These languages are typologically close to Fula with
respect to verbal extensions. Swahili may express benefactive, male-
factive and allative in a similar fashion to the one described above for
Fula. However, I will not undertake a thorough comparative analysis
here, the Swahili applicative extension being extremely polysemous.
In addition to the possibilities mentioned, it serves to express among
other things manner, purpose, and instrument. In Fula, these latter
three form parts of an independent verbal extension.
6. Benefactive by means of serial GIVE
Typologically very different from extensional languages, serializing

languages commonly apply the verb for GIVE or alternatively SHOW
to express not only recipiency-of-thing but also beneficiary-of-action.
Such serial verb constructions consist of two verbs most commonly
sharing the same subject and direct object, the GIVE verb, which is
the second one in the string, introducing the RECIPIENT or beneficiary
role into the complex predicate. The verb is rendered in English as
the prepositions for or to. These languages normally have no
productive prepositions corresponding directly to the English ones.
The serial constructions below are complex predicates com-
parable to the Fula verbal extensions, being able to express both the
recipiency-of-thing (11 and 12i) and the beneficiary-of-action (12ii-
15) types of benefactive. Introducing several serial languages at this
point shows that the same benefactive construction type is common
throughout Africa and further. Whether all of these languages posit
equal polysemous structures for the the GIVE verb remains an object
of investigation. It is not necessarily the case.
(11) Kofi to-o boll no ma-α Ama. (Akan; Saether0

Kofi throw-Pst ball Def GIVE-Pst Ama 1997: 19)
'Kofi threw the ball to Ama.'
(12) Megbe sefofo ac[e na wo. (Ewe; Essegbey: p.c.)

I-pluck flower indef. GIVE you
(i) Ί picked a flower to you.'
(ii) Ί picked a flower for you.'
(13) Kofi og na Paramaribo gi mi. (Sranan; Jansen et al.

Kofi go to? Paramaribo GIVE me 1978 in Holm 1988)
'Kofi went to Paramaribo for me.'
(14) Mi wroko gi en. (Sranan; Voorhoeve 1975

I worked GIVE him in Lord 1993)
Ί worked for him.'
(15) Ali tono koro la a naa. (Dagaare;

Ali work+imp GlVE+imp a.m. a.def. chief Bodomo: p.c.)
'Ali works for the chief.'
The item expressing the benefactive meaning in Fula is a productive

verbal extension, while the Akan, Ewe, Sranan, and Dagaare items
expressing benefactive are words which are identical to independent
GIVE verbs in the respective languages. The benefactive marker in
Ewe, see (12), although glossed as 'for' and reported to be morpholo-
gically reduced in some serialization contexts (Eli Saether0 Andenes:
p.c.), corresponds phonologically to the GIVE verb. All the mentioned
GIVE verbs are elsewhere in the syntax used as full-fledged inde-
pendent verbs.
Examples (11) and (12i) are instances of complex predicates
where the GIVER and the THING are identical to the subject and direct
object referents of the host verb. GIVE introduces a new participant
212 Öle T. Fagerli
on the scene, namely the RECIPIENT, the indirect object referent to

whom the THING is being transferred concretely. The only detail
differing from prototypical literal GIVE in the two sentences is the
lack of physical contact between the GIVER and the RECIPIENT in
(11), due to the nature of the host verb THROW. The second reading
of (12) would normally involve concrete recipiency, but it need not.
It could be that the RECIPIENT is simply a beneficiary of the action,
and that the THING is being transferred to somebody else. A more
accurate translation into English would then be picked a flower on
your behalf.
Examples (13) to (15) are instances where the RECIPIENT is
clearly a beneficiary of the action, no other THING than the action of
the GIVER affecting the RECIPIENT.
It is a controversial issue whether one should call the instances of
GIVE above prepositions or case markers rather than verbs because
their use has been extended to include the beneficiary-of-action type
of benefactive not typically associated with the verb. From the point
of view of many European languages, they also occupy positions
normally associated with prepositions or case markers. It is ascer-
tained that GIVE has developed into uninflected benefactive markers
in some languages, for instance the Senufo language Karaboro (Lord
1993: 42). I believe serial GIVE as in the quoted examples are good
candidates for being grammaticalized, but they still retain too much
of their verbal morphophonological properties in these languages to
deserve other category labels than verbs. For a discussion of category
status of serial verbs, cf. Lord (1993) and Osam (1994).
The linguistic material available presenting GIVE in serializing
contexts is relatively rich, I have presented a selection of data only.
Languages such as Mandarin and Hokkien Chinese, Thai, Viet-
namese, Yao Samsao, Chamorro, Siya, and Yoruba also posit the
same potential: GIVE can be used as a benefactive marker, cf. Lord
(1993: 43) and Newman (1996: 212-223).
7. Benefactive by means of SHOW
GIVE is not the only possible verbal source for benefactive marking;
among others SHOW is used in serializing and compounding
languages to denote benefactive events. In Akan (Twi) it is possible
to use a verb corresponding to SHOW in such circumstances:
(16) wo-tow tuo kyere borohene. (Twi; Christaller 1881

they-fire gun SHOW governor in Lord 1993)
They fire guns in honor of the governor.'
(17) ama kd-d asem kyere-ε kofi. (Akan; Saether0:

Ama say-PST matter SHOW-PST Kofi p.c.)
'Ama told Kofi something.'
According to Eli Sasther0 (p.c.) (17) represents a standard way of

expressing TELL in Akan. Schematically, SHOW exhibits the same
type of participants as well as a similar flow of action as GIVE.
However, I will not pursue the SHOW case further in this study.
8. Malefactive by means of serial GIVE
Using GIVE to express malefactive seems like a contradiction. It is

certainly a marked interpretation of sentences like (18) below, where
the beneficiary-of-action type is preferred. However, just like in the
case of verbal extensions in Fula, serial GIVE can serve as a
malefactive marker, as in (19):
(18) meh gbecfuc[o na wo (Ewe; Essegbey:

I-collect rubbish GIVE you p.c.)
(i) Ί collected rubbish for you.' (='on your behalf)
(ii) '?*! collected rubbish to you.' (marked interpretation)
214 OleT.Fagerli
(19) Ο ngma la zirii ko Amai oi yideme yele.

he cut a.m. lies GIVE Ama she housepeople matter
'He lied to Ama about her family.' (Dagaare; Bodomo: p.c.)
(19) shows that the patient is negatively affected by the lies sent out
by the malefactor. It is structurally and semantically similar to the
examples in (11) and (12), and the use of GIVE is perhaps not sur-
prising, although unusual. (20) is different and more surprising in
that the primary verb indicates motion of the THING away from the
male-ficiary, seemingly contradictory to the direction expressed by
GIVE:
(20) Kofi kibri wan sani gi Gado. (Sranan; in

Kofi hide one something GIVE God Boretzky 1983)
'Kofi is hiding something from God.'
Hiding something in this case means to keep it out of reach or out of

sight of someone who normally has access to it. I believe it is right to
apply the general term ablative for this relation, but more specifically
the situation of the patient (God) is malefactive. This is so in spite of
the use of GIVE as a second verb. The construction is very similar to
the ablative examples in (22-23). As previously seen, GIVE implies
movement of some entity from the benefactor to the beneficiary/reci-
pient, in the case of (21) from Kofi to God. This is the opposite direc-
tion of that one signaled by HIDE. This apparent paradox can be dealt
with if we assume that there is movement at two levels. HIDE implies
that the entity here referred to as 'something' moves (or is kept
away) from someone, while GIVE has the function of showing that
the action of hiding affects God. It is merely an affectedness marker
in this case. Affectedness may be extracted as a subpart of the
meaning of GIVE, involving a flow of action from an energy source to
the affectee. Any idea of benevolence as an attribute to GIVE is
suppressed, and this is exactly what happened in (19), too. Thus this
is a maleficiary-of-action sort of construction, a negative counterpart
of the beneficiary-of-action type discussed in previous examples.
Pike (1970: 10) cites an example from Bariba, conforming to the

HIDE type above:
(21) ba kuku - na (tonal morpheme for "benefactive")(Pike 1970: 10)

They hid for each other (from each other).'
Examples of this kind of malefactive or adversative are not easy

found, languages using GIVE in serial verb constructions do not al-
ways go as far as to allow for semantic extension of GIVE to the
extent necessary in malefactive constructions. In Akan one has not
encountered serial ma 'GIVE' in malefactive constructions of the TELL
LIES or HIDE type (Saether0 Andenes: p.c.). They would rather be
interpreted in the beneficiary-of-action sense. In Yoruba, however,
one can find a clear structural and semantic parallell in the ablative
examples below. Ablative is normally used to denote an in-flective
case relation which indicates the source of a movement. Fun (fu) is
the word for GIVE in the language. By the grammarians it is glossed
as 'from' below, evidently because it suits the reading of the sentence
better than GIVE or FOR:
(22) nw5 sa fü u (Yoruba; Rowlands 1976

she run-away from him in Boretzky 1983)
(23) mo yera fü u (Yoruba; Boretzky 1983)

I myself withdraw from him
First of all, the complex predicate here lacks the usual kind of THING
participant associated with the GIVE scene. Syntactically there is no
direct object, if we assume that the postverbal part functions as an
oblique object. Furthermore, it is again striking to see the contrast in
energy flow comparing the host verb and GIVE in this context. In an
utterance with GIVE we expect movement of something starting with
the subject referent of the sentence, i.e the energy source, and having
the direct object referent as the endpoint. Fun can be used that way
too in Yoruba; as an independent verb and as a marker of recipiency
or benefactive in complex predicates. But in (22) and (23) the subject
216 OleT.Fagerli
referent moves away from the object referent. It is clear that the
change in direction is due to the verbs sa and yera which imply
motion away from the direct object. This does not necessarily mean
that the ordinary attribute of direction (to beneficiary/recipient)
associated with GIVE is suppressed, but that there are two levels of
direction; the agent's motion away (from someone), indicated by the
first verb, and the connection this motion has to a specified patient of
the motion, indicated by fü. Fü here has a character more like an
ablative case marker than a full-fledged verb. The semantics of these
sentences is similar to that of the ones marking malefactive affected-
ness above. Whether the object referent is negatively affected by the
act however, such that we could apply the term malefactive, is uncer-
tain from the scarce data I possess. My knowledge of Yoruba does
not go far beyond the examples quoted here, and that keeps me from
drawing any conclusion.
(24) below may also be read in the ablative sense because
something is being moved away from the patient.
(24) abai di wosu da mi (Saramaccan; Grimes/Glock

he-buy the house GIVE me 1970 in Boretzky 1983)
'He bought the house from me.'
We would expect GIVE in this case to have a benefactive/recipient

reading, like in (11-12), thus meaning 'He bought a house for/to me'.
Allegedly, Saramaccan uses GIVE differently in this case from the
other serializing languages we have looked at. GIVE here has lost its
idea of benevolence, and probably functions more like a marker of
affectedness. Despite the inherent ablative characteristics of BUY,
there is no real negative affectedness associated with the act. The
patient is nevertheless affected by the buying act, in the sense that he
loses control over a thing.
Thus, the forward (allative) direction normally indicated by GIVE
needs not stand in conflict with the ablative direction indicated by the
verbs HIDE, WITHDRAW, and BUY. The malefactive and ablative
interpretations of the sentences are due to these verbs.
9. The polysemy of GIVE morphemes
We may consider both the serial GIVE and the benefactive extension
approach at coding the related types of events described in the pre-
vious sections as special instances of a cognitively and semantically
basic GIVE scene. Note that the uses of GIVE in the serializing langua-
ges encountered here do not complete the picture of polysemous
structures for GIVE in the respective languages. We have only seen a
few examples of occurences of GIVE in a limited syntactic environ-
ment, that of serial verb constructions. And this fact would naturally
limit the range of potential semantic extensions of the verb as a
whole. On the other hand, it has been shown that in the serial verb
construction, where two or more verbs share semantic and syntactic
structures, GIVE tends to be viewed as an affectedness marker of the
complex predicate. With the help of the meaning of the host verb or
its complement, GIVE can be viewed as a malefactive or ablative
marker. As far as I know, and plausibly on semantic grounds, GIVE as
a simple predicate can never mark malefactive or ablative.
In Newman (1996: 233, 249) an overview of possible figurative
extensions of GIVE typologically is presented within the framework
of Cognitive Grammar. Newman in general considers both GIVE as
an independent full-fledged verb and instances of dative case, serial
verbs and directional prepositions. Benefactive and allative verbal
extensions are not included, however. The main categories of
semantic extensions are as follows:
(i) interpersonal communication (give somebody a message)

(ii) emergence / manifestation (give good crops)
(iii) causation / purpose (focus on the GIVER'S role as volitioner or causer)
(iv) enablement (enable somebody to do or be in control of something)
(v) schematic interaction (agent-patient interaction, give the car a wash)
(vi) recipient / benefactive
(vii) movement (into or away)
(viii) completedness (relatable to inherent perfectivity of GIVE)
218 OleT.Fagerli
It would take us too far from the scope of this study to comment on
all of these semantic extensions, suffice here to refer to Newman
(1996) for a detailed analysis. However, the absence of the malefac-
tive in the list deserves some comments. Malefactive would certainly
not count as a main category of semantic extension of GIVE, I have
argued that it belongs to an overall recipient category, thus falling
into group (vi) above. Newman counts harmful things given to
recipients as atypical of the GIVE scenario, so no specific status is
given to that kind of malefactive. But he notes that the dative case in
for instance German, Czech, and Polish may encode both benefactive
and malefactive effects on the recipient (Newman 1996: 115-117).
The dative is also an issue in Janda's (1997) article on GIVE, HAVE,
and TAKE in Slavic. Experiencer is a cover term used by Newman
and Janda to denote the beneficiary and the maleficiary. Whichever
the label, what is shown to be the case is that an apparent opposition
in meaning can be given a natural semantic explanation. The GIVE
morpheme by itself never seems to mark malefactive, it is the context
in which it appears which triggers the overall malefactive reading.
Schematic meaning of
entity to which an act
is directed
Schematic meaning of Schematic meaning of

person affected by goal to which move-
an act ment is directed
Affectee not Affectee involved Goal at end of path

involved in act in act as recipient
of concrete thing
The lizard ran towards
the tree
Beneficiary Maleficiary Beneficiary
of act of act of recipiency
Abbo works for me Abbo lied against me 1 picked a mango I picked a rotten
He took it from me to her mango to her
Figure 2. Schematic network of Fula -AN-

What about allative (goal)? In most cases, this function is not

considered a figurative extension of GIVE. It is rather a semantic sub-
part of literal GIVE, both participants, the THING moving and the
GOAL to which it moves, are usually concrete (whether animate or
inanimate).
The verbal extension in Fula has been given a fairly complete
treatment of polysemous structures. The meanings related to the -AN-
suffix are summarized in the schematic network in Figure 2. Serial
GIVE as we have looked at it, would follow approximately the same
network of meanings.
I have attempted to present a rather fine gradation of how different
malefactive and benefactive events are distributed in the relatedness
network. The maleficiary-of-action type could be further divided by
two different kinds of verb meanings involved in the complex predi-
cate, one like TAKE, which is of the ablative kind, and another like
LIE, which is inherently an adversative act. Somehow, literally or
figuratively, all these subcategories of -AN- are instantiations of the
recipiency part of GIVE events.
10. Additional remarks
We have seen that what is in the unmarked cases a benefactive

marker, whether it is a verbal extension or serial use of GIVE, can
extend its semantic space in appropriate circumstances to denote
malefactive situations. To a large extent the two differing structural
types, serial verb constructions and verbal extensions, match each
other semantically. So far I have not come across serial GIVE in
expressions of allative the way it may be used in Fula. It is perhaps
not plausible to think of the verb corresponding to GIVE in such
languages as the most natural source for denoting an event of motion
towards a goal. Serializing languages rather apply directional verbs
like GO or BE DIRECTED TO for this purpose.
Actually, Fula dialects may apply the verb GO in serial con-
structions. Although not used productively as a benefactive marker,
220 le T. Fagerli
various expressions of allative as well as concrete recipiency can

readily be found:
(25) Mi neldi derewol yah-ugo Nigeria. (Fula)

I send-VAP sheet GO-INF Nigeria
Ί sent a letter to Nigeria.'
(26) Mi winnd-ii leeter faa-de e Aminata. (Pulaar;

I write-TNS letter GO-INF PREP Aminata Sylla 1979)
Ί wrote a letter to Aminata.'
(Pulaar = Fula of Senegal)
The beneficiary-of-action type and malefactive is not expressed

through GO. We would not single out the grammaticalization
potential of the verb, however. In a historical perspective, it is
reasonable to hypothesize that the source for the destinative
extension in Fula could be some directional verb like GO rather than
GIVE. This may explain the difference in semantic function between
the extension in Fula and serial GIVE in other languages. The allative
function would probably not be traced back to GIVE, while it is easy
to imagine that it could originate in a former verb for GO.4 GO is
presumably more basic conceptually than GIVE.
The polysemy of the destinative extension and of serial GIVE does
not end there, though. A matter of further investigation is the re-
lationship between benefactive and causative markers, both in seria-
lizing and extensional languages.
It is clear that the word for GIVE in several languages is used to
express causative relations. In Newman's list of main categories of
semantic extensions of GIVE in the above section, causation is one of
them. Examples (27)-(28) from Akan show the word ma 'GIVE' used
with a causative function. In (28) it is even used twice in one
utterance, first with a causative (manipulative) effect, then with a
clear beneficiary-of-action meaning.
Additional examples are available from Akan in Saether0 (1997)
and from diverse languages cited in Newman (1996: 17Iff.). Exam-
ples of a combined causative and benefactive verbal affix are
discussed for Ngiti and Lendu by Kutsch Lojenga (1997), and for
Hausa by Mark Lobben (conference paper, June 1998).
(27) Kofi ma-α abofra no bo-o famu.

Kofi GIVE-PST child def fall (=hit)-PST ground(in)
'Kofi made the child fall down.' (Akan; Saether0 1997)
(28) o-ma-a me ye-ε adwuma ma-α no.

s/he-GIVE-PST me do-past work GIVE-PST him
'He made me work for him.' (Akan; Saether0 1997)
No direct parallels are found with benefactive extensions in Fula of

today. However, Endresen (1994) has looked at the etymological
relationship between the Fula benefactive and causative suffixes, -AN-
and -(I)N-, respectively. Phonologically, they could have the same
origin, the vowel in the -Vn- suffix being subject to vowel assimila-
tion. Endresen's grammaticalization hypothesis implies that an alla-
tive element has developed in stages to include first the recipient
function, then benefactive, and lastly the causative function. I assume
that the malefactive function as described in this paper accompanies
the recipient and benefactive stages.
Notes
1. Capitalized letters are used to represent basic-type categories, such as the verb
GIVE as well as participant categories within the verb.
2. The term extension in this paper is used both to denote verbal extensions as a
morphological phenomenon in some languages and semantic extension. The
context will hopefully clarify any potential confusion caused by the use of this
(polysemous) term.
3. Broken Fula in Aadamaawa (or pidginized Fula, sometimes referred to as
Bilkiire) tends to use the form -in- instead of -an-, a linguistic variant not noted
earlier by grammarians as a systematic one. The -in- suffix coincides with the
causative suffix -(i)n-, a matter of confusion to learners of Fula. Goal, however,
is never rendered by any other form than -an-.
222 Öle T. Fagerli
4. According to Eli Saether0 Andenes (p.c.), Akan uses serial ma 'GIVE' in cases
where the recipient is a person (recipient of thing), e.g. 'throw to Ama', while
the verb Jfo 'go' is used for inanimate goals.
The Norwegian and Russian reflexive-middle-
passive systems and Cognitive Grammar
Hans-Olav Enger and Tore Nesset
The paper presents an extensive analysis and comparison of the

Norwegian and Russian reflexive-middle-passive systems, drawing
especially on Kemmer (1993), but presenting some improvements.
Further, the paper shows why indirect reflexives with beneficiary
and recipient participants are best analysed as two different
situation types. With reference to the Norwegian distinction between
seg and seg selv, the term "superheavy marker" is introduced. It is
shown how the categorisation networks of Cognitive Grammar
(Langacker 1987α, 1990, 1991) are valuable for the typological
study of voice systems.
1. Introduction: Aims and organisation
This paper has both an empirical and a theoretical aim. The former is
to present an extensive analysis of the Norwegian and Russian
reflexive-middle-passive systems. In this respect we draw on
Kemmer's (1993) analysis of Norwegian and Janda's (1993) of
Russian. However, our account goes beyond these in that we
consider a wider range of data. In particular, we consider the
important distinction between Norwegian seg and seg selv, which has
been amply studied in generative work on Scandinavian syntax (cf.
e.g. Dalrymple 1993; Hellan 1988; Hestvik 1992), but which
Kemmer has chosen not to discuss.
The middle voice has been a topical research problem in cognitive
linguistics (cf. e.g. Janda 1993; Maldonado 1988, 1993; Manney
1990, 1992, 1995, 1998). Whereas the focus in most of these works
has been on particular properties of the middle voice in particular
224 Hans-Olav Enger and Tore Messet
languages, our theoretical aim is to show how categorisation

networks of the type employed in Cognitive Grammar, cf. Langacker
(1987a, 1990, 1991) provide a valuable tool for typological research.
Our analysis yields two implications for the typological study of
voice systems:
• We improve Kemmer's terminology by introducing the term

"superheavy marker".
• We demonstrate why indirect reflexives with beneficiary
and recipient participants are best analysed as different
situation types.
The paper is organised as follows. After presenting the relevant data

in section 2, we establish categorisation networks for the two
languages in section 3 and section 4. The former focuses on
extension relations, the latter on instantiation relations. Section 4
concludes and summarises the contributions of the present paper.
1. The Norwegian and Russian systems: Situation types and

markers
In what follows we shall explore 18 situation types and see how they
are marked in Norwegian and Russian. This provides us with the data
to be accommodated in terms of Cognitive Grammar categorisation
networks in section 2 and section 3, but it also allows us to make two
points of more general interest for typology. First, we shall improve
Kemmer's (1993) terminology by advancing the term "superheavy
marker". Secondly, we shall argue that reflexives with beneficiary
and recipient markers represent different situation types, a question
that is left open by Kemmer.
We shall discuss the functions of the Norwegian and Russian
pronouns seg (selv) and sebja and the related affixes -s and -sja.1 We
shall refer to them jointly as "reflexive-middle-passive markers". In
keeping with Kemmer's (1993) terminology, we shall label the
pronouns "heavy markers" (H) and the affixes "light markers" (L)
Reflexive-middle-passive 225
since the pronouns display more phonological body (cf. Kemmer

1993: 25). The inventory of forms is set out in Table 1. As indicated
by the table, Norwegian seg is in some sentences (e.g. (1) below)
accompanied by selv, whereas in other cases it occurs in isolation (cf.
e.g. (7) below). In what follows we shall refer to both seg and seg
selv as "heavy markers", but when it is necessary to distinguish
between them, the former will be referred to as a "bare heavy
marker" and the latter as a "superheavy marker". We believe this to
be a significant improvement to existing terminology. Kemmer
(1993), who does not treat the distinction between Norwegian seg
and seg selv, lacks appropriate terminology. Admittedly, she
discusses e.g. Surselvan sesez under the heading "emphatic
reflexive", but this term is unfortunate from a methodological point
of view. In trying to clarify the various meanings or functions of
certain forms, it is essential to apply a terminology that itself refers to
form only. Otherwise, one runs the risk of circularity in taking the
relation for granted before it is established on independent grounds.
The term superheavy has the advantage of being consistent with
Kemmer's heavy and light marker.
Table L Markers in Norwegian and Russian

Heavy marker (H) Light marker (L)
Norwegian seg (selv) -s
Russian sebja -sja
Our analysis relies on the notion of "situation type", which Kemmer

(1993: 7) defines as "sets of situational or semantic/pragmatic
contexts that are systematically associated with a particular form of
expression". Kemmer (1993: 7) emphasises that "semantic/pragmatic
contexts" are not to be equated with "simple 'real-world contexts'
existing independently of the language-user; situational contexts
include real-world information, but that information is necessarily
filtered through the conceptual apparatus of the speaker. This
conception of situational contexts thus allows for the obvious role of
the language-user in construing particular 'real-world' situations in
different ways (cf. ... Langacker 1987a) [emphasis original]."
226 Hans-Olav Enger and Tore Nesset
Table 2 summarises the situation types and markers to be

discussed in the present study. The list of situation types is adapted
from Geniusiene (1987), Pederson (1991), and Kemmer (1993), but
we exclude types not attested in either language.
Table 2. The uses of Norwegian -s and seg and Russian -sja and sebja
Norwegian Russian
1. Prototypical reflexive HH H
2. Indirect reflexive (recipient) HH H
3. Indirect reflexive (beneficiary) H H
4. Natural collective H L
5. Natural reciprocal — L
6. Autocausative H L
7. Grooming H L
8. Reflexive causative H L
9. Converse reflexive H L
10. Mental events H L
11. Spontaneous events H L
12. Modal-deagentive — L
13. Subject demotion — L
14. Object demotion — L
15. Absolutive — L
16. Quasi-passive L L
17. Reflexive passive L L
18. Impersonal L —
HH: Superheavy marker (seg selv) H: Heavy marker (seg, sebja)
L: Light marker (-s, -sja) —: No marker attested
We are not going to discuss the characteristics of each situation type

here, as they are already covered in the works cited above (see
especially Kemmer 1993, chapters 3 and 4). However, representative
examples from Norwegian and Russian are presented below for each
situation type. A familiar problem is how to distinguish one situation
type from another. We endorse Kemmer's (1993: 5ff.) approach
which is to let cross-linguistic data decide. Thus, if one wonders
whether what Kemmer refers to as "grooming" and "reflexive"
represent one or two situation types, one should examine whether
they are marked differently in some languages. Given that they are in
e.g. Russian (Ivan moetsja 'Ivan washes' vs. Ivan nenavidit sebja
'Ivan hates himself), one may conclude that one is dealing with
different situation types. There may be practical problems involved
in Kemmer's approach, but it nevertheless provides a well-motivated
criterion.
Examples for the attested situation types are given in (1) through
(18).2 Notice that the numbering of the examples corresponds to the
numbering of situation types in Table 2. Consider first situation types
1 to 3, prototypical and indirect reflexives, the only situation types
where Russian has the heavy marker sebja. Norwegian evinces the
superheavy marker in 1 and 2, but the bare heavy marker in 3. In the
typological literature, indirect reflexives with recipient and
beneficiary participants are normally regarded as subtypes of one and
the same situation type. However, Kemmer (1993: 74-75) notes that
recipients and beneficiary participants normally are typologically
distinct, and hints at the possibility that "[c]loser typological
investigation may result in the necessity of treating these two
situation types as distinct". For the purposes of her study, however,
Kemmer treats the two as one type. Given that recipient and
beneficiary NPs are marked differently in indirect reflexives in
Norwegian, we submit that they must be considered different
situation types according to the criterion for delimiting situation
types discussed above. Thus our analysis allows us to answer a
question left open by Kemmer (1993).
(1) PROTOTYPICAL REFLEXIVE

a. Jeg elsker meg selv og har ingen rivaler.
I love segselv.l.SG and have no rival:PL
Ί love myself and have no rivals.' (Fredrik Stabel)
b. On nenavidit sebja.
he:NOM hate:3.SG sebja:ACC
'He hates himself.' (Haiman 1983: 804)
(2) INDIRECT REFLEXIVE (RECIPIENT)

a. Han gav seg selv en bok.
he gave seg selv a book
'He gave himself a book.'
b. On podaril sehe knigu.
he:NOM gave:MASC.SG sebja-.ΌΑΎ book:ACC.SG
'He gave himself a book.'
(3) INDIRECT REFLEXIVE (BENEFICIARY)

a. Jeg vil bygge meg en g rd.
I will build seg:l.SG a farm
Ί shall build myself a farm.' (children's song)
b. On kupil sebe dom.
he:NOM bought:MASC.SG sebja-.ΌΑΎ houserACC.SG
'He bought himself a house.'
We now turn to situation types 4 through 11, which differ from the
previous in that Russian evinces the light marker -sja instead of the
heavy sebja. Norwegian uses the bare heavy marker.3
(4) NATURAL COLLECTIVE

a. Slektningene samlet seg rundt bruden.
relative.the:PL gathered seg around bride.the
The relatives gathered around the bride.'
b. Sobralos' mnogo znakomych.
gathered: 3. SG.sja many acquaintance :GEN.PL
Ά lot of acquaintances gathered.' (Usakov 1935)
(5) NATURAL RECIPROCAL

a. (not attested)
b. Korabli vstrecajutsja u vchoda
ship:NOM.PL meet:3.PL.5/a at entrance:GEN.SG
v gavan'.
in harbour:ACC.SG
'The ships meet at the entrance of the harbour.' (Usakov 1935)
(6) AUTOCAUSATIVE
a. Jon reiste seg motstrebende,
Jon arose seg reluctantly
'Jon arose reluctantly.' (Hellan 1988: 114)
b. On lozilsja v poster.
he:NOM lay:MASC.SG.5/a in bed:ACC.SG
'He lay down on the bed.' (Usakov 1935)
(7) GROOMING
a. Jeg vasker meg.
I wash seg:l.SG
Ί wash.'
b. Ja kazdyj den' brejus'.
I:NOM every:ACC.SG day.ACC.SG shave:l.SG.sja
Ί shave every day.'
(8) REFLEXIVE CAUSATIVE

a. Har du kuppet de g hos Nilsen Fris0r?
have you:SG cut seg:2.SG at Nilsen Fris0r
'Have you had your hair cut at Nilsen Fris0r?'
b. Ona operirovalas' v novoj klinike.
she:NOM operated:3.SG.sja in new:LOC.SG clinic:LOC.SG
'She had an operation at the new clinic.' (Janda 1993: 313)
(9) CONVERSE REFLEXIVE

a. Trtzrne speiler seg i vannet.
tree:PL mirror seg in water.the
The trees are mirrored in the lake.'
b. Nebo otrazaetsja v ozere.
sky:NOM.SG mirror:3.SG.sja in lake:LOC.SG
The sky is reflected in the lake.' (Geniusiene 1987: 228)
(10) MENTAL EVENTS

a. Alle barn gleder seg til jul.
all children look-forward seg to Christmas
'All children look forward to Christmas.'
b. / vot teper' my vse raduemsja

and here now we:NOM all:NOM.PL are-happyil.PL.sja
i veselimsja,
and enjoy 1.PL.sja
'And now we are happy and enjoy ourselves.' (Usakov 1935)
(11) SPONTANEOUS EVENTS

a. D0ren pnet seg langsomt.
door.the opened seg slowly
The door opened slowly.' (Hellan 1988: 113)
b. Dveri zakryvajutsja.
door:NOM.PL close:3.PL.5/a
The doors are closing.' (Janda 1993: 313)
Russian displays the light marker in examples (12)-(15). No marker

is attested in Norwegian; the Russian system is somewhat more
elaborate. Notice the dative case in the experiencer in (12) and (13),
and the use of genitive in (13) and instrumental in (14).
(12) MODAL-DEAGENTIVE
a. (not attested)
b. Mne ne rabotaetsja.
I:DAT not work:3.SG..s/a
Ί cannot work.' (Geniusiene 1987: 289)
(13) SUBJECT DEMOTION

a. (not attested)
b. Mne chocetsja orechov.
I:DAT want:3.SG.s/<2 nuts:GEN.PL
Ί want nuts.'
(14) OBJECT DEMOTION

a. (not attested)
b. On brosaetsja kamnjami.
he:NOM throw.'S.SG.^'a stone:INST.PL
'He throws stones.' (Geniusiene 1987: 48)
(15) ABSOLUTIVE
a. (not attested)
b. Sobaka kusaetsja.
dog:NOM.SG bite:3.SG.J/a
The dog bites.' (Isacenko 1982: 460)
In types 16 through 18, Norwegian shifts from heavy to light marker;

both languages evince light markers. However, the impersonal is not
attested within the Russian reflexive-middle-passive system, cf. (18).
(16) QUASI-PASSIVE
a. Det h0res musikk fra nabohuset.
it:NOM hears music from house-next-door.the
'Music is heard from the house next door.'(Berkov 1989: 154
b. Nasa masina choroso zavoditsja.
our:NOM car:NOM well start:3.SG.5/'a
Our car starts well.' (Janda 1993: 314)
(17) REFLEXIVE PASSIVE

a. Lederen skrives av redakt0ren.
editorial.the write.* by editor.the
The editorial is written by the editor.' (Dyvik 1980)
b. Takie knigi citajutsja sirokimi
such:NOM.PL book:NOM.PL read:3.SG..s/a broad:INST.PL
massami.
mass:iNST.PL
'Such books are read by the broad masses.'(Janda 1993: 313)
(18) IMPERSONAL
a. Det danses,
it dance.s
'It is danced.'
b. (not attested)
2. Extension relations: Networks for Norwegian and Russian
We shall now accommodate the data considered above in Cognitive

Grammar networks with Schemas interrelated by extension relations
(see especially Langacker 1987a: 369-386, 1990: 261-272). We shall
emphasise the heuristic value of such networks. Not only do they
allow us to account for the relevant data about each language, they
also facilitate cross-linguistic comparison.
The networks in Figure 1 are based on semantic maps from
Kemmer (1993: 202) and, especially, Pederson (1991: 92).4 The
network to the left represents Norwegian and the right one Russian.
Each schema in the networks corresponds to one situation type. As is
usual in Cognitive Grammar, the extension relations that connect the
Schemas are symbolised by dotted arrows. Three boxes are not
connected to any other box because they are either not included or
represented as unconnected in Pederson's (1991: 92) network.
The merits of the networks can be summarised in four points. First
of all, they allow us to account for the relations between the relevant
situation types. As can be seen from Figure 1, we assume them to
constitute a radial category with "prototypical reflexive" as the most
central subcategory. The distance on paper reflects iconically the
conceptual distance between a certain subcategory and the prototype.
Secondly, by the use of shading and heavy, solid and dashed lines
in the boxes that represent the various situation types, the networks
account for the distribution of the various markers in the two
languages. We designed the notation for the purposes of this paper.
Thirdly, as a consequence of the two points already mentioned,
networks like those in Figure 1 enable us to formulate empirical
hypotheses about the conceptual structure of the reflexive-middle-
passive category. The Norwegian collective is a case in point. As can
be seen from Figure 1, it is not directly related to the rest of the
reflexive-middle-passive category, because the intermediate situation
type of reciprocals arguably does not display the heavy marker seg in
Norwegian. Given that the network is a potential model of the
conceptual structure of a category, one would expect seg to occupy a
spatially coherent area in the network. But, as we have seen, this is
not the case for the collective. One might then think of various
possibilities. For example, the spatial incoherence might in principle
indicate a situation where a diachronic change is likely to occur,
although this is improbable in this particular case, since similar
patterns are attested in other languages without showing any signs of
diachronic instability (Kemmer 1993: 100).
Norwegian: Russian:
Λ. ,
object reflexive object reflexive
demotion causative demotion causative
3r 3!
modal- spontaneous modal- spontaneous
deagentive «·- events -* absolutive deagentive «-- events -> absolutive
1 F
t t *
t
subject- quasi- subject quasi -
demotion passive demotion passive
1
I Superheavy marker Heavy marker Light marker No marker
Figure 1. Networks for Norwegian and Russian
An alternative might be to alter the network proposed in Figure 1,

e.g. by assuming a direct connection between collective and reflexive
situations. We shall not pursue this issue; our point is to illustrate the
heuristic merits of Cognitive Grammar networks; they force us to
reconsider linguistic data and improve our analyses.
Taken together, these three points suggest that extensive analyses
of the reflexive-middle-passive system of each language are possible
in terms of Cognitive Grammar networks. In addition, however,
Figure 1 illustrates the value of such networks in typological
research. The juxtaposition of the networks for Norwegian and
Russian brings out the similarities and differences between the two
languages in a simple and straightforward way, and thus facilitates
cross-linguistic comparison.5
4. Instantiation relations: Reflexive, middle and passive
The networks presented in the previous section are very similar to the
semantic maps employed by typologists like Pederson (1991) and
Kemmer (1993). In this section we shall see that Cognitive Grammar
goes beyond the typological tradition, in that the instantiation
relation allows us to integrate the cover terms reflexive, middle and
passive in the networks. This enables us to explicate the similarities
and differences in the distribution of the markers in the two
languages with regard to the cover terms. Once again, then, the
Norwegian and Russian evidence supports our general claim about
the value of Cognitive Grammar networks in typological research.
Sentences (19)-(21) (also listed as (Ib), (7b) and (17b) above)
deserve further discussion; they provide good examples of reflexive,
middle and passive voice. Following Kemmer (1993: 201ff.) we may
characterise these three cover terms by means of the five features
listed in Table 3. The three situation types are united in that the chief
participant is directly affected by the verbal action. In the case of
reflexives and middles like (19) and (20), the affectedness is due to
the agent subject being identical to the patient of the action, whereas
in the case of the passive in (21), the patient is the subject of the
sentence.6
(19) PROTOTYPICAL REFLEXIVE

On nenavidit sebja.
'He hates himself.' (Haiman 1983: 804)
(20) MIDDLE (GROOMING)

Ja kazdyj den' brejus'.
Ί shave every day.'
(21) PASSIVE
Takie knigi citajutsja sirokimi massami.
'Such books are read by the broad masses.' (Janda 1993: 313)
The second feature, "degree of elaboration of events", is introduced

by Kemmer (1993) in order to distinguish between middle and
reflexive voice. The intuition behind this feature is that reflexive
situations like hating are situations that are normally directed
towards other persons or entities, whereas for middle situations like
shaving direction towards oneself is the expected case. Since in
middle situations the action is normally directed towards oneself, the
two participants involved are distinguished to a lesser degree. The
less the participants are distinguished, the less the situation is
elaborated. Kemmer arranges reflexive and middle situations on a
cline ranging from transitive two-participant events to intransitive
one-participant events. Reflexive situations are placed in the middle
of the cline and are accordingly assigned an intermediate degree of
elaboration, while middle situations are placed between reflexive and
intransitive situations and hence show low degree of elaboration of
events. (Kemmer (1993: 205) considers the passive to be "midway
between a two- and a one-participant event", and we have therefore
assigned the value "intermediate" as far as the degree of elaboration
of events is concerned.)
Features c) and d) set the passive apart from the middle and
reflexive. As opposed to these voices, the passive has an external
agent insofar as the subject and the adjunct in the instrumental case
are not coreferential in (21). Carrying a patient role the subject of
passive sentences like (21) is obviously not volitional. The last
feature, which concerns the salience of the participant that "might be

conceived of as the originator of the event" (Kemmer 1993: 205),
unites the passive and middle as opposed to the reflexive. Passive
sentences like (21) involve demotion of the agent from subject to
adjunct, and at least certain middle situations, e.g. those of emotion
and spontaneous events, defocus the originator of the event in a
similar way (cf. Kemmer 1993: 205).
Table 3. Characterisation of reflexive, middle and passive

Reflexive Middle Passive
a) Dir. affectedness of chief part.: + + +
b) Deg. of elaboration of events: intermediate low intermediate
c) External agent: +
d) Volitionality of chief part.: + + -
e) Salience of originator of event: high low low
If we apply the features to the data in Table 2, situation types 1-3

represent reflexive voice, 4-15 middle, and 16-18 passive.7 Given
this, we are in a position to relate the use of the heavy and light
marker in the two languages to the functions of reflexive, middle and
passive. This is done in Table 4.
Table 4. The marking of reflexive, middle and passive

Norwegian Russian
Reflexive: HH H
Middle: H L
Passive: L L
On the basis of this table, generalisations about the relation between

meaning and form can be made. If we take meaning as the starting
point, we can state that both languages mark reflexive by a heavy
(superheavy or bare heavy) marker and passive by a light marker, but
that the languages differ in the marking of the middle voice, insofar
as Norwegian displays a heavy and Russian a light marker. If we take
form as the starting point, we can say that in Norwegian the super-
heavy and bare heavy markers are reflexive and middle markers and
the light marker a passive marker, whereas in Russian the heavy mar-
ker is a reflexive marker and the light a middle and passive marker.
Having considered the relationship between the situation types
and the cover terms, we must ask how this relationship can be
modelled. The typological approaches of Pederson (1991), Kemmer
(1993) and others do not address this question, insofar as their
semantic maps do not include explicit reference to the cover terms.
Cognitive Grammar, on the other hand, offers a simple and
straightforward answer to the question; the cover terms can be related
to the situation types by means of instantiation and thus be
incorporated in the networks. This is shown in Figure 2, where the
instantiation relations are represented as solid arrows. In this network
"reflexive", "middle" and "passive" represent Schemas that each sub-
sume a number of the functions from the networks given in Figure 2.
The figure is simplistic in that the schema "middle" subsumes more
than the two Schemas "autocausative" and "grooming" and that the
extension relations are excluded in order to avoid crossing lines. The
latter problem would not arise in a three-dimensional representation
where situation types and cover terms would constitute different
planes connected by instantiation relations. However, the simplified
two-dimensional Figure 2 suffices to illustrate that Cognitive
Grammar offers a straightforward modelling of the relationship be-
tween situation types and the terms "reflexive", "middle" and
"passive". This way Cognitive Grammar provides a welcome addi-
tion to the notional apparatus employed in the typological tradition.
REFLEXIVE
PROTOTYPICAL IND. REFL. IND. REFL. QUASI- REFLEXIVE IM-

REFLEXIVE (RECIP.) (BENEF.) PASSIVE PASSIVE PERSONAL
Figure 2. Instantiations of reflexive, middle and passive
The instantiation relation also allows us to explicate the similarities

and differences in the distribution of the markers in Norwegian and
Russian with regard to the Schemas of "reflexive", "middle" and
238 Hans-Olav Enger and Tore N esset
"passive". If we let these notions be represented by the properties

given in Table 3, we are in a position to characterise the two systems
in terms of the networks in Figures 3 and 4.
dir. affec.
non-high elab.
no extern, ag.
volit. chief part.
... seg...
Figure 3. Reflexive, middle and passive in Norwegian
The mid level in the networks brings out the differences between the
Norwegian and the Russian reflexive-middle-passive systems; in
Norwegian, the middle goes with the reflexive, whereas in Russian it
goes with the passive.8
We see that not only does the instantiation relation provide a
means for modelling the relationship between "reflexive", "middle"
and "passive" and situation types; it also facilitates cross-linguistic
comparison. This again illustrates the value of Cognitive Grammar in
typological studies.
5. Conclusion
The contribution of the present paper is twofold. Firstly, we have

presented extensive analyses of the reflexive-middle-passive systems
in Norwegian and Russian. Our analyses have both enabled us to
improve Kemmer's (1993) terminology by introducing the term
"super-heavy marker" and to show why indirect reflexives with
recipient and beneficiary participants are best analysed as different

situation types (a question that is left open by Kemmer).
As a second main point, we have shown how Cognitive Grammar
categorisation networks allow us not only to account for the
reflexive, middle and passive voices in each language, but also to set
out the similarities and differences between the two languages in a
lucid and straightforward way. Therefore, Cognitive Grammar proves
very valuable in typological research.
dir. affec.
non-high elab.
low sal. of orig.
sja
REFLEXIVE PASSIVE
... sebja sja
Figure 4. Reflexive, middle and passive in Russian
Abbreviations:
dir. affec.: direct affectedness of chief participant
non-high elab.: non-high degree of elaboration of events
low sal. of orig.: low salience of originator of event
no extern ag.: no external agent
volit. chief part.: volitional chief participant
Notes
1. Whether -sja is an affix or a clitic is disputed. We follow Schoorlemmer (1997)

and Nesset (1998a, b) who conclude that -sja is best regarded as an affix. It is
worth mentioning that -sja shows some allomorphy, but since allomorphy is
irrelevant for present purposes, the affix will be referred to as -sja throughout
the paper. Norwegian seg is inflected for number and person, but in Table 1
below we give only the 3. person singular/plural form. There is also a
possessive form sin which will not be discussed in the present paper. Full
paradigms are given in Hellan (1988: 60). According to Pederson (1991),
Norwegian -s must be analysed as several homonymous suffixes. However, we
follow Enger (this volume) who argues that it is best considered one
polysemous suffix.
2. When no reference is given, Russian examples are checked with a native
consultant. Some of the Norwegian examples are taken from secondary
sources, but in most cases we rely on our own native speaker competence. In
some examples the Norwegian heavy marker is meg (1st person singular) and
deg (2nd person singular). The Russian light marker in some examples occurs
in its postvocalic form -s'. Throughout the paper, Russian examples are cited in
orthography transliterated according to the Scando-Slavica system.
3. In the case of a restricted and relatively small set of verbs, the light marker is
possible in these situations, but we will not treat such exceptional marking in
this paper. For thorough discussion, see Enger and Nesset (1999).
4. For a smaller and somewhat different network for the Russian light marker, see
Janda (1993: 312).
5. One may speculate whether the similarity between the Russian and
Scandinavian reflexive-middle-passive systems may be due to a Sprachbund
around the Baltic Sea (Dahl and Koptjevskaja-Tamm 1992: 17ff.). Although
this question is interesting in itself, it does not bear on any conclusion to be
drawn in this paper; Cognitive Grammar networks are equally valuable for
comparison of linguistic systems whether their similarities are due to contact or
not.
6. It is debated in the literature whether affectedness really is an appropriate label.
Dimitrova-Vulchanova (1996: 162-163) is skeptical towards Kemmer's notion
of affectedness in connection with the reflexives: "it is highly disputable
whether being affected by the act in question can be assumed for the elements
encoded as subject." However, Dimitrova-Vulchanova accepts Kemmer's point
about "low elaboration". For a discussion of affectedness and low individuation
(low elaboration) from the perspective of cognitive linguistics, see Manney
(1995).
7. Admittedly, this is simplistic in one respect; as can be seen from Table 2,
situation type 3 has a heavy marker in both languages. Thus, this situation type
does not behave like a reflexive in Norwegian. We submit that it represents a
middle situation in Norwegian, but a reflexive in Russian. This move may
appear completely ad hoc, but is in fact supported by empirical evidence. In
Norwegian, these verbs do not occur with a non-coreferential indirect object:
(i) *Jens bygde meg en gärd.

'Jens built me a farm.'
This suggests that indirect reflexives with a beneficiary marker denote actions
normally directed towards the self, and thus qualify as middle, not reflexive in
Kemmer's terms. In Russian, on the other hand, indirect reflexives do occur
with non-coreferential indirect objects, behaving like true reflexives, cf. (ii):
(ii) Ivan postril mnefermu.

'Ivan built me a farm.'
For an extended discussion, see Enger and Nesset (1999: 45-48).

8. Our analysis in figure 5 of -sja as signalling low salience of the originator of
the event is in harmony with an analysis presented by Adger Williams, who
characterises -sja as "a device that enables the speaker to bring an entity into
focus; i.e. to talk about it, without having to be distracted by the mention of
other entities" (Williams 1993: 181).
Norwegian reflexive-middle-passive -s:
Homonymy/polysemy and morphology
Hans-Olav Enger
This paper discusses the Norwegian reflexive-middle-passive marker

-s. This suffix is inflectional when it functions as a passive marker,
and derivational in all other functions. In §2, it is discussed whether
the various s-suffixes are polysemously related or mere homonyms.
The answer is "polysemy". This has some interesting consequences.
It creates problems for many morphological models; apparently, a
"split morphology" must be rejected. Another consequence relates
to the view on polysemy/homonymy propounded by Pederson (1991),
who argues that whenever the direct connection between two nodes
in a network is cut off, homonymy must be posited for those pans of
the network that no longer cohere. On the basis of the Norwegian
case, this principle is questioned, and theoretical arguments against
it are adduced. In §3, the question is raised, why only the passive is
inflectional. It is argued that the other "meanings" of -s (notably
reflexive and reciprocal) are of higher relevance, in the sense of
Bybee (1985). Therefore, it is no surprise that they are derivational.
This explanation is indirectly an argument for Bybee's relevance
concept.1
1. Facts and questions
In Norwegian Bokmäl, the verb suffix -s serves to mark a number of

functions or meanings;
• passive, as in b0ker skrives 'books are written',

• reflexive, as in herrene samles rundt sigarene 'the gentlemen
gather around the cigars', and
• reciprocal, as in vi tales i neste uke 'we'll talk next week'.2
244 Hans-Olav Enger
The suffix -s can be called a reflexive-middle-passive marker (Enger

and Nesset this volume). It is sometimes derivational and sometimes
inflectional. As a passive marker, -s is inflectional and productive. In
all other functions, -s is derivational and fairly improductive. Two
questions arise:
1. Should we posit one polysemous suffix -5, or several

homonymous suffixes?
2. If we prefer polysemy: why is -s inflectional only when it is
passive?
2. Homonymy or polysemy — and some implications
2.L Homonymy versus polysemy
We begin with question 1 above: are the various -s'es polysemically

connected or mere homonyms? The homonymy-polysemy distinc-
tion is notoriously tricky, but we shall consider some standard
criteria.
A first criterion of polysemy is semantic relatedness. Semantical-
ly, there is a relatively close relationship between the reflexive and
the reciprocal (cf. e.g. Langacker 1991: 370-371). The passive is
related to both of these, since in the reflexive and the reciprocal, the
grammatical subject is "both affected and affector", whereas in the
passive, the subject is "affected". So this criterion indicates
polysemy.
We now turn to the second criterion, cross-linguistic comparison:
as Croft (1990: 166) says, "if many diverse languages independently
have the same pattern of "homonymy", then the meanings are
closely related." The idea is that homonymy, being accidental formal
identity, should not recur in many languages. This idea is defended
at length by Haiman (1985). For example, the English preposition to
can be used both to indicate direction towards some location (We
took the bus to Hadeland) and to indicate transfer of possession (He
Norwegian reflexive-middle-passive marker -s 245
gave a book to his father). Now, English also has a homophonous

numeral two. However, it can be shown that the formal identity
between direction towards location and transfer of possession is
quite common, cross-linguistically. By contrast, the formal identity
between direction towards location and the numeral '2' is
uncommon. Therefore, the former case of formal identity qualifies as
polysemy, the latter as phonological homonymy. Now, the passive
and the middle can be expressed by the same suffix in Norwegian, as
they can in Russian (cf. Enger and Nesset this volume), and the
explanation probably has nothing to do with genetic affiliation. This
indicates polysemy.
A third criterion and argument for polysemy is the fact that within
the Norwegian grammatical tradition, verbal -s is (more or less
explicitly) treated as a polysemous suffix (see e.g Bei to 1986 or
Faarlund, Lie, and Vannebo 1997). This may not be a strong
argument in itself, but it is surely relevant.
A fourth criterion and argument for polysemy is methodological.
Other things being equal, polysemy is preferable to homonymy for
two reasons. Firstly, semantic distinctions can be maximised
indefinitely, so unless polysemy is given priority, homonymy can be
posited almost everywhere (Lyons 1977: 550-569). Secondly, to
posit homonymy is to say that a case of formal similarity is a mere
coincidence. To posit polysemy, by contrast, is to say that there is
some reason for the formal similarity. A priori, scientists look for
reasons and do not postulate coincidence unless forced to. Thus,
polysemy should normally be the more attractive hypothesis. This
argument is especially important for those linguists that prefer to
focus on the non-arbitrary aspects of language. Leiss (1997: 136)
even talks of the imperative of a non-arbitrary grammar: Die Form
ist unbedingt ernst zu nehmen! ("The form must absolutely be taken
seriously!")3
The upshot is that the Norwegian verbal suffix -s is a case of
polysemy rather than homonymy. To specialists in Norwegian, this
conclusion is unsurprising, but it has some interesting consequences,
to which we now turn.
246 Hans-Olav Enger
2.2. A morphological problem
Within a Word-and-Paradigm (WP, or realisational) model of mor-

phology (e.g. Anderson 1990, 1992; Matthews 1972, 1991; Aronoff
1994) it is difficult to show the similarity between -s as an inflec-
tional suffix and as a derivational one. The reason is that the two
suffixes are found in different "modules" in the grammar. Within a
WP model, two forms have been created either by one and the same
rule, or by two different rules, and there is no convincing way of
generalising over rules. Thus, the derivational suffix -s is introduced
by a derivational rule, the inflectional suffix -s by an inflectional
rule, and the two implicitly emerge as unrelated, i.e., as homonyms.4
There might be various ways of trying to save a WP model here.
One could be to use the "meta-rules" suggested by Matthews (1991:
201-204). This would amount to postulating a third rule, a rule on a
meta-level, which generalises over the inflectional rule that introdu-
ces -s and the derivational rule that introduces -s. There are prob-
lems with meta-rules, however. They are introduced as an extra de-
vice in order to cope with some problematic examples, and do not
have any independent empirical justification. Furthermore, if linguis-
tic theory strives for psychological plausibility, the meta-rules may
seem ad hoc,5 and a defense that relies on meta-rules is too weak.
Another possible defense of the WP model might rely on mor-
phomes, as we know them from Aronoff (1994) -s might be ana-
lysed as a morphome, essentially an element of form without mean-
ing. This morphome would then be unspecified with respect to the
inflection/derivation distinction (just like Aronoffs (1994: 34-36)
Latin example of a morphome). The problem is that whereas mor-
phomes are meaningless, -s is polysemous (cf. §2.1); it does have a
meaning, and it should therefore not be analysed as a morphome.
The root of this problem may be that WP theorists have ex-
aggerated the distinction between inflection and derivation. This is
understandable, as the distinction is both controversial and valuable.
Yet it should not be considered categorical, i.e., even if we accept
the inflection/derivation distinction, we should not accept a "split
morphology", where inflection and derivation are seen as categori-
cally different (see e.g. Wurzel 1984: 40-51; Bybee 1985: 81-110;
Dressier 1989; Booij 1993; van Marie 1996). It is the sharp and cate-
gorical split between the two kinds of morphology that ultimately
makes it difficult to generalise over the different uses of -s in WP
models.6
A sketchy and incomplete Cognitive Grammar description of
Norwegian -s is given in Figure 1.
direct affectedness of chief

participant
non-high elaboration of events
Figure L Norwegian -s
A more comprehensive study is presented by Enger and Nesset (this

volume), so we need not go into all the details here. However, it may
be worth pointing out that at least some problems can be avoided
within Cognitive Grammar. Firstly, the inflection/derivation distinc-
tion need not be seen as totally discrete (cf. Langacker 1987a: 18).
Secondly, in Figure 1, the labels "reflexive", "reciprocal" and "pas-
sive" denote the semantic content of the suffix. These semantic con-
tents are not identical, but they are fairly similar, and they can be
subsumed under a common label (viz. 'direct affectedness of chief
participant' and 'non-high elaboration of events'). In Cognitive
Grammar, this is described by positing a relation of instantiation.
Thus, it is shown that the three meanings are related without being
one and the same. As for phonological form, all the three look alike,
as indicated by Figure 1. The passive is boldtyped in Figure 1, to
show that it is more entrenched. This description seems promising,
but, admittedly, the inflection/derivation distinction is not directly
brought out in Figure 1.
248 Hans-Olav Enger
Chaining Collective
They followed themselves out of the room. They gathered themselves in the square.
Prototype/ "Dative" reflexives

Coref w/ sub-ord. subject -recipient
He said himself to do it. Semantic ref.
He hit/saw himself -benefactive/affected
-oossessor reflexive
Grooming
He combed himself.
Autocausatives Reflexive causative
- Body posture self-raise / sewed myself a shirt
- Trans, motion self-run at the tailor's.
Non-trans, motion self-bow
Partitive
He buttoned himself up.
Inchoative/
Deagentive/ Converse reflexives
Anticausative The trees reflect themselves
The door opened itself in the water.
Modal-deagentive
The book read itself
quickly by me.
Absolut) ve
Quasi-Passives: Watch out! That dog
bites itself.
Subject Demotion Potential
It doesn 't work itself for me. The books read themselves well.
I
Necessity
The "gh" pronounces itself like an "/".
ψ
Resultative
, The shoes wore themselves out. ι
Reflexive passives:
With external force

The door opened itself from the wind.
With external agent

The door opened itself by me.
Impersonal
Itself signs the document.
Figure 2. Network of potential functional extensions of the reflexive construction

2.3. Networks and homonymy/polysemy
A clear tendency in Cognitive Grammar has been to maximise poly-

semy. It is therefore surprising that according to a principle sug-
gested by Pederson (1991), who is inspired by Cognitive Grammar,
Norwegian -s should be analysed as homonymous. Pederson (1991:
92) posits a semantic map, a network of functions for "reflexive
markers"; see Figure 2.
Furthermore, Pederson (1991: 348) claims that when the con-
nection between two nodes is cut off, we must posit homonymy for
those parts of the network that no longer cohere: "distal functions,
relatable only via now lost central functions, are no longer concep-
tually linked. To wit, these severed distal functions are synchronical-
ly homonymous, despite the clear diachronic motivation for their
expression by the same form." Norwegian -s would be a case in
point, as the so-called semantic reflexive is expressed by seg. In
Pederson's terms, the semantic reflexive is the "now lost central
function". (For further details, see Enger and Nesset this volume).
The implication of Pederson's words, then, is that -s should be
analysed as homonymous, after all.
However, a number of arguments for describing Norwegian -s as
polysemous were given in §2.1 above; and some may be added here:
generally, Pederson a priori denies the possibility of so-called
doughnut categories ever being polysemous, and this may seem
unfortunate, in the perspective of grammaticalisation theory. In the
case at hand, Pederson's argument entails treating the passive -5 and
other verb suffixes -s as homonymous suffixes. They are just as
unrelated as, say, passive -s and the so-called possessive -s that is
added enclitically to nouns in Norwegian (as in Pedersons avhand-
ling 'Pederson's dissertation'). This seems somewhat counter-intui-
tive. More importantly, even if the connection between the passive
and the middle may be weak, it simply does not follow that the con-
nection must be non-existent: there is a difference between "weak
semantic connection" and "no semantic connection". In other words,
Pederson's argument presupposes that the homonymy-polysemy
distinction is a simple binary dichotomy. This is unfortunate; the
250 Hans-Olav Enger
distinction is better seen as a gradient (cf. Lyons 1977: 552;

Langacker 1990: 268).
Enger and Nesset (this volume) argue that even though the
semantic maps as employed by Pederson and others yield valuable
insights, they are somewhat too simple, since they only model
extension relations (i.e., relations between a prototype and its
extensions). This leaves no room for "classical" criterial-attribute
relations, which presumably also matter in categorisation; they are
referred to as instantiation relations in Cognitive Grammar. Hence,
categories may have something in common semantically, even if the
direct connection between them on a semantic map has been cut off.
For example, it is characteristic of all s-forms in Norwegian that the
chief participant is directly affected by the verbal action, no matter
whether the verb is "passive", "reflexive", "reciprocal" or something
else. Even if the common meaning is fairly abstract, this is never-
theless an additional argument against characterising the different
uses of -s as mere homonyms.7
To conclude, the Norwegian verbal suffix -s should be considered
polysemous. However, if we regard the polysemy/homonymy
distinction as gradient, we ought to ask about the degree of polysemy
(Sandra and Rice 1995: 98). In fact, we could follow Pederson and
use networks for this purpose: degrees of polysemy could be
captured by counting "empty" nodes. If, for example, the connection
is broken off at several points in a network, this may be a "weak"
polysemy, i.e., closer to homonymy. This is a tentative suggestion,
but it would be useful in the case at hand: the derivational categories
expressed by -s are closer to each other than to the passive, and this
is shown clearly both in Pederson's, Kemmer's and Enger and
Nesset's networks. Thus, even if we claim polysemy for the verbal
suffix -5, we have to accept the semantic distance being larger
between the passive and the other categories than between for
example the reflexive and the reciprocal.
3. Relevance
S-forms with a passive meaning are inflectional; other s-forms are

derivational (cf. §1). Why are only the passives inflectional? The
answer may be that the passive is less "relevant", in Bybee's (1985)
sense. In simplified terms, the passive makes a smaller semantic
contribution than the other categories. This explanation takes seman-
tics as its starting-point. Question 2 in §1 above can hardly be
answered any other way, and it is not obvious (to me, at least) that
any other current approach to morphology has any answer.
3.1. The concept of relevance
According to Bybee, two parameters decide whether a category is

expressed as inflection or as derivation. One of these parameters is
relevance. A category is relevant for the verb to the extent that its
meaning directly affects the lexical content of the verb stem (Bybee
1985: 15). The larger change of meaning a category induces, the
more relevant it is.8 The other parameter is generality, i.e., semantic
compatibility. A category that can be combined with many verbs is
high in generality.
Whether a category is expressed inflectionally or not is decided
by these two semantic parameters. A category has to be so relevant
that it can be inflectional; a category of very low relevance will not
be grammaticalised. On the other hand, for a category to be
inflectional, it must be relatively general as well, so it cannot be too
relevant, either. Derivational categories may be of higher relevance.
Relevance and generality are not completely independent; high
relevance tends to detract from generality (Bybee 1985: 17). More
relevant suffixes are expected to occur closer to the verb stem than
less relevant ones.
252 Hans-Olav Enger
3.2. What are the meanings of -s ?
We shall see that the passive is less relevant than the other categories
expressed by the suffix -s. However, the immediate problem then is
to decide what other "meanings" are expressed by -s. The literature
gives no unanimous answer; Beito (1986) operates with four diffe-
rent meanings, Faarlund, Lie, and Vannebo (1997) with five. At this
stage, however, we shall simply follow Faarlund, Lie, and Vannebo
(1997): their grammar is reasonably new and authoritative, and has
the advantage of being easily "translatable" into Bybee's terms.
Faarlund, Lie, and Vannebo (1997: 511-512) say that the following
five categories are expressed by -s: a) the reflexive, b) the reciprocal,
c) the inchoative, e.g. dages 'become day', d) others, e) the passive.9
Nothing sensible can be said here about the category "others". What
we need to show, is that the reflexive, the reciprocal and the
inchoative are more relevant than the passive.
3.3. Bybeean relevance meets -s
To begin with, the reflexive and the reciprocal are more relevant
than the passive; they change the action denoted by the verb more
than the passive does. This is to say that the semantic "distance" is
shorter between the active and the passive than between the active
and the reflexive and the reciprocal. The reason is that both in the
active and in the passive, the agent is unaffected by the action. In the
reflexive and the reciprocal, by contrast, the agent is both carrying
out the action and is affected by it (Bybee 1985: 20).
The higher relevance of the reciprocal and the reflexive can be
illustrated by examples. Admittedly, even if a category is more
relevant in one example, it need not follow that it always must be
more relevant, but the examples are probably fairly representative.
We begin with the reciprocal. Consider tales. It has two slightly
different meanings. One example, viz. the reciprocal, was given in
§1: Vi tales i neste uke 'we'll talk next week'. The other meaning is
shown in e.g. Det tales sä mye om morfologi na om dagen lot is
said about morphology these days'. This second meaning is passive.

The reciprocal is derivational here, the passive is inflectional.
Compare paraphrases with bli 'become', another sort of passive in
Norwegian: Del blir talt sä mye om morfologi na om dagen is
acceptable, but *Vi blir talt i neste uke is not.
The semantic distance between reciprocal tales and the active
form tale is larger than the semantic distance between the passive
tales and the active form. The reason is as follows: it is possible to
tale without there being any conversation involved: Hitler talte til
massene 'Hitler spoke to the masses'. Likewise, it is possible to tales
—in the inflectional, passive sense—without there being any con-
versation involved: Det tales sä mye om morfologi, men del er bare
enetale. lot is said about morphology, but it is just a monologue'.
In the case of the derivational tales, by contrast, there must be a
conversation. This indicates that the semantic distance between the
active and the reciprocal is larger than that between the active and
the passive; i.e., that the reciprocal is more relevant.
At this stage, a possible objection should be considered: despite
its intuitive appeal, the idea of 'semantic distance' may be difficult
to measure. Even though the active form tale is similar to the passive
form tales in that none of the forms obligatorily refer to a
conversation, the passive form differs from the active form in that
the agent is not only unexpressed, but also unspecified, and cannot
easily be supplied ("Dei tales sä mye om morfologi av meg lot is
said about morphology by me' is not perfectly grammatical). By
contrast, reciprocal tales and active tale are similar as regards the
specificity of the agents.
This objection is interesting, but it may be discarded if it can be
argued that the active-passive similarity is more important than the
active-reciprocal similarity. Fortunately, such an argument can be
supplied: conversation is somehow a different activity from
monologue. Thus, the reciprocal -s induces a change in meaning that
has to do with the world as we perceive it (whatever that may be!);
in Bybee's (1985: 5) terminology, we may talk of a "semantic
change". By contrast, whether an agent is expressed or specified,
does not really change the activity denoted by tale. Thus, the
254 Hans-Olav Enger
objection can be discarded; the passive -s is less relevant than the

reciprocal.
We now turn to the reflexive; consider sarnies. An example of the
passive sense is papir sarnies inn pä mandager 'paper is collected on
Mondays'. (Compare the near-synonymous papir blir samlet inn pä
mandager.) This is different from the reflexive example given in §1
above, herrene sarnies rundt sigarene. The reflexive simply cannot
be paraphrased by a construction with bli. In the reflexive sarnies,
there is an element of intentionality that is absent in the passive. In
this way, the passive is more similar to the active, because the active
samle is usually combined with an inanimate object. Notice also that
the reflexive can be paraphrased by a different verb, e.g. stimle
sammen. The reflexive is thus more relevant than is the passive.10
The inchoative, finally, is clearly fairly relevant. "To begin to do
something" is different from "to do something". (A methodological
problem here, however, is that those verbs that have -s as an
inchoative suffix, do not usually have counterparts without -s.)
We have now seen that the inchoative, the reflexive and the
reciprocal are more relevant than the passive. In consequence, it
comes as no surprise that these categories are derivational while the
passive is inflectional. However, on the basis of Bybee (1985), we
cannot predict that these categories must be derivational in
Norwegian. Thus, our explanation is not "nomothetic"; rather, it is
an explanation in terms of motivation. The notion of motivation is
explicated at great length by a number of authors (see e.g. Lakoff
1987; Nesset 1998b: 27-34), so it will not be discussed in depth here.
Suffice it to say that what is motivated, is not completely predictable,
but still far from arbitrary.
An explanation in terms of motivation may not fit with wide-
spread positivist ideas of what constitutes a 'scientific explanation',
since motivation does not give us any basis for prediction. Lass
(1980) in particular has emphasised prediction as the ultimate
criterion for scientific explanation. However, strict insistence on
prediction may perhaps, somewhat irreverently, be characterised as a
case of 'physics envy': after all, meteorologists do not make 100%
correct predictions either (at least not in the Scandinavian
countries!); yet it would be absurd to deny meteorology the status of

a science. An explanation in terms of motivation also qualifies as an
explanation. To the best of my knowledge, no other morphological
theory than Bybee's comes anywhere near explaining why only the
passive -s is inflectional in Norwegian.
3.4. Relevance and the order of suffixes
The value of relevance becomes even clearer when we consider the

fact that Norwegian children sometimes rearrange the order of -s and
other suffixes. The standard past tense of synes 'think, feel' is syntes.
The past tense suffix t thus comes "inside" the -s. Since past tense is
usually seen as inflectional in Norwegian, whereas -s here is
derivational, this is anomalous.11 It is therefore no surprise that
Norwegian children sometimes produce synste, thereby setting the
order of the suffixes right, as it were, by getting more relevant
suffixes closer to the root (cf. §3.1). In the terminology of
Haspelmath (1993), this is a case of "externalisation of inflection"
(see Enger 2000).
Similarly, Icelandic children sometimes produce the "incorrect"
forustum instead of "standard" forumst 'we perish', cf. Anderson
(1990: 270, fn. 5). Anderson comments that "[s]uch forms [as
forustum, HOE] apparently represent the misanalysis of -st as a
simple derivational suffix, to which inflectional suffixes are then
added, as opposed to the structure-making constituent argued for in
the subsequent text here." By Anderson's analysis, then, forustum is
just a mistake, and not terribly interesting, either. However, the
parallelism with Norwegian indicates that there is more to this
example. In fact, Birkmann (1997) presents good arguments for
analysing Icelandic -st as a derivational suffix, and Ottosson (1992)
has made a similar case for Old Icelandic -st.
Furthermore, there are parallels to Norwegian synste in the dialect
(or perhaps language) of Dalarna, Sweden, as described by Levander
(1928: 168). Levander notes that in that dialect, -s (the cognate of
Norwegian -s) "is often inserted in front of the suffix proper" (my
256 Hans-Olav Enger
translation), such as in minnsom 'we remember', where the -s is

placed closer to the root than the person/number agreement marker -
om. This development, like the Norwegian one, makes perfect sense
if we take -s to be derivational, and assume Bybee's relevance to be
at work here.
In fact, Kemmer (1993: 189), who does invoke Bybeean
relevance in the Icelandic example, points to a parallel to the
Icelandic development in Surselvan (a Romance language). Since an
analysis by means of relevance make cases like forustum something
more than an isolated error, this analysis seems superior to one
which cannot.12
4. Conclusion
This paper set out to answer two questions (cf. §1: 1) Should the
Norwegian verb suffix -s be analysed as several homonyms or one
polysemous suffix? We have found that the answer should be "one
polysemous suffix". 2) Why is -s inflectional only when it is
passive? The answer is that the passive is of lower relevance, in
Bybeean terms.
These issues should be of interest not only for the study of
Norwegian. More specifically, even though the distinction between
inflection/derivation is accepted, the paper provides arguments
against a "split morphology"; in §2.2, it turned out that the cate-
gorical split between the two kinds of morphology is the reason why
it is so difficult to generalise over the different uses of -s in WP
models.
While the paper is based in part on networks (also known as
semantic maps), it has shown why we should be cautious in using
them as the only evidence for homonymy/polysemy. The paper has
also presented support for Bybee's (1985) relevance concept. These
conclusions have bearing on both morphological theory and
cognitive semantics.
Notes
1. I am indebted to Rolf Theil Endresen, Staffan Hellberg, Ronald Langacker,

Tore Nesset, Tomas Riad, an anonymous referee, and the participants at The
verb in Cognitive Grammar for very helpful comments on different versions of
this paper. Thanks also to Osten Dahl for making me aware of -s in Dalarna,
and to Mark Aronoff for e-mail clarification of the nature of morphomes. The
mistakes that no doubt remain are my own.
2. We concentrate on Norwegian Bokmäl, but Nynorsk, Swedish and Danish
seem similar in most relevant respects.
3. Some linguists take common etymology as a criterion for polysemy (e.g. Beard
1995: 33), but this is highly disputable (cf. Lyons 1977: 550-551, and, from a
different angle, Leiss 1997: 145-147). Anyway, the -s suffixes can all be traced
back to one common etymon sik.
4. This problem is probably not restricted to WP models. For example, Halle and
Mohanan (1985: 63) posit two distinct English suffixes -ing (one adjective
ending, one verb ending). The two suffixes are located at different "strata", and
not explicitly connected. They are thus treated as if they were homonyms.
However, this paper focuses on WP only.
5. The arguments against meta-rules are taken from Nesset (1994: 17-18) and
Enger (1998a: 44-45).
6. Anderson (1990) has argued that Icelandic st-verbs, which are the cognates of
Norwegian s-verbs, create problems for a morpheme-based model of
morphology; problems that do not arise in a WP-type model (such as
Anderson's "a-morphous morphology"). Yet Norwegian -verbs create
problems for a WP-type model, too.
7. An additional objection is that Pederson's argument relies crucially on the
architecture of his network. If the network had been drawn differently, what
now looks like homonymy would instead look like polysemy. It is therefore
important that the network approach in cognitive linguistics is not entirely
unproblematic, cf. Sandra and Rice (1995). Sandra and Rice (1995: 94) even
say that "there seem to be as many network models as there are network
modellers"; Pederson (1991: 90) himself says that his diagram is greatly
simplified, and his diagram differs in some minor details from Kemmer's
(1993). Therefore, it seems premature to base decisions about homonymy
versus polysemy on network evidence only.
8. Admittedly, the concept of semantic change is not quite as well developed in
Bybee's theory as one might have wished. This problem cannot be pursued
here, however, and we shall see that Bybeean relevance is still interesting.
258 Hans-Olav Enger
9. The label "others" may seem strange, but it is not unfair. In the original
wording, Faarlund, Lie, and Vannebo (1997: 512) say: "d) Andre s-verb som
ikke faller inn i betydningsgruppene ovenfor", in my translation: "Other s-
verbs that do not fall into the semantic groups above".
10. An interesting complication here is that the semantic distance between the
active and the passive form of tale may be larger than that between the active
and the passive form of samle. However, this complication does not challenge
the analysis presented here.
11. Faarlund, Lie, and Vannebo (1997: 508) argue that, since -s comes "outside"
the inflectional suffix in such cases, it is problematic to consider -s
derivational. However, there is independent evidence for taking -s in these
cases to be derivational, as has traditionally been done (cf. Enger 2000 for a
more detailed discussion). Moreover, derivational suffixes do sometimes occur
inside inflectional suffixes (see e.g. Wurzel 1984: 41, Dressler 1989, Booij
1993, Rainer 1996). After all, if derivation and inflection were defined in terms
of position only, then it would amount to a non-empirical truism to say that
derivation is inside inflection.
12. Recently, Bybee's concept of relevance has been challenged by Beard (1995:
54-55). He argues that affixes are ordered either according to some universal
principle such as Bybeean relevance or according to language-specific rules (as
assumed by Anderson 1992). Beard concludes that "the actual ordering of
features is, as Anderson maintains, a variable of the parameter set by individual
languages. The lack of consistent relevance of ordering across languages casts
doubts on Bybee's hypothesis". However, it is not obvious why affix ordering
cannot be influenced by both universal and language-specific factors. Stump
(1997: 239, fn. 3) explicitly points out this possibility. Hence, Beard's
argument may be an example of what Langacker (1987a: 28) has dubbed "the
exclusionary fallacy". Much of morphology may in fact be subject to such
partly conflicting influences between universal and language-specific factors;
cf. Wurzel (1984). Without Bybeean relevance, it is also difficult to account for
synste.
Control and transitivity: A study of the
Norwegian verb love 'promise'
Kristian Emil Kristoffersen
The Norwegian verb love 'promise' can appear in several different

complementation patterns, among them the following two: (i) subject
and infinitival complement, (ii) subject, nominal object and
infinitival complement. These two patterns exhibit a phenomenon
known as argument control. In this paper I propose an account of
these two patterns in terms of construction grammar. Furthermore, I
present some data which demonstrate that pattern (i) above is far
more frequent than pattern (ii). Finally, I suggest an explanation of
this difference in terms of type frequency.'
I. Overview
This paper is an investigation of the relationship between transitivity

and argument control in one Norwegian verb; love 'promise'. The
term TRANSITIVITY refers to the ability of a predicator to take one or
more objects, whereas ARGUMENT CONTROL refers to a relation of
coreference between a logical, (obligatorily) unexpressed subject of a
complement clause and another argument in the linguistic context of
this subject. Consider the following Norwegian examples:
(1) a. Politiet päbyr folk a sykle pä fortauet.

police-the orders people to cycle on pavement-the
b. Kari nekta Per a reise til Roma.
Kari denied Per to go to Rome
260 Kristian Emil Kristoffersen
(2) a. Jeg 0nsker ä reise til Tyskland.

I wish to go to Germany
b. Jeg planlegger ä reise til Tyskland.
I plan to go to Germany
In the two sentences in (1) the logical subject of the infinitival

complement is coreferent with the direct object of the matrix verb;
thus, they are instances of object control. In contrast, the sentences in
(2) display what is called subject control, where the coreference
relation holds between the logical subject of the complement verb
and the matrix subject. In both cases, however, control arises from
lexical semantic properties of the matrix predicates in combination
with the presence of an obligatorily subjectless complement.
The examples in (l)-(2) are part of a more general pattern in Nor-
wegian and other languages, showing a strong correlation between on
the one hand monotransitive syntax and subject control, as in (1), and
on the other ditransitive syntax and object control, as in (2). In §2 be-
low I analyse the syntactic and semantic patterns shown in (l)-(2)
within the theoretical framework of CONSTRUCTION GRAMMAR. In
my analysis these two sets of sentences instantiate two distinct con-
structions, which I call the OBJECT CONTROL CONSTRUCTION and
SUBJECT CONTROL CONSTRUCTION respectively. It will be argued
that the different control relations expressed by the two sets of
sentences follow from the semantic characteristics of these construc-
tions.
However, not all Norwegian control verbs appear to fit easily into
either of the two constructions mentioned above. The Norwegian
verb love is a case in point here. Consider the examples in (3):
(3) a. V-INF
De hadde lovet ä komme.
they had promised to come
b. V-NP-1NF
De hadde lovet meg a komme.
they had promised me to come
The Norwegian verb love 'promise' 261
(3a) has monotransitive syntax, and can thus be analysed as an

instance of the subject control construction. (3b), on the other hand,
has the syntax of the object control construction, while at the same
time expressing a subject control relation. In other words, there
seems to be a conflict between syntax and semantics in (3b). In §3 I
suggest a solution to this apparent conflict in terms of the notion of
MULTIPLE INHERITANCE.
Concerning the complementation of love, (3a and b) are not the
only possible syntactic patterns associated with this verb. Compare
these sentences with the pair in (4):
(4) a. V-S
De hadde lovet at de skulle komme.
they had promised that they would come
b. V-NP-S
De hadde lovet meg at de skulle komme.
they had promised me that they would come
If we consider the data in (3) and (4) against the correlation noted
above between monotransitive syntax and subject control, and
ditransitive syntax and object control, we might expect the pattern
illustrated in (3 a) (V-INF) to be more frequent in actual usage than
the pattern illustrated by (3b) (V-NP-INF). Furthermore, we expect
that the pattern in (4b) (V-NP-S) is more frequent than the pattern in
(3b) (V-NP-INF). In §4 I present some data showing that these
expectations are to a large extent borne out. Finally, in §5 I suggest
an explanation of the difference in frequency between the V-INF and
the V-NP-INF patterns in terms of Bybee's (1985, 1995a) correlation
between high type frequency and productivity.
In sum,I have the following four aims with this paper:
1. To propose an account of argument control in terms of two

distinct control constructions (§2),
2. to propose an analysis of the two different syntactic patterns the
verb love can appear in as a "control verb", i.e., V-INF and V-
NP-INF (§3),
3. to present a set of data that demonstrate that the V-INF pattern

of love is more frequent than the V-NP-INF pattern of the
same verb (§4), and
4. to explain this difference in frequency (§5).
2. A construction grammar analysis of argument control
In this paragraph I first briefly present relevant aspects of the theore-

tical framework employed in this paper, construction grammar
(§2.1).2 Next, I propose an analysis of argument control within this
framework (§2.2).
2.1. Construction grammar
Within construction grammar syntactic constructions are regarded as

basic linguistic units. Consider the following four sentences:
(5) a. Mary gave Jill some flowers.

b. Peter sent Mary the book.
c. Jill told the children a story.
d. John promised Peter a book.
These four sentences are taken to be instances of an abstract con-

struction, which Goldberg (1995) calls THE ENGLISH DITRANSITIVE
CONSTRUCTION. This construction is related to, but different from,
the individual verbs that appear in it. It carries a particular meaning,
which is expressed by a particular form. Within this framework there
is no substantial difference between constructions and words, as both
types of unit can be characterised as "form-meaning-function com-
plexes [differing] only with regard to internal complexity" (Michaelis
andLambrecht 1996: 216).
The English ditranstive construction, which is instantiated by the
four sentences in (5), is an example of a LINKING CONSTRUCTION. A
basic property of linking constructions is that they express systematic
correspondences between the thematic arguments conventionally

associated with these constructions, and the syntactic elements that
express these roles. For example, on Goldberg's analysis of the
English ditransitive construction, the construction is argued to be a
mapping between the three arguments agent, recipient and patient,
and the syntactic functions subject, object and secondary object.
Labels like agent, recipient and patient refer to what is called
ARGUMENT ROLES. Argument roles belong to constructions. How-
ever, also the verbs which enter into linking constructions are associ-
ated with roles; these latter roles are referred to as PARTICIPANT
ROLES. For example, Goldberg (1995: 50-51) calls the three
participant roles of the English ditransitive verb hand, which appears
in sentence (6), 'hander' (Mary), 'handee' (Peter), and 'handed'
(book), while the corresponding argument roles are agent, recipient
and patient.
(6) Mary handed Peter a book.
When a verb occurs in a construction, its participant roles are said to

be fused with the corresponding argument roles. Thus, in sentence (6)
the participant roles 'hander', 'handee' and 'handed' of hand are
fused with the argument roles agent, recipient and patient of the
ditranstive construction. A general constraint on fusion is that for a
participant role and an argument role to be fused, they must be
semantically compatible, i.e., the participant role must be an instance
of the argument role (The semantic coherence principle, Goldberg
1995: 50). In other words, the participant role 'hander' can be fused
with the argument role agent since the former can be construed as an
instance of the more general agent argument role, i.e., a 'hander' is a
kind of agent. Likewise, a 'handee' is a kind of recipient, and
'handed' is a kind of patient.
Another aspect of syntactic constructions is that they are said to
LICENSE the elements that instantiate them. The ditransitive construc-
tion, for example, licences three syntactic arguments, subject, object
and secondary object.
As noted above constructions are pairings of form and meaning.

Concerning the meaning of constructions, it has been argued that
many constructions have polysemous senses. A case in point is the
English ditransitive construction which, according to Goldberg
(1995: 38), has (7a) as its central sense, and (7b-c) as two of its
related senses. Verbs motivating the various senses are given in (8).
(7) a. Agent successfully causes recipient to receive patient.

b. Agent intends to cause recipient to receive patient.
c. Agent acts to cause recipient to receive patient at some
future point in time.
(8) a. Peter gave Mary a cake,

b. Peter baked Mary a cake.
c. Peter left Mary a letter.
The fact that constructions are associated with several distinct, but
systematically related senses is referred to as CONSTRUCTIONAL
POLYSEMY. This is connected to the claim made within construction
grammar that there is no substantial difference between words and
constructions, cf. the following statement by Goldberg (1995: 32):
"[SJince constructions are treated as the same basic data type as
morphemes, that they should have polysemous senses is expected."
2.2. A constructional account of argument control
As noted in the introductory paragraph the term ARGUMENT CON-

TROL designates a relation of coreference between the unexpressed
(logical) subject of an infinitival complement and a syntactic
argument (subject, object or oblique) in the linguistic context of this
complement. In what follows I propose an account of control in
which the notion of linking construction plays a central role.
Compare first the sentences in (9) with those in (10):
(9) a. Karl nekta Per ä reise til Roma. (= (lb))

b. Jeg planlegger ä reise til Tyskland. (= (2b))
(10) a. Kari nekta Per mat.

Kari denied Per food
b. Jeg planlegger en reise.
I plan a journey
An important difference between the two pairs of sentences is that

the pair in (9) contains controlled complements, whereas the pair in
(10) does not. This contrast can be described from two different
perspectives. On the one hand one could see the two instances of
nekte 'deny' as two verbs with distinct but related senses. One sense,
which involves control, is present in (9a); the other, which does not
involve control, is found in (lOa). Thus, control is seen as part of the
semantics of the verb. However, this description fails to account for
the fact that control in (9a) appears to arise from the combination of
the argument structure of nekte (three-place) and a particular
syntactic form of the third argument (infinitival complement).
An alternative analysis which presents a solution to this problem
would be to analyse the difference between (9a) and (lOa) as involv-
ing one and the same verb which participates in two distinct syntactic
contructions. This is the basis for the analysis proposed below
(§2.2.1-2.2.2), and argued more extensively for in Kristoffersen
(1996), where object and subject control are analysed in terms of two
types of constructions, the object control construction and the subject
control construction.
2.2.1. The object control construction
Object control refers to a coreference relation between a grammatical

object and the logical subject of a complement infinitival which is a
coargument of this object. Consider again the two examples of object

control in (1) together with a third instance, all shown in (11):
(11) a. Politiet päbyr folk ä sykle pä fortauet. (= (la))

b. Kan nekta Per ä reise til Roma. (= (lb))
c. Politiet tillot folk ä sykle pä fortauet.
police-the permitted people to cycle on pavement-the
Most Norwegian sentences involving object control refer to events in

which one participant tries to get another participant to perform some
subevent (expressed by the infinitival complement), as in (lla), or to
prevent another participant from performing some subevent, as in
(lib), or to enable another participant to perform some subevent, as
in (lie). In accord with this I suggest to analyse the meaning of the
object control construction as metaphorical extensions of Talmy's
(1988) concept of a force dynamic opposition between an antagonist
and an agonist. Together these extensions constitute the polysemous
meaning of the object control construction, described in (12).
(12) THE SENSES OF THE OBJECT CONTROL CONSTRUCTION

• Antagonist acts to get agonist to perform the subevent.
• Antagonist acts to prevent agonist from performing subevent.
• Antagonist acts to enable agonist to perform subevent.
The object control construction is a linking construction specifying

three semantic arguments, antagonist, expressed by the subject,
agonist, expressed by the NP object, and a propositional argument
expressed by the infinitival complement. Accordingly, this con-
struction licences three arguments. That the object argument of this
construction refers to the controller follows from the relationship
between antagonist on the one hand and agonist and subevent on the
other.
2.2.2. The subject control construction
Subject control refers to a coreference relation beteween a

grammatical subject and the logical subject of a complement
infinitival which is a coargument of this subject. Two examples
which involve subject control are given in (13):
(13) a. Je g 0nsker ä reise til Tyskland.

I wish to go to Germany
In Kristoffersen (1996) I analyse verbs like 0nske in terms of a

semantic element 'desire' which is attributed to the subject referent,
who has a desire to realise the subevent. A verb like planlegge, on
the other hand, can be taken to express a semantic element 'inten-
tion', again attributed to the subject referent, who intends to realise
the subevent.
Several linguists have pointed to the fact that 'desire' and 'inten-
tion' are semantically related, often with further reference to expres-
sions of a grammatical meaning future (Bybee 1985: 166; Bybee,
Perkins, and Pagliuca 1994: 256; Heine, Claudi, and Hunnemeier
1991: 173-175; Wierzbicka 1988: 32). Accordingly, these two
meaning elements can be analysed as constituting the polysemous
meaning of the subject control construction. I follow Bybee, Perkins,
and Pagliuca (1994: 256) in their view that the two elements are
related by metonymy.
2.2.3. Summary
The subject control construction is like the object control construc-

tion in being a linking construction, in this case licencing two argu-
ments, one (experiencer with the 'desire' verbs and committor3 with
the "intention" verbs) expressed by the subject and one by the infini-
tival complement. The properties of the two constructions are sum-

marised in (14) and (15).
(14) THE OBJECT CONTROL CONSTRUCTION

• is a polysemous linking construction based on metaphori-
cal extensions of a force dynamic opposition; and
• licences three arguments, antagonist, agonist, and soa-arg,4
and
• has the direct object as controller.
(15) THE (MONOTRANSITIVE) SUBJECT CONTROL CONSTRUCTION

• is a polysemous linking construction comprising two
meanings, "desire" and "intention", which are metonymi-
cally related; and
• licences two arguments, experiencer/committor and soa-
arg, and
• has the subject as the controller.
The reason for using the term monotransitive subject control con-
struction in (15) will become apparent in the next paragraph, where a
distinction is made between monotransitive and ditransitive subject
control constructions.
3. A constructional analysis of love
3.L The exceptional status of love as a control verb
One Norwegian verb which does not fit easily into the description of
argument control given in the preceding paragraph is love 'promise'.
Compare sentences (3b) and (la) above, repeated here as (16a and b):
(16) a. V-NP-INF
De hadde lovet meg ä komme.
b. V-NP-INF
Politiet pabyr folk ä sykle pa fortauet.
The contrast between (16a and b) poses the following problem:

These two sentences appear to have the same syntax. However, they
differ with respect to the controller. Paby Order' has object control,
while love has subject control. Thus, if we consider the contrast
between (16a and b) in isolation, love is clearly exceptional, at least
from a purely syntactic point of view (cf. Bresnan 1982: 355; Larson
1991). The view that this verb is exceptional gains strength by the
fact that whereas päby is one of several ditransitive verbs in
Norwegian which involve object control, love is an isolated instance
of a ditransitive verb with subject control.5
Within the constructional approach to control presented in the
previous paragraph the difference regarding controller choice
between these two sentences cannot be straightforwardly accounted
for, since they have the same syntactic form. If a particular syntactic
construction is taken to express a particular meaning, and if
controller choice is taken to follow from semantic properties of this
construction, we should expect (16a) to have object control, and not
subject control. However, this sentence has subject control, a fact
which from a constructional point of view raises the following
problem: According to the description in (15) the subject control
construction licences two arguments, viz. the subject and the
infinitival complement, whereas love in (16a) has three arguments.
Thus, we may ask how the recipient object is licenced in (16a).
3.2. Multiple inheritance
To answer that question, consider first (3a), repeated here as (17):
(17) V-INF
This sentence is an instance of the subject control construction,

expressing among other things an intention on the part of the subject
referent to perform the action referred to by the complement, cf. (15).
The two arguments are licenced by the construction.
Next, consider (16a). As mentioned, the problem in this case is
that the subject control construction cannot license the recipient
object. However, this argument is clearly licenced when love appears
in another construction, the DITRANSITIVE CONSTRUCTION, cf. (18):
(18) Fetter lovet meg en bok.

Peter promised me a book
Thus, the use of love observed in (16a)can be related to two different

syntactic constructions. Stated in terms of argument licencing, we
could say that two of the arguments of love in (16a) are licenced by
one construction, the subject control construction, whereas the third
is licenced by another construction, the ditransitive construction. This
view of love as used in the pattern in (16a) (V-NP-INF) can be hand-
led in terms of the notion of MULTIPLE INHERITANCE, which is charac-
terised as follows by Goldberg (1995: 97):6 "Multiple inheritance
allows us to capture the fact that instances of some construction types
seem to resist being uniquely characterized in a natural way [...]."
A relation of inheritance holds between two constructions A and B
if the two have formal and semantic properties in common. The
relationship between A and B is asymmetric, in the sense that one
inherits from the other. Multiple inheritance, on the other hand,
occurs when one construction inherits from more than one other
construction, i.e., if it has formal and semantic properties in common

with more than one other construction.
Argument licensing in examples like (16a) can be accounted for as
a case of multiple inheritance. I call the construction instantiated by
this sentence the DiTRANSiTiVE SUBJECT CONTROL CONSTRUCTION,
in contrast with the monotransitive subject control construction
described in (15) above. This construction stands in a relation of
multiple inheritance to two other constructions, which jointly licence
the three syntactic arguments: the monotransitive subject control con-
struction (15) licenses the subject argument and the prepositional
argument, and the ditransitive construction licenses the recipient
object. Furthermore, the subject control construction contributes with
the subject control interpretation. This relationship of multiple inheri-
tance is outlined in Figure 1; with argument licensing indicated by
arrows.
To recapitulate, (16a), in which love is used as a control verb in a
ditransitive pattern can be taken to instantiate two distinct syntactic
constructions; on the one hand the subject control construction (see
(15)) and on the other the ditransitive construction. The three argu-
ments of love in (16a) are licenced by both of these constructions.
MONOTRANSITIVE SUBJECT DITRANSITIVE

CONTROL CONSTRUCTION CONSTRUCTION
NP V VP NP V NP NP
NP V NP VP
DITRANSITIVE SUBJECT
CONTROL CONSTRUCTION
Figure 1. Argument licencing in the V-NP-INF pattern of love

4. Love: patterns of use in written and spoken Norwegian
In §3 two complementation patterns of love were discussed, V-INF

and V-NP-INF. In (4) two patterns were exemplified, with the
infinitival complements (cf. (16)) replaced by finite complement
sentences. Examples of all patterns are repeated in (19):
(19) a. V-INF
De hadde lovet a komme.
b. V-NP-INF
c. V-S
De hadde lovet at de skulle komme.
they had promised that they would come
d. V-NP-S
De hadde lovet meg at de skulle komme.
they had promised me that they would come
In §1 I formulated the following two assumptions concerning the

actual usage of the different patterns in (19) on the basis of the
correlation between transitivity and control:
• The V-INF pattern in (19a) is more frequent than the V-NP-

INF pattern in (19b).
• The V-NP-S pattern in (19d) is more frequent than the V-NP-
INF pattern in (19b).
The first assumption accords with the fact that in the V-INF pattern
the correlation between monotransitive syntax and subject control
holds, whereas this is not the case with the V-NP-INF pattern;
because of this one would think that speakers and writers would
prefer the former pattern. Similarly, the second assumption addresses
the fact that of the two patterns V-NP-S and V-NP-INF only the
latter displays the mentioned "conflict" between syntax and

semantics.
To see if the two assumptions are supported by actual usage, I
have (a) examined a corpus of Norwegian texts, and (b) given two
elicitation tasks to speakers of Norwegian concerning the difference
between the patterns in (19). In what follows I present the results.
4.1. Love in a corpus of Norwegian texts7
All sentences in the corpus containing love with either finite or

infinitival complement have been classified according to two
variables, first, whether the verb takes a finite or infinitival
complement, and, second, whether or not the sentence also contains a
nominal object expressing the recipient. Sentences with finite
complements have been included only if the complement subject is
coreferent with the subject of love. Tables 1 and 2 present the number
of occurrences for each of the four patterns illustrated in (1) and (2).
Table 1. Choice of complement when nominal object is absent (corpus)

Finite sentence 5 12.8% (19c)
Infinitival complement 34 87.2% (19a)
Total 39 100.0 %
Table 2. Choice of complement when nominal object is present (corpus)

Finite sentence 5 45.5% (19d)
Infinitival complement 6 54.5% (19b)
Total Π 100.0 %
These tables show that in the majority of the examples without a

nominal object (Table 1: 87.2 %) love takes an infinitival
complement, not a finite sentence. However, when a nominal object
is present, the corpus does not point in favour of any of the two
alternative complement types, cf. Table 2. Note, however, that the
total number of occurrences here is low, only 11 in all.
4.2. Elicitation tasks
The aim of the elicitation tasks was to find out what complement
type subjects choose when asked to complete a sentence beginning
with Jeg lover deg ... Ί promise you ....' or Jeg lover ... Ί promise
...'. The subjects were given the Norwegian equivalents of one of the
following two sets of instructions:8
• Imagine the following situation: You want to make a promise

to someone about something. What you promise has to be an
action performed by yourself. You start by saying: "I promise
you..." How would you finish this sentence?
• Imagine the following situation: You want to make a promise
about something. What you promise has to be an action
performed by yourself. You start by saying: "I promise ..."
How would you finish this sentence?
Tables 3 and 4 summarise the results of these tests:
Table 3. Choice of complement when nominal object is absent (elicitation)

Finite sentence 10 22.2 % (19c)
Infinitival complement 35 77.8% (19a)
Total 45 100%
Table 4. Choice of complement when nominal object is present (elicitation)

Finite sentence 89 62.7 % (19d)
Infinitival complement 53 37.3 % (19b)
Total 142 100%
These results give a somewhat clearer picture of the variation

between finite and infinitival complementation with love than the
results from the corpus study. Table 3, presenting the results for
sentences without nominal object, supports the results from the
corpus study (Table 1), whereas Table 4 shows that a majority of my
subjects prefers finite complement in the presence of a nominal
object (Table 2). Table 5 summarises the results from the corpus
study and the elicitation tasks:
Table 5. Choice of complements with love (corpus and elicitation)

Pattern Corpus Elicitation
V-INF (19a) 34 87.2 % 35 77.8 %
V-S (19c) 5 12.8 % 10 22.2 %
V-NP-INF (19b) 6 54.5 % 53 37.3 %
V-NP-S (19d) 5 45.5 % 89 62.7 %
To sum up: I pointed to two possible differences in frequency

concerning the four complementation patterns illustrated in (19); (i)
that the V-INF pattern is more frequent than the V-NP-INF pattern,
and (ii) that the V-NP-S pattern is more frequent than the V-NP-
INF pattern. As Table 5 shows, (i) is supported by the results both
from the corpus examination and the elicitation tasks. The second
difference, however, is supported only by the results from the
elicitation tasks. I have no answer to the question why this should be
so. It may possibly be related to the fact that the diverging results
correlate with the difference between written and spoken language.
However, the number of instances of sentences with a nominal object
present in the written corpus is too low to draw strong conclusions.
5. Why is V-INF more frequent than V-NP-INF?
If we consider the difference between the V-INF pattern on the one

hand and the V-NP-INF on the other, another question emerges: Why
do so many speakers and writers of Norwegian prefer the former
pattern to the latter? It seems to me that a plausible answer to this
question can be given in terms of Bybee's (1985, 1995a) use of type
frequency in explaining productivity. On the basis of morphological
data from several languages Bybee argues that grammatical affixes
which in a language L have high type frequency are more productive
than affixes in L having low type frequency.
What, then, would the type frequency of a construction be? In
construction grammar grammatical constructions are considered the
syntactic counterparts of grammatical affixes. Goldberg (1995: 22-
23) says the following about the correspondence between these two
units: "On the present account, the closed class grammatical
morpheme is analogous to the English skeletal construction; the verb
stem plays the role of the main verb. The semantic integration of verb
and morpheme is analogous to the integration of construction and
verb in English."
From this point of view, the type frequency of a given
construction in a language L is the same as the number of verbs
appearing in that construction in L (cf. Goldberg 1995: 134).
Let us now turn to the construction types which have been
discussed in this paper. Consider first the monotransitive subject
control construction (MSC), illustrated by the sentences in (2) and
described in (15). (20) lists 16 other Norwegian verbs participating in
this construction:
(20) avslä 'refuse'; elske 'love'; foreslä 'propose'; foretrekke

'prefer';/0rs0A:e 'try'; hate 'hate'; kreve 'demand'; like 'like';
nekte 'refuse'; planlegge 'plan';pr0ve 'try'; sverge 'sverge';
tenke 'intend'; trenge 'need'; velge 'choose'; 0nske 'want'
Against the 16 verbs in (20), love is one of three Norwegian verbs

which can appear with ditransitive syntax and subject control. The
two others are tilby Offer' and garantere 'guarantee', which under
special circumstances can appear with subject control (Engh [1982];
Runde (1997), cf. note 5 above). This means that the construction
instantiated by for example the sentences in (2) has much higher type
frequency than the construction instantiated by for example (19a)
(the ditransitive subject control construction, DSC), cf. Table 6 on
the opposite page, based on the number of verbs in (20), and the three
verbs love, tilby and garantere).
In view of this difference in type frequency it is expected that love
as a control verb primarily appears without recipient object. If, on the
other hand, one wishes to express the nominal object as well, one can
instead invoke the pattern instantiated by (19d), which in fact is what
the majority of my subjects do (cf. Table 5), although this result is
not corroborated by the results from the corpus examination (cf.

Table 2).
Table 6. Type frequency of the two subject control constructions

Construction type MSC DSC
Number of verbs 16 3
Type frequency 84.2 % 15.8 %
6. Conclusion
In this paper I have analysed the Norwegian verb love 'promise'

when used as a control verb, as in the sentences in (21):
(21) a. V-INF
b. V-NP-INF
Both sentences express a relation of subject control, i.e. a coreference

relation between the subject of the matrix verb (V) and the logical
subject of the infinitival complement (INF). Concerning the proper-
ties of this verb I formulated four aims for the paper in § 1:
1. To propose an account of argument control in terms of two

distinct control constructions (§2),
2. to propose an analysis of the two different syntactic patterns the
verb love can appear in as a "control verb", i.e. V-INF and V-
NP-INF (§3),
3. to present a set of data that demonstrate that the V-INF pattern
of love is more frequent than the V-NP-INF pattern of the
same verb (§4), and
4. to explain this difference in frequency (§5).
In §2 two different patterns of argument control were discussed;

subject control (22a), and object control (22b):
(22) a. Kari nekta Per ä reise til Roma.

In the proposed analysis of argument control (22a) was analysed as

an instance of a linking construction, called the object control con-
struction, with three argument roles, antagonist, agonist and soa-arg.
(22b) was analysed as an instance of another linking construction, the
subject control construction expressing one of two metonymically
related argument roles, committor and experiences
The analysis treats argument control not as a property of indi-
vidual verbs but as a property of the two constructions. As far as I
can see this is a natural way of accounting for the fact that many
"control verbs" can be used in syntactic contexts where control is not
involved, compare (22) with (23) and (24):
(23) Per nekta Kari mat.

Per denied Kari food
(24) a. Jeg planlegger en reise.

I plan a journey
b. Jeg planlegger at jeg skal reise.
I plan that I shall go
Within lexical approaches to argument control like, i.a., Bresnan

(1982), Larson (1991), and Pollard and Sag (1994), the variation
between (17) on the one hand and (23) - (19) on the other needs to be
treated in terms of variable verb meanings. On the analysis proposed
here a verb like planlegge is the same both in (22b) and (24);
semantic differences between the use of this verb in the three
sentences are to be attributed to the constructions it appears in.
§3 was devoted to the second aim, in particular the question how

the pattern instantiated by (21b) should be analysed with respect to
the control constructions described in §2. The proposed analysis here
builds on the notion of multiple inheritance: Sentence (21b) is taken
to instantiate two different linking constructions, the subject control
construction and the ditransitive construction. The three syntactic
arguments of this sentence was taken to be licenced by both con-
structions simultaneously, the subject and the infinitival complement
by the subject construction and the recipient object by the ditransitive
construction.
My analysis of (2la and b) implies that the V-INF pattern (2la) is
the basic one with respect to control, whereas the V-NP-INF pattern
(21b) is exceptional in the sense that it inherits one more construction
than the basic pattern does. This is the opposite perspective of that
taken by most lexical theories of the control properties of love, where
V-NP-INF is seen as the basic pattern, and the V-INF as a variant
pattern in which the recipient object is omitted.9
In §4, devoted to the third aim, I presented some data which
showed that among speakers and writers of Norwegian the V-INF
pattern is more frequent than the V-NP-INF pattern. This is to be
expected if the former is the basic control pattern associated with
love. Finally, in §5 I looked at differences between the two patterns
in terms of type frequency, and concluded that it is expected that love
as a control verb primarily appears without recipient object, since the
type frequency of the monotransitive control construction is higher
than that of the ditransitive subject control construction.
Notes
1. Thanks to Jan Engh, Anders N0klestad, Kjell Johan Saeb0 and two anonymous
reviewers for useful comments. Errors are mine.
2. Here I build primarily on Goldberg (1995); other recent studies employing this
framework are Zwicky (1994), Michaelis and Lambrecht (1996) and Kay and
Fillmore(1999).
3. The term committor (Pollard and Sag 1994: 287) refers to a participant who
commits to perform the action denoted by the infinitival complement.
4. The term soa-arg (Pollard and Sag 1994: 287) refers to an argument expressing
a state of affairs.
5. Two other ditransitive verbs in Norwegian, tilby Offer' and garantere 'guaran-
tee' can under special circumstances appear with subject control, see Engh
[1982] and Runde (1997), However, these two verbs differ from love in that the
latter has subject control as the unmarked reading.
6. See also Michaelis and Lambrecht (1996: 237) and Kay and Fillmore (1999)
for discussion of this concept.
7. The corpus contains approximately 1.5 million word forms, and consists of (A)
three issues of the Norwegian newspaper Aftenposten: 940116; 940115;
930322, and (B) the following Norwegian (bokmal variant) novels: A. Briseid:
Kjedereaksjoner, 1977; A. Eriksen: Promenade, 1937; A. Jensen: Ikaros, 1957;
A. K. Elstad: Magret, 1977; B. Vestre: Sparet αν en sti, 1957; Ε. Griffiths:
Lenket til kjcerligheten, 1957; G. Allbjart: Flukten til livet, 1937; G.
Brantenberg: Egalias d0tre, 1977; G. Johannessen: H0st i mars, 1977; H.
Geelmuyden: Periferi og sentrum, 1937; H. S. Dehlin: Men skyggen f0lger
efter, 1937; I. Heiberg: Hjeml0s, 1957; J. Bj0rkelund: Ragnarokk! 1977; J.
Brinchmann: Mannen som kom tilbake, 1937; J. Bj0rneboe: Under en h rdere
himmel, 1957; J. H. Jensen: D0den er stamgjest, 1977; K. Alnaes: Gaia, 1977;
K. Askildsen: Davids bror, 1977; K. Bj0rnstad: Vinterbyen, 1977; K. Fasting:
Det gode kj0bmannskap, 1937; L. S. Christensen: Amat0ren, 1977; N. Grieg:
Spansk sommer, 1937; R. Arntzen: N r alt er nytt, 1937; R. Mager0y: Gunhild,
1957; S. Christophersen: Demningen, 1957; T. Stigen: Frode Budbcereren,
1957.
8. The two sets of instructions were given to two different groups of subjects;
thus, no informant has answered both sets.
9. Pollard and Sag (1994) represent a notable exception to this trend, when they
state that some subject control verbs allow a third argument in addition to the
subject argument and the infinitival complement.
References
Aasen, Ivor
1864 Norsk Grammatik [Norwegian Grammar]. [Reprinted 1965, by
Universitetsforlaget, Oslo.]
Abelin, Asa
1999 Studies in sound symbolism. Gothenburg: Department of
Linguistics, Göteborg University. [Gothenburg Monographs in
Linguistics 17]
Anderson, Stephen R.
1990 The grammar of Icelandic verbs in -st. In Joan Maling and Annie
Zaenen (eds.), Syntax and Semantics 24, modern Icelandic syntax,
235-287. San Diego: Academic Press.
1992 A-morphous morphology. Cambridge: Cambridge University Press.
Arnott, D. W.
1970 The nominal and verbal systems ofFula. Oxford: Clarendon Press.
Aronoff, Mark
1976 Word formation in generative grammar. Cambridge, MA: MIT
Press.
1994 Morphology by itself. Cambridge, MA: MIT Press.
Baayen, R. H.
1991 Quantitative aspects of morphological productivity. In Geert Booij
and Jaap van Marie (eds.), Yearbook of morphology 1991, 109-149.
Dordrecht: Kluwer.
1992 On frequency, transparency and productivity. In Geert Booij and
Jaap van Marie (eds.), Yearbook of morphology 7992, 181-208.
Dordrecht: Kluwer.
Bauer, Laurie
1983 English word-formation. Cambridge: Cambridge University Press.
1988 Introducing linguistic morphology. Edinburgh: Edinburgh
University Press.
282 References
Beard, Robert
1995 Lexeme-morpheme-base morphology. Albany, NY: SUNY Press.
Beito, Olav T.
1986 Nynorsk grammatikk. Lyd- og ordlcere [Nynorsk Grammar.
Phonology and Morphology]. Oslo: DetNorske Samlaget.
Bennett, David
1992 Stratificational grammar. In W. Bright (ed.), International encyclo-
pedia of linguistics, vol 4, 77-79. Oxford: Oxford University Press.
Berko, Jean
1958 The child's learning of English morphology. Word 14: 150-177.
Berkov, Valerij Pavlovic
1989 Die Reflexiva in den skandinavischen Sprachen. In Ernst Walter
and Hartmut Mittelstadt (eds.), Altnordistik: Vielfalt und Einheit.
Erinnerungsband für Walter Baetke, 144-160. Weimar: Hermann
Böhlaus Nachfolger.
Birkmann, Thomas
1997 Das neuisländische Mediopassiv: Flexion oder Wortbildung? In T.
Birkmann et. al. (eds.), Vergleichende germanische Philologie und
Skandinavistik. Festschrift für Otmar Werner, 81-89. Tübingen:
Max Niemeyer.
Bjerkan, Kirsten Meyer
2000 Verbal morphology in specifically language impaired children:
Evidence from Norwegian. (Acta Humaniora 73.) Oslo: Unipub.
Bjerkan, Kirsten Meyer and Hanne Gram Simonsen
1995 Past tense processing in Norwegian: Experimental evidence from
children and adults. Paper presented at NELAS 4 and Workshop on
Dyslexia and Early Language Development, Jyväskylä, Finland, 7-
9 September 1995.
1996 Prosessering av preteritumsformer i norsk: Eksperimentell evidens
fra barn og voksne [Processing of past tense in Norwegian:
Experimental evidence from children and adults]. Norsk Lingvistisk
Tidsskrift 14: 189-207.
Bloomfield, Leonard
1935 Language. London: George Allen & Unwin.
References 283
Bokmälsordboka [The Bokmäl Dictionary]

1986 Mark Ingebj0rg Landr0 and Boye Wangensteen (eds.). Oslo:
Universitetsforlaget.
Booij, Geert
1993 Against split morphology. In Geert Booij and Jaap van Marie
(eds.), Yearbook of morphology 1993, 27-49. Dordrecht: Kluwer.
Boretzky, Norbert
1983 Kreolsprachen, Substrate und Sprachwandel. Wiesbaden: Otto
Harrassowitz.
Breedveld J. O.
1996 Form and meaning in Fulfulde: A morphophonological study of
Maasinakore. Leiden: CNWS.
Bresnan, Joan
1982 Control and complementation. In Joan Bresnan (ed.), The mental
representation of grammatical relations, 282-390. Cambridge, MA:
MIT Press.
Brown, Roger
1973 A first language: The early stages. Cambridge, MA: Harvard
University Press.
Brown, Roger and Ursula Bellugi
1964 Three processes in the child's acquisition of syntax. In E.
Lenneberg (ed.), New directions in the study of language.
Cambridge MA: MIT Press.
Burani, Cristina and Alfonso Caramazza
1987 Representation and processing of derived words. Language and
Cognitive Processes 2: 217-227'.
Burani, Cristina and Alessandro Laudanna
1992 Units of representation for derived words in the lexicon. In R. Frost
and L. Kats (eds.), Orthography, phonology, morphology, and
meaning, 361-376. Amsterdam: Elsevier.
Bybee, Joan L.
1985 Morphology: A study of the relation between meaning and form.
Amsterdam: John Benjamins.
1988 Morphology as lexical organization. In Michael Hammond and
Michael Noonan (eds.), Theoretical morphology. Approaches in
modern linguistics, 119-141. New York: Academic Press.
284 References
1995a Regular morphology and the lexicon. Language and Cognitive

Processes 10: 425-455.
1995b Diachronie and typological properties of morphology and their
implications for representation. In Laurie Beth Feldman (ed.),
Morphological aspects of language processing, 225-246. Hillsdale,
NJ: Lawrence Erlbaum.
Bybee, Joan, Revere Perkins and William Pagliuca
1994 The evolution of grammar. Tense, aspect, and modality in the
languages of the world. Chicago: University of Chicago Press.
Bybee, Joan L. and Dan I. Slobin
1982 Rules and Schemas in the development and use of English past
tense. Language 58: 265-289.
Caramazza, Alfonso, Alessandro Laudanna and Cristina Romani
1992 Lexical access and inflectional morphology. Cognition 28: 297-
332.
Carstairs-McCarthy, Andrew
1992 Current morphology. London: Routledge.
Chafe, Wallace
1994 Discourse, consciousness, and time: The flow and displacement of
conscious experience in speaking and writing. Chicago: University
of Chicago Press.
Cho, Euiyon
1991 Notes on some tests for subjecthood in Korean. Studies in the
Linguistic Sciences 21.2: 19-29.
Chomsky, Noam
1986 Knowledge of language. Its nature, origin, and use. New York:
Praeger.
Collins, A. and M. Quillian
1969 Retrieval time from semantic memory. Journal of Verbal Learning
and Verbal Behavior 8: 240-247.
Cook, Kenneth W.
1988 A cognitive analysis of grammatical relations, case, and transitivity
in Samoan. San Diego: University of California doctoral
dissertation.
Creider, Chet and Richard Hudson
1999 Inflectional morphology in word grammar. Lingua 107: 163-187.
References 285
Croft, William
1990 Typology and universals. Cambridge: Cambridge University Press.
Cutler, Anne
1980 Productivity in word formation. Papers from the Regional Meeting
of the Chicago Linguistic Society 16: 45-51.
Cutrer, Michelle
1994 Time and tense in narrative and in everyday language. San Diego:
University of California doctoral dissertation.
Dahl, Osten and Maria Koptjevskaja-Tamm
1992 Language typology around the Baltic Sea: A problem inventory.
Stockholm: Stockholm University.
Dalrymple, Mary
1993 The syntax of anaphoric binding. Stanford, CA: CSLI.
Davis, Garry W.
1992 What's in a word? On integrating recent approaches to secondary
associations, submorphemic units and morphological segmentation.
Word 43: 197-216.
De Mauro, Tullio, Federico Mancini, Massimo Vedovelli and Miriam Voghera
1993 Lessico difrequenza dell 'italiano parlato. Milano: Etas.
Derwing, Bruce L. and WJ. Baker
1986 Assessing morphological development. In Paul Fletcher and
Michael Garman (eds.), Language acquisition. Studies in first
language development. [2nd ed.] Cambridge: Cambridge
University Press.
Derwing, Bruce L. and Royal Skousen
1989 Morphology in the mental lexicon: A new look at analogy. Year-
book of morphology 2: 55-71.
1994 Productivity and the English past tense. Testing Skousen's analogy
model. In D. Lima, R.L. Corrigan and G.K. Iverson (eds.), The
reality of linguistic rules, 193-218. Amsterdam: John Benjamins.
Dimitrova-Vulchanova, Mila
1996 Verb semantics, diathesis and aspect. Trondheim: Norwegian
University of Science and Technology doctoral dissertation.
286 References
Dressler, Wolfgang Ullrich

1989 Prototypical differences between inflection and derivation. Zeit-
schrift für Phonetik, Sprachwissenschaft und Kommunikations-
forschung 42: 3-10.
Dyvik, Helge
1980 Har gammelnorsk passiv? [Does Old-Norwegian have passive?] In
Even Hovdhaugen (ed.), The Nordic languages and modern
linguistics 4, 81-107. Oslo: Universitetsforlaget.
Elman, Jeffrey L.
1990 Finding structure in time. Cognitive Science 14: 179-211.
Emmory, Karen and Victoria Fromkin
1988 The mental lexicon. In F. Newmeyer (ed.), Linguistics. The
Cambridge survey, vol 3, 124-149. Cambridge: Cambridge
University Press.
Endresen, Rolf Theil
1982 Morfologi [Morphology]. In Harald Bjorvand, Even Hovdhaugen
and Hanne Gram Simonsen (eds.), Sprakvitenskap. En elementcer
innf0ring, 87-126. [Linguistics. An elementary introduction]. [2nd
ed.] Oslo: Universitetsforlaget.
1991 Fonetikk og fonologi. Ei elementar innf0ring [Phonetics and
phonology. An elementary introduction]. Oslo: Universitets-
forlaget.
1992 Norske verb og kognitiv morfologi [Norwegian verbs and cognitive
morphology]. Oslo: Unpublished ms., Department of Linguistics,
University of Oslo.
1994 The etymological relationship between the benefactive and
causative suffixes in Fula. Trondheim: Trondheim Working Papers
in Lingustics 22: 89-100.
Endresen, Rolf Theil, Hanne Gram Simonsen and Andreas Sveen
1996 Innf0ring i lingvistikk [Introduction to linguistics]. Oslo: Uni-
versitetsforlaget.
Enger, Hans-Olav
1998a The classification of strong verbs in Norwegian with special
reference to the Oslo dialect. (Acta Humaniora 26). Oslo:
Universitetsforlaget.
References 287
1998b Forholdet mellom presens -form og Infinitiv s-form [The

relationship between present -form and infinitival s-form]. Maal
og Minne 2 (1998): 193-206.
2000 Verbendelsen -s i norsk bokmäl: B0ying eller avledning? [The
verbal ending -s in Bokmäl: Inflectional or derivational?] Norsk
Lingvistisk Tidsskrift 18: 9-36.
Enger, Hans-Olav and Tore Nesset
1999 The value of cognitive grammar in typological studies: The case of
Norwegian and Russian passive, middle and reflexive. Nordic
Journal of Linguistics 22: 27-60.
Engh, Jan
[1982] 1996 On control. Norskrift 89: 19-48.
Faarlund, Jan Terje, Svein Lie and Kjell Ivar Vannebo
1997 Norsk referansegrammatikk [Norwegian Reference Grammar].
Oslo: Universitetsforlaget.
Fagerli, Ole Torfinn
1994 Verbal derivations in Fulfulde. Trondheim: University of
Trondheim cand.philol thesis. [Trondheim Working Papers in
Lingustics 21].
1995 Benefactive constructions: A typological study. Trondheim Work-
ing Papers in Linguistics 25: 100-118.
Fauconnier, Gilles and Eve Sweetser (eds.)
1996 Spaces, worlds, and grammar. Chicago: University of Chicago
Press.
Fodor, Jerry
1983 The modularity of mind. An essay on faculty psychology.
Cambridge, MA: MIT Press.
Forrest, Linda B.
1996 Discourse goals and attentional processes in sentence production:
The dynamic construal of events. In Adele E. Goldberg (ed.),
Conceptual structure, discourse and language, 149-161. Stanford:
CSLI.
288 References
Frauenfelder, Uli H. and Robert Schreuder

1992 Constraining psycholinguistic models of morphological processing
and representation: The role of productivity. In Geert Booij and
Jaap van Marie (eds.), Yearbook of morphology 1991, 165-183.
Dordrecht: Kluwer.
Fromkin, Victoria
1988 Grammatical aspects of speech errors. In F. Newmeyer (ed.),
Linguistics. The Cambridge survey, vol 2, 117-138. Cambridge:
Cambridge University Press.
Fromkin, Victoria and Robert Rodman
1988 An introduction to language. [4th ed.] New York: Holt, Rinehart
and Winston.
Geniusiene, Emma
1987 The typology of reflexives. Berlin: Mouton de Gruyter.
Givon, Talmy
1984 Syntax: A functional-typological introduction, vol. 1. Amsterdam:
John Benjamins.
Goldberg, Adele E.
1995 Constructions. A construction grammar approach to argument
structure. Chicago: University of Chicago Press.
Goldberg, Adele E. (ed.)
1996 Conceptual structure, discourse and language. Stanford: CSLI.
Gopnik, Myrna
1994 Impairment of tense in a familial language disorder. Journal of
Neurolinguistics 8: 109-133.
Greenberg, Joseph H.
1963 Some universale of grammar with particular reference to the order
of meaningful elements. In Joseph H. Greenberg (ed.), Universals
of language, 73-113. Cambridge, MA: MIT Press.
Gundersen, Helge
1995 Linjedansere og pantomine pa sirkhus. Folkeetymologi som morfo-
logisk omtolking [Line dancers and pantomine at the circhouse.
Folk etymology as morphological reanalysis]. Oslo: Novus.
References 289
1997 Objektivisme i lingvistikken [Objectivism in linguistics]. In Anita

Leirfall and Thor Sandmel (eds.), Festskrift til Johan Arnt Myrstad
i anledning 50-arsdagen [Festschrift for J. A. M. on the occasion
of his 50th birthday], 147-181. Trondheim: Tapir.
Hagen, J. E.
1994 Svake verb pä -Vde og -Vie i bokmäl og deres paradigmetilh0righet
[Weak verbs ending in -Vde and -Vte in Bokmäl and their
paradigm membership]. In J. Myking, H. Sand0y and I. Utne (eds.),
Helsing til Lars Vassenden pä 70-arsdagen [Greeting for L.V. on
his 70th birthday], 57-67. Bergen: Nordisk institutt.
Haiman, John
1983 Iconic and economic motivation. Language 59: 781-819.
1985 Natural syntax. Cambridge: Cambridge University Press.
Halle, Morris
1973 Prolegomena to a theory of word formation. Linguistic Inquiry 4: 3-
16.
Halle, Morris and K.P. Mohanan
1985 Segmental phonology of Modern English. Linguistic Inquiry 16:
57-116.
Halliday, Michael
1985 An introduction to functional grammar. London: Arnold.
Hare, M., J.L. Elman and K.G. Daugherty
1995 Default generalisation in connectionist networks. Language and
Cognitive Processes 10: 601-630.
Haspelmath, Martin
1993 The diachronic externalization of inflection. Linguistics 31: 279-
309.
Heine, Bernd, Ulrike Claudi and Frederike Hünnemeier
1991 Grammaticalization. A conceptual framework. Chicago: University
of Chicago Press.
Hellan, Lars
1988 Anaphora in Norwegian and the theory of grammar. Dordrecht:
Foris.
Hestvik, Arild
1992 LF movement of pronouns and antisubject orientation. Linguistic
Inquiry 23: 557-594.
290 References
Hockett, Charles F.
1954 Two models of grammatical description. Word 10: 210-234.
1958 A course in modern linguistics. New York: Macmillan.
van Hoek, Karen
1995 Conceptual reference points: A cognitive grammar account of
pronominal anaphora constraints. Language 71: 310-340.
1996 A cognitive grammar account of bound anaphora. In Eugene H.
Casad (ed.), Cognitive linguistics in the Redwoods: The expansion
of a new paradigm in linguistics, 753-791. Berlin: Mouton de
Gruyter.
1997a Anaphora and conceptual structure. Chicago: University of
Chicago Press.
1997b Backwards anaphora as a constructional category. Functions of
Language 4: 47-82.
to appear Quantifier scope: A cognitive grammar perspective.
Holm, John A.
1988 Pidgins and Creoles, vol 1. Theory and Structure. Cambridge, UK:
Hudson, Richard
1971 English complex sentences. An introduction to systemic grammar.
Amsterdam: North Holland.
1976 Arguments for a non-transformational grammar. Chicago:
University of Chicago Press.
1984 Word grammar. Oxford: Blackwell.
1990 English word grammar. Oxford: Blackwell.
1995 Word meaning. London: Routledge.
1996 Sociolinguistics. [2nd ed.] Cambridge: Cambridge University Press.
1998 English grammar. London: Routledge.
2000 */ amn 't. Language: 297-323.
Isacenko, Aleksandr V.
1982 Die russische Sprache der Gegenwart. Formenlehre. Munich: Max
Hueber Verlag.
Jackendoff, Ray
1975 Morphological and semantic regularities in the lexicon. Language
51: 639-671.
References 291
1997 The architecture of the language faculty. Cambridge, MA: MIT

Press.
Jaeger, Jeri T., A.H. Lockwood, D.L. Kemmerer, R.D. Van Valin, B.W. Murphy
and H.G. Khalak
1996 A positron emission tomographic study of regular and irregular
verb morphology in English. Language 72: 451-497.
Janda, Laura A.
1993 Cognitive linguistics as a continuation of the Jakobsonian tradition.
The semantics of Russian and Czech reflexives. In Robert A.
Maguire and Alan Timberlake (eds.), American contributions to the
eleventh international congress of slavists, 310-319. Columbus,
Ohio: Slavica.
1997 GIVE, HAVE, and TAKE in Slavic. In John Newman (ed.), The
linguistics of giving, 249-265. Amsterdam: John Benjamins.
Jensen, John T.
1990 Morphology: Word structure in generative grammar. Amsterdam:
John Benjamins.
Johannessen, Ole-J0rgen
1993 Ordlaging. In Jarle R0nhovd (ed.), Norsk morfologi, 125-153. Oslo:
Ad No tarn Gyldendal.
Kay, Paul and Charles J. Fillmore
1999 Grammatical constructions and linguistic generalizations: The
What'sXdoing Y? construction. Language 75: 1-33.
Keenan, Edward L. and Bernard Comrie
1977 Noun phrase accessibility and universal grammar. Linguistic
Inquiry 8: 63-99.
Kemmer, Suzanne
1993 The middle voice. Amsterdam: John Benjamins.
Kirchner, D. M. and R. L. Klatzky
1985 Verbal rehearsal and memory in language-disordered children.
Journal of Speech and Hearing Research 28: 556-565.
Kristoffersen, Kristian Emil
1996 Infinitival phrases in Old Norse: Aspects of their syntax and
semantics. Oslo: University of Oslo dr.art thesis.
292 References
Kumashiro, Toshiyuki
in prep. The conceptual basis of grammar: A cognitive approach to
Japanese clause structure. San Diego: University of California
doctoral dissertation.
Kuroda, S.-Y.
1972 The categorical and the thetic judgment. Foundations of Language
9: 153-185.
Kutsch Lojenga, Constance
1997 Verbal derivation in Ngiti and Lendu. Handout for paper presented
at the 2nd World Congress of African Linguistics, Leipzig.
Labov, William
1972a Language in the inner city. Philadelphia: University of
Pennsylvania Press.
1972b Sociolinguistic patterns. Philadelphia: University of Pennsylvania
Press.
1994 Principles of linguistic change, vol 1. Internal factors. Oxford:
Blackwell.
Lakoff, George
1987 Women, fire and dangerous things. What categories reveal about
the mind. Chicago: University of Chicago Press.
Lakoff, George and Mark Johnson
1980 Metaphors we live by. Chicago: University of Chicago Press.
Lamb, Sidney
1966 Outline of stratificational grammar. Washington, DC: Georgetown
University Press.
Lambrecht, Knud
1987 Sentence focus, information structure, and the thetic-categorical
distinction. Proceedings of the Annual Meeting of the Berkeley
Linguistics Society 13, 366-382.
1994 Information structure and sentence form: Topic, focus, and the
mental representations of discourse referents. Cambridge:
1996 On the formal and functional relationship between topics and
vocatives. Evidence from French. In Adele E. Goldberg (ed.),
Conceptual structure, discourse and language, 267-288. Stanford:
CSLI.
References 293
Langacker, Ronald W.
1986 Abstract motion. Proceedings of the annual meeting of the Berkeley
Linguistics Society 12: 455-471.
1987a Foundations of cognitive grammar, vol. 1. Theoretical
prerequisites. Stanford, CA: Stanford University Press.
1987b Nouns and verbs. Language 63: 53-94.
1990 Concept, image and symbol. The cognitive basis of grammar.
Berlin: Mouton de Gruyter.
1991 Foundations of cognitive grammar, vol. 2. Descriptive application.
Stanford, CA: Stanford University Press.
1993 Reference-point constructions. Cognitive Linguistics 4: 1-38.
1994 Cognitive Grammar. In R. Asher (ed.), Encyclopedia of language
and linguistics, 590-593. Oxford: Pergamon.
1995a Conceptual grouping and constituency in cognitive grammar. In
Linguistic Society of Korea (ed.), Linguistics in the morning calm
3, 149-172. Seoul: Hanshin.
1995b Possession and possessive constructions. In John R. Taylor and
Robert E. MacLaury (eds.), Language and the cognitive construal
of the world, 51-79. Berlin: Mouton de Gruyter.
1997a Constituency, dependency, and conceptual grouping. Cognitive
Linguistics 8: 1-32.
1997b A dynamic account of grammatical function. In Joan L. Bybee,
John Haiman and Sandra A. Thompson (eds.), Essays on language
function and language type dedicated to T. Givon, 249-273.
1999a Double-subject constructions. In Sung-Yun Bak (ed.), Linguistics
in the morning calm 4, 83-104. Seoul: Hanshin.
1999b Virtual reality. Studies in the Linguistic Sciences 29.2: 77-103
Larson, Richard K.
1991 Promise and the theory of control. Linguistic Inquiry 22: 103-139.
Lass, Roger
1980 On explaining language change. Cambridge: Cambridge University
Press.
Leira, Vigleik
1992 Ordlaging og ordelement i norsk [Morphology and morphemes in
Norwegian]. Oslo: Det Norske Samlaget.
294 References
Leiss, Elisabeth
1997 Synkretismen und Natürlichkeit. Folia Linguistica 31: 133-161.
van der Lely, Heather K.J. and Michael T. Ullman
in press Past tense morphology in specifically language impaired and nor-
mally developing children. Language and Cognitive Processes.
Leonard, Laurence B.
1998 Children with specific language impairment. Cambridge, MA: MIT
Press
Levander, Lars
1928 Dalmalet. Beskrivning och historia II [The Dal dialect. Description
and history]. Uppsala: Appelbergs.
Li, Charles N. and Sandra A. Thompson
1976 Subject and topic: A new typology of language. In Charles N. Li
(ed.), Subject and topic, 457-489. New York: Academic Press.
Lieber, Rochelle
1989 On percolation. Yearbook of morphology 2: 95-138.
Lobben, Marit
1998 Force as a metaphor in grammar, with special emphasis on the
Hausa verbal marking of causative and benefactive. Paper
presented at The Verb in Cognitive Grammar, Gran, Norway.
Lord, Carol
1993 Historical change in serial verb constructions. Amsterdam: John
Benjamins.
Liibben, August
1993 Mittelniederdeutsches Handwörterbuch. Facsimile. Darmstadt:
Wissenschaftliche Buchgesellschaft.
Luger, George and William Stubblefield
1993 Artificial intelligence. Structures and strategies for complex
problem solving. [2nd ed.] New York: Benjamin/Cummings.
Lyons, John
1977 Semantics, vol. 2. Cambridge: Cambridge University Press.
MacWhinney, Brian
1977 Starting points. Language 53: 152-168.
MacWhinney, B. and J. Leinbach
1991 Implementations are not conceptualizations: Revising the verb
learning model. Cognition 40: 121-157.
References 295
Maldonado, Ricardo
1988 Energetic reflexives in Spanish. Proceedings of the annual meeting
of the Berkeley Linguistics Society 14: 153-165.
1993 Dynamic construals in Spanish. Studi haliani di Linguistica
Teorica e Applicata 22: 531-566.
Manney, Linda
1990 Mental experience verbs in Modern Greek: A cognitive explanation
for active versus middle voice. Proceedings of the annual meeting
of the Berkeley Linguistics Society 16: 229-240.
1992 States, results, and the active/passive/middle continuum in Modern
Greek. Journal of Modern Greek Studies 10: 205-233.
1995 Pragmatic motivation for inflectional middle voice in Modern
Greek. Functions of Language 2: 159-188.
1998 The reflexive archetype and its various realizations in Modern
Greek. Studies in Language 22: 1-48.
Marantz, Alec
1992 Review of Jensen 1990. Language 68: 413-416.
Marchman, Virginia A.
1997 Children's productivity in the English past tense: The role of
frequency, phonology and neighborhood structure. Cognitive
Science 21: 283-304
Marcus, Gary F., Ursula Brinkmann, Harald Clahsen, Richard Wiese and Steven
Pinker
1995 German inflection: The exception that proves the rule. Cognitive
Psychology 29: 189-256.
Marcus, Gary F., Steven Pinker, Michael Ullman, Michelle Hollander, T. John
Rosen and Fei Xu
1992 Overregularisation in language acquisition. Monographs of the
society for research in child development 57.4, serial No. 228.
van Marie, Jaap
1996 The unity of morphology. In Geert Booij and Jaap van Marie (eds.),
Yearbook of morphology 1995, 67-82. Dordrecht: Kluwer.
Marslen-Wilson, William
1980 Speech understanding as a psychological process. In J.C. Simon
(ed.), Spoken language generation and understanding. Boston:
Reidel.
296 References
Marslen-Wilson, William and Lorraine Tyler

1980 The temporal structure of spoken language understanding.
Cognition 18: 1-71.
1981 Central processes in speech understanding. Philosophical
Transactions of the Royal Society of London B295: 317-332.
Martin, James
1992 Systemic grammar. In W. Bright (ed.), International encyclopedia
of linguistics, vol 4, 120-122. Oxford: Oxford University Press.
Matcovich, Paola F.
1998 The acquisition and processing of past participle morphology in
Italian. An experimental study. Oslo: University of Oslo cand.
philol thesis.
Matsumoto, Yo
1996a How abstract is subjective motion? A comparison of coverage path
expressions and access path expressions. In Adele E. Goldberg
(ed.), Conceptual structure, discourse and language, 359-373.
Stanford: CSLI.
1996b Subjective motion and English and Japanese verbs. Cognitive
Linguistics!: 183-226.
1996c Subjective-change expressions in Japanese and their cognitive and
linguistic bases. In Gilles Fauconnier and Eve Sweetser (eds.),
Spaces, worlds, and grammarl24-l56. Chicago: University of
Chicago Press.
Matthews, P. H.
1972 Inflectional morphology. Cambridge: Cambridge University Press.
1991 Morphology. Cambridge: Cambridge University Press.
1993 Grammatical theory in the United States from Bloomfield to
Chomsky. Cambridge: Cambridge University Press.
McCarthy, John J.
1982 Prosodic templates, morphemic templates, and morphemic tiers. In
Harry van der Hülst and Norval Smith (eds.), The structure of
phonological representations, part 1,191-223. Dordrecht: Foris.
Michaelis, Laura A. and Knud Lambrecht
1996 Toward a construction-based theory of language function: The case
of nominal extraposition. Language 72: 215-247.
References 297
Naes, Oddvar
1972 Norsk grammatikk. Elementare strukturer og syntaks [Norwegian
Grammar. Basic structures and syntax]. Oslo: Fabritius.
Nesset, Tore
1994 Russian stress: Stress as an inflectional formative in Russian noun
paradigms and Bybee 's cognitive morphology. Oslo: Novus.
1998a Affiks eller klitikon? Om elementet -sja i moderne standard russisk
[Affix or clitic? On the element -sja i modern Standard Russian].
Norsk Lingvistisk Tidsskrift 16: 185-206.
1998b Russian conjugation revisited: A cognitive approach to aspects of
Russian verb inflection. Oslo: Novus.
Newman, John (ed.)
1996 Give: A cognitive linguistic study. Berlin: Mouton de Gruyter.
1997 The linguistics of giving. Amsterdam: John Benjamins.
Nichols, Johanna, Gilbert Rappaport and Alan Timberlake
1980 Subject, topic, and control in Russian. Proceedings of the Annual
Meeting of the Berkeley Linguistics Society 6: 372-386.
Nida, Eugene
1949 Morphology: The descriptive analysis of words. [2nd ed.] Ann
Arbor: University of Michigan Press.
N0klestad, Anders
1996 A connectionist modelling approach to past tense acquisition in
Norwegian. Oslo: University of Oslo cand.philol thesis.
Noreen, A.
1923 Altnordische Grammatik. I. 5., unveränderte Auflage. Tübingen:
Max Niemeyer. [Reprinted 1970.]
Noye, Dominique
1989 Dictionnaire Foulfoulde-Fraâis. Dialecte peul du Diamare,
Nord-Cameroun. Paris: Librairie Orientaliste Paul Geuthner.
Nuyts, Jan and Eric Pederson (eds.)
1997 Language and conceptualization. Cambridge: Cambridge
University Press.
van Oosten, Jeanne
1977 Subjects and agenthood in English. Papers from the Regional
Meeting of the Chicago Linguistic Society 13: 459-471.
298 References
1986 The nature of subjects, topics and agents: A cognitive explanation.

Bloomington: Indiana University Linguistics Club.
Orsolini, Margherita and William D. Marslen-Wilson
1997 Universals in morphological representation: Evidence from Italian.
Language and Cognitive Processes 72: 1-47.
The Oslo corpus of tagged Norwegian texts
http://www.tekstlab.uio.no/norsk/bokmaal/english.html
Osam, E.K.
1994 From serial verbs to prepositions and the road between. Sprach-
typologie und Universalienforschung 47: 16-36.
Osthoff, Hermann
1878 Formassociation bei Zahlwörtern. In Hermann Osthoff and Karl
Brugmann (eds.), Morphologische Untersuchungen auf dem
Gebiete der indogermanischen Sprachen, 92-132. Leipzig: Hirzel.
Ottosson, Kjartan G.
1992 The Icelandic middle voice: The morphological and phonological
development. Lund: Department of Scandinavian Languages.
Pederson, Eric William
1991 Subtle semantics: Universals in the polysemy of reflexive and
causative constructions. Berkeley: University of California Ph.D.
dissertation.
Perlmutter, David M. (ed.)
1983 Studies in relational grammar, vol. 1. Chicago: University of
Chicago Press.
PICA
1962 Proceedings of the Institute on Childhood Aphasia. San Francisco:
National Society for Crippled Children and Adults.
Pike, Kenneth L.
1970 Tagmemic and matrix grammar applied to selected African
languages. Summer Institute of Linguistics: Publications in
Linguistics and Related Fields, no. 23.
Pinker, Steven
1994 The language instinct. Harmondsworth, Middlesex: Penguin.
1997 How the mind works. Harmondsworth, Middlesex: Penguin.
1999 Words and rules. London: Weidenfeld and Nicolson.
References 299
Pinker, Steven and Alan Prince

1988 On language and connectionism: Analysis of a parallel distributed
processing model of language acquisition. Cognition 28: 73-193.
Plank, Frans
1981 Morphologische (Ir-)Regularitäten. Aspekte der Wortstruktur-
theorie. Tübingen: Narr.
Plunkett, Kim and Virginia A. Marchman
1991 U-shaped learning and frequency effects in a multi-layered per-
ceptron: Implications for child language acquisition. Cognition 38:
43-102.
1993 From rote learning to system building: Acquiring verb morphology
in children and connectionist nets. Cognition 48: 21-69.
1996 Learning from a connectionist model of the English past tense.
Cognition 61: 299-308.
Pollard, Carl and Ivan A. Sag
1994 Head-driven phrase structure grammar. Chicago: University of
Chicago Press.
Postal, Paul M.
1969 Anaphoric islands. Papers from the Regional Meeting of the
Chicago Linguistic Society 5: 205-239.
Prasada, Sandeep and Steven Pinker
1993 Generalisation of regular and irregular morphological patterns.
Language and Cognitive Processes 8: 1-56.
Prideaux, Gary D.
1984 Psycholinguistics: The experimental study of language. London:
Croom Helm.
Ragnarsdottir, Hrafnhildur, Hanne Gram Simonsen and Kim Plunkett
1997 The acquisition of past tense inflection in Icelandic and Norwegian
children. In Eve Clark (ed.), Proceedings of the 28"' Annual Child
Language Research Forum, 259-270. Stanford: CSLI.
1999 The acquisition of past tense morphology in Icelandic and
Norwegian children: An experimental study. Journal of Child
Language 26: 577-618.
Rainer, Franz
1996 Inflection inside derivation. In Geert Booij and Jaap van Marie
(eds.), Yearbook of morphology 1995, 83-91. Dordrecht: Kluwer.
300 References
Reisberg, Daniel
1997 Cognition. Exploring the science of the mind. New York: Norton.
Rhodes, Richard
1994 Aural images. In Leanne Hinton, Johanna Nichols and John J.
Ohala (eds.), Sound symbolism, 276-292. Cambridge: Cambridge
University Press.
Rhodes, Richard A. and John M. Lawler
1981 Athematic metaphors. Papers from the Regional Meeting of the
Chicago Linguistic Society 17: 318-342.
Rumelhart, D.E. and J.L. McClelland
1986 On learning the past tense of English verbs. In J. L. McClelland,
D. E. Rumelhart and the PDF Research Group (eds.), Parallel
distributed processing: Explorations in the micro structure of
cognition, vol. 2. Psychological and biological models. Cambridge,
MA: MIT Press.
Runde, Älov
1997 En semantisk-pragmatisk beskrivelse av noen norske kontrollverb
[A semantic-pragmatic description of some Norwegian control
verbs]. Oslo: University of Oslo cand.philol thesis.
Sandra, Dominiek and Sally Rice
1995 Network analyses of prepositional meaning: Mirroring whose mind
- the linguist's or the language user's? Cognitive Linguistics 6: 89-
130.
Saether0, Eli
1997 'Minimal' verbs in Akan. A study of the contribution of the verb
and construction in meaning. Trondheim: University of Trondheim
cand.philol thesis.
Say, Tessa
1997 Regular and irregular inflection in the mental lexicon: Evidence
from Italian. Essex Graduate Students Papers in Linguistics 2: 55-
70.
Schoorlemmer, Maaike
1997 Russian sja and the affix-clitic distinction. In Martina Lindseth and
Steven Franks (eds.), Annual workshop on formal approaches to
Slavic linguistics. The Indiana meeting, 7996, 253-274. Ann Arbor:
Michigan Slavic Publications.
References 301
Shibatani, Masayoshi
1985 Passives and related constructions: A prototype analysis. Language
61: 821-848.
Simonsen, Hanne Gram
1993 Norwegian Bokmäl: Strong and weak verbs. Oslo: Unpublished
ms., Department of Linguistics, University Oslo.
Simonsen, Hanne Gram and Kirsten Meyer Bjerkan
1998 Testing past tense inflection in Norwegian: A diagnostic tool for
identifying SLI children? International Journal of Applied
Skousen, Royal
1989 Analogical modeling of language. Dordrecht: Kluwer.
1992 Analogy and structure. Dordrecht: Kluwer.
Söderbergh, Ragnhild
1967 Svensk ordbildning [Swedish morphology.]. Stockholm: Svenska
B okför laget/Nor stedts.
Steele, Susan M.
1977 On being possessed. Proceedings of the Annual Meeting of the
Berkeley Linguistics Society 3: 114-131.
van Stek, Rob
1983 Enkelte aspekter ved suffiks-avledning i norsk [Some aspects of
suffixal derivation in Norwegian]. Studier i skandinavistik 1: 45-75.
Stump, Gregory T.
1997 Template morphology and inflectional morphology. In Geert Booij
and Jaap van Marie (eds.), Yearbook of morphology 1996, 217-241.
Dordrecht: Kluwer.
Sweetser, Eve
1997 Role and individual interpretations of change predicates. In Jan
Nuyts and Eric Pederson (eds.), Language and conceptualization,
116-136. Cambridge: Cambridge University Press.
Sweetser, Eve and Gilles Fauconnier
1996 Cognitive links and domains: Basic aspects of mental space theory.
In Gilles Fauconnier and Eve Sweetser (eds.), Spaces, worlds, and
grammar, 1-28. Chicago: University of Chicago Press.
302 References
Swinney, D.
1979 Lexical access during sentence comprehension: (Re-)consideration
of context effects. Journal of Verbal Learning and Verbal Behavior
5: 219-227.
Sylla, Yero
1979 Grammatical relations and Fula syntax. Ph.D. dissertation,
University Microfilms International, London: Ann Arbor.
Talmy, Leonard
1988 Force dynamics in language and cognition. Cognitive Science 12:
49-100.
1996 Fictive motion in language and "ception". In Paul Bloom, et al.
(eds.), Language and space, 211-276. Cambridge, MA: MIT Press.
Taylor, F. W.
1932 A Fulani-English dictionary. Oxford: Clarendon Press.
Taylor, John R.
1989 Linguistic categorisation: An essay in cognitive linguistics. Oxford:
Oxford University Press.
1996 Possessives in English: An exploration in cognitive grammar.
Oxford: Oxford University Press.
Tesniere, Lucien
1959/66 Elements de syntaxe structurale. Paris: Klincksieck.
Thorell, Olof
1981 Svensk ordbildningslära [A course in Swedish morphology.].
Stockholm: Esselte.
Tlearn neural network simulator
http://crl.ucsd.edu/innate/tlearn.html
Tomblin, J. Bruce, Nancy L. Records, Paula Buckwalter, Xuyang Zhang, Elaine
Smith and Marlea O'Brien
1997 Prevalence of specific language impairment in kindergarten
children. Journal of Speech, Language, and Hearing Research 40:
1245-1260.
Tomlin, Russell S.
1995 Focal attention, voice, and word order. In Pamela Downing and
Michael Noonan (eds.), Word order in discourse, 517-554.
References 303
1997 Mapping conceptual representations into linguistic representations:

The role of attention in grammar. In Jan Nuyts and Eric Pederson
(eds.), Language and conceptualization, 162-189. Cambridge:
Usakov, Dmitrij Nikolaevic (ed.)
1935 Tolkovyj slovar' russkogo jazyka [Monolingual Dictionary of
Russian]. Moscow: Sovetskaja enciklopedija.
Vangsnes, 0ystein Alexander
1999 Substantivfraser: Om kapittel 3 i Norsk referansegrammatikk
[Noun phrases: On chapter 3 of The Norwegian reference
grammar]. Nordica Bergensia 21: 34-50.
Verfaillie, Karl and Anja Daems
1996 The priority of the agent in visual event perception: On the cogni-
tive basis of grammatical agent-patient asymmetries. Cognitive
Linguistics 7: 131 -147.
Western, August
1921 Norsk Riksmals-Grammatikk [Norwegian Riksmäl Grammar].
Oslo: Aschehoug.
Wierzbicka, Anna
1988 The semantics of grammar. Amsterdam: John Benjamins.
Williams, Adger
1993 The argument structure of / -predicates. Journal of Slavic
Williams, Edwin
1981 On the notions 'lexically related' and 'head of a word'. Linguistic
Inquiry 12: 245-274.
Winter, Werner
1969 Analogischer Sprachwandel und semantische Struktur. Folia
Linguistica 3: 29-45.
Wurzel, Wolfgang Ullrich
1984 Flexionsmorphologie und Natürlichkeit. Berlin: Akademie-Verlag.
Yelland, G. W.
1994 Word recognition and lexical access. In R. Asher (ed.), Encyclo-
pedia of language and linguistics, 5009-5014. Oxford: Pergamon.
304 References
Zwicky, Arnold
1994 Dealing out meaning: Fundamentals of syntactic constructions.
Proceedings from the annual meeting of the Berkeley Linguistics
Society 20:611-625.
Index
Aasen, L, 86 Bodomo, A.B. ,211,214

Abelin, A., 110 Bokm l, 74,76,78, 83-85, 89,
absolutive, 231 90-91,93-94,132,168
access, 50 Boretzky, N., 214
activity level, 171 Breedveld, J. O., 206
adversative, 203 Brown, R., 130
affectedness, 214 building-block metaphor, 96
affectedness marker, 204 building-block model, 95
agent, 234 Burani, C, 199
Akan, 211, 213,215,220-221 Bybee, J. L., 99-122, 129-131, 134,
allative, 206,209 155,159, 166, 171, 184,189,
analogical reinterpretation, 101 192
Andenes.E. S.,211 Baayen, R. H., 168, 170
applicative, 204,210
Arnott, D. W., 206
Aronoff, M., 102 consonant class, 90
assemblies of symbolic structures, Caramazza, A., 199
18 Carstairs-McCarthy, A., 113
Atlantic, 204 categorical judgement, 44
Augmented Addressed Morphology causation / purpose, 217
Model, 199 C-class, 86
Chafe, W., 41
Chamorro, 212
Baker, W. J., 130 clause-internal discourse topic, 45
Bantu, 204,210 clipping, 105
Bariba, 215 clipping compounds, 105
base, 19 cognitive grammar, 153
Bauer, L., 102, 105, 106 cognitive morphology, 99
Beito, Ο. Τ., 86 collective, 228, 232
Bellugi.U., 130 completedness, 217
benefactive, 203, 210, 213 composite symbolic structure, 22
benefactor, 206 compounding, 105
beneficiary, 206, 227 Comrie, B., 49
Berko, J., 130, 134 concatenator, 118
Bjerkan, K. M., 92, 115, 130, conceptual reification, 38
140-141, 167-168, 170,177-179, conceptualist semantics, 18
181,183 conceptualizer, 19
blends, 105 conceptually dependent, 41
Bloomfield, L., 97, 102, 109, 110, conjugation classes, 190
112 Conjugation I, 190, 193, 195
306 Index
Conjugation II, 190,193,196 energy flow, 205, 209,215

Conjugation III, 190 energy sink, 205, 209
connectionist models, 166,170, 189 energy source, 205,215
connectionist network simulation, Enger, H.-O., 83, 90, 107
146 English, 26, 33,43-46,105,109,
consonantal conjugation, 86 111-116,126,130-134,150,165
constituency, 23 entrenchment, 149, 162
construal, 18 Essegbey, J., 213
constructional schema, 23 Ewe, 211,213
control phenomena, 13 experiencer, 218
Cook,K.W.,51 extension of sense, 204
correspondence, 22 extension, 232
cranberry morphs, 101
Creole, 204
Cutler, A., 106 Fagerli, O. T., 206
Cutrer, M., 52 Fauconnier, G., 52
Czech, 218 fictive motion, 50
finding things in space, 40
focal prominence, 21, 31
Daems, A., 34 folk etymology, 101
Dagaare, 211,214 force-dynamics, 205, 209
Dalrymple, M., 223 Forrest, L. B., 33
Danish, 115 frame, 25
dative, case 218 Frauenfelder, U. H., 199
dative event, 206 French, 13,14,116
De Mauro, T., 194 frequency, 135
Derwing, B. L., 130, 131, 199 Fromkin, V., 97
destinative, 210 Fula, 210, 213,219,221
dominion, 24 Faarlund, J. T., 86,94, 109,116
double-subject constructions, 15
dual mechanism account, 130
dual mechanism model, 147, 189, generative analysis, 104
199 generative grammar, 106
Dutch, 115 generative linguistics, 96
dynamic, 24 GeniuSiene, E., 226
German, 111,113,115-117,218
Germanic, 133
elaboration, 18 giver, 204
Elman, J. L., 52 Givon, T., 12
emergence / manifestation, 217 global productivity, 168
enablement, 217 goal, 205
Endresen, R. T., 73, 86, 109, 115, Gopnik, M., 162
127,132-136, 155, 161,173, grammatical class, 21
221 grammatical construction, 22
Index 307
grammaticalisation, 81,212 Japanese, 17

Greenberg, J. H., 34 Jensen, J. T., 97, 99
Gundersen, H., 91, 95,101, 117 Johannessen, O.J., 116
Gur languages, 204 Johnson, M, 95, 96
Hagen, J. E., 92, 170 Karaboro, 212

Halle, M., 97 Keenan, E. L., 49
Harris, Z. S., 108 Kemmer, S., 223, 225
Hausa, 221 Korean, 15, 17
head, 23 Kumashiro, T., 51
heavy markers 224 Kwa languages, 204
Hellan, L., 223
Hestvik, A., 223
Hockett, C. F., 102, 107 Lakoff, G., 95, 96
van Hoek, K., 26, 35, 37, 52 Lambrecht, K., 14, 43,44,50
Hokkien Chinese, 212 landmark, 21,30, 124
Holm, J. A., 211 Langacker, R. W., 96, 100-111,
123-124, 153, 160, 166, 171,
189-190,204,224
Icelandic, 149 larger weak (WL) class), 86, 90, 136,
iconic, 45 147, 154, 160
identification, 40 Laudanna, Α., 199
Indonesian, 14 Lawler.J. M., 109, 110
infinitive, 139 Leira,V., 115, 116
inflection, 190 van der Lely, H. K. J., 162
initial and secondary foci, 30 Lendu, 221
initial focus, 24 Leonard, L. B., 153
inner subject, 17 level of specificity, 18
input factors, 134, 142, 146, 167 levels of organization, 42
input layer, 171 lexical connections, 99
instantiation, 234, 237 Li,C. N., 12, 14
integrated, 22 Lie, S., 86, 109
interpersonal communication, 217 Lieber, R., 97
irregular forms, 190 light markers, 224
irregular inflection, 131 Lobben.M., 221
irregular verbs, 83, 193,198 Lojenga, K., 221
irregular weak verbs, 84 Lord, C., 211,212
Italian, 146, 149,190 Low German, 113, 115, 116, 117
Luiseno, 15, 17
Liibben, A., 116
Jackendoff, R., 113
Jaeger, J. T., 130
Janda,L. A., 218, 223 MacWhinney, B., 41
308 Index
Maldonado, R., 223 non-processual relations, 21

malefactive, 203,213 Noreen, A., 86
Mandarin, 212 Norwegian, 13,73- 86, 90- 95,
Manney, L., 223 100-104,109-117,126,129-
Marantz, A., 97 136,140-149,167,223
Marchman, V. A., 130, 131, 150 Noye,D.,210
Marcus, G. F., 130-131, 166,183, Nynorsk, 74,78, 83, 85,93-94, 104
186, 189 108
Marslen-Wilson, W. D., 150, 199
Matcovich, P. F., 130,146,150
Matsumoto, Y., 52 object, 23
Matthews,?. H.,97, 108, 111 objectivist philosophy, 95
McCarthy, J. J., 97,108, 113 Old Norse, 86
McClelland, J. L., 130,131 van Oosten, J., 47
mental contact, 24 Orsolini, M., 150,199
mental lexicon, 99 Osam.E. K.,212
mental scanning, 50 Osthoff, H., 112
metaphor, 205 outer subject, 17
metaphorical extension, 206 output layer, 171
metonymy, 26,40 overgeneralisation, 131,165, 180
middle, 234,235,238 overgeneralisation errors, 141, 156,
morpheme-based structuralism, 97 185
morphemes as emergent, 111 overlapping, 105
morphemes as epiphenomenal, 111 overlapping items, 101, 104
morphological ambiguity, 101
Morphological Race Model, 199
morphological redundancy rules, 113 passive, 234-235,238
movement, 217 past participle, 190
past tense, 130, 153, 165, 193
patient, 234
Naes, O., 86 Pederson, E. W., 226,232
Nesset, T., 98, 239 Perlmutter, D. M., 49
network, 17Ί, 234 perspective, 18
network in the mental lexicon, 113 phonestheme, 109,112
network models, 189 phonological coherence, 136, 144,
network training, 173 155
Newman, J., 204, 212,217-218, 220 phonological difference, 136
Ngiti, 221 phonological factors, 135
Nichols,!., 13 phonological openness, 136, 144,
Nida, E., 97, 98, 99 148
Nida's principles, 98 phonological properties, 143, 192
N0klestad, A. 92, 130, 146, 167 phonological residues, 101
nonce items, 195 phonological similarity, 136, 145,
nonconcatenative morphology, 122 195, 199
Index 309
phonological structure, 18 Ragnarsdottir, H., 129-130, 134,

pidgin, 204 136,149-150, 154,167-168,
Pike, K. L., 215 170, 177,182-183, 186-187
Pinker, S., 131,165-166, 186, 189 Rappaport, G., 13
pivot, 12 recipient / benefactive, 217
Plank,F., I l l recipient, 204, 227
Plunkett, K., 129-131, 134,150, 154, reciprocal, 228
166- 168, 170, 177, 182-183, redundancy rule, 113
186-187 reference point, 24, 30
point of access, 35 reference point chain, 31, 36
poles, 18 reference tracking, 40
Polish, 218 reflexive, 227,234-235, 238
polysemy, 205,217 reflexive-middle-passive systems,
portmanteau words, 105 223
position verbs, 81 regular forms, 190,198
possession, 40 regular inflection, 131
possessive constructions, 26 regular verbs, 83, 193
Postal, P. M., 52 relationship, 20
Prasada, S., 131,165, 166,186, 189 residual morphs, 101
present tense, 139 Rhodes, R. A., 109,110
Prideaux, G. D., 106 Rodman, R., 97
primary and secondary clausal topic, Romani, C, 199
12 root-forming morphs, 110, 112
primary and secondary focal rule, 162
participants, 30 Rumelhart, D. E., 130, 131
primary focal participant, 21 Russian, 13,223
Prince, A., 131,166,186
process, 21
processing time, 24 Saether0,E.,211,221
product-oriented definition, 90 salience, 143, 149, 182
profile, 19 Saramaccan, 216
profile determinant, 22 Say, T., 199
profiling, 19 Scandinavian, 116
prominence, 13, 18 schema, 112,148,160,162, 171,
pronominal anaphora, 26,40 232
pronouns, 224 schematic interaction, 217
prototypical act of giving, 206 schematicity, 18
prototypical GIVE scene, 204 Schoorlemmer, M., 239
Pulaar, 220 Schreuder, R., 199
secondary focus, 25
secondary focal participant, 21
radial category, 232 segmentability, 182
semantic structure, 18
Senufo, 212
310 Index
serial verb constructions, 210

serializing languages, 203, 210 Talmy, L., 52
Shibatani, M., 49 target, 24, 30
Simonsen, H. G., 73, 86,92, 115, Taylor, F. W., 206
127,132-135, 154-155, 161, Taylor, J. R., 27
167-170,177-187 telescoped words, 105
single mechanism account, 130 Thai, 212
single mechanism models, 148, 186, thematic vowel, 190
189,198 thetic judgement, 44
situation type, 224, 225,226 thing, 20, 21,27, 204
Siya, 212 Thompson, S. A., 12, 14
Skousen, R., 131, 199 Thorell.O., 116
Slobin, D. I., 129, 130, 134, 155, 159 Timberlake, A., 13
smaller weak (WS) class, 86, 90, tlearn, 173
136,154 token frequency, 135, 140, 145-146,
Söderbergh, R. 116 148-149,155, 173, 187,193,
source-oriented, 90 199
South East Norwegian phonology, Tomblin, J. B., 153
74 Tomlin, R. S., 33
specific language impairment, 153 tönernes, 76
spotlights of focal prominence, 31 topic construction, 28
Sranan,211,214 trajector, 124, 21, 30
starting point, 41 trajector/landmark alignment, 46
Steele, S. M., 15 Twi, 213
vanStek, R. 118 type frequency, 90,135,143, 148,
strong inflection, 131 155,191,193
strong verbs, 83, 86,136
structuralist linguistics, 96
subject, 23 Ullman, M. T., 162
subject honorification, 15 unique morphs, 101
subjective motion, 19 usage-based, 149
submorphemic morphology, 106
submorphemic relations, 101
superheavy marker, 225 vowel class, 90
Surselvan, 225 Vangsnes, 0. A., 109
Sveen, A., 161 Vannebo, K. I., 86, 109
Swahili, 210 V-class, 86
Swedish, 110, 115-116 verbal extensions, 206
Sweetser, E., 52 Verfaillie, K., 34
Sylla, Y., 220 Vietnamese, 212
symbol, 160 viewpoint shift, 50
symbolic, 18 vocabulary size, 176
symbolization, 40 vocabulary spurt, 173
vocalic conjugation, 86
Index 311
weak inflection, 131 word tone, 132, 145, 149

weak verbs, 83, 90 WS class, 86
Western, A., 86,94
Williams, E., 97
Winter, W., 112 Yao Samsao, 212
WL class, 86 Yoruba, 212,215
word accents, 76, 78, 132

Untitled

Uploaded by

Copyright:

Available Formats

Untitled

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Untitled

Uploaded by

Copyright:

Available Formats

A Cognitive Approach to the Verb

© Printed on acid-free paper

Die Deutsche Bibliothek — Cataloging-in-Publication Data

A cognitive approach to the verb : morphological and construc-

Section I: Basic issues

Topic, subject, and possessor 11

Language as a cognitive network 49

Section II: Morphological aspects of the verb

The Norwegian verb 73

Building blocks or network relations:

Past tense acquisition in Norwegian:

Individual variation in past tense inflection:

A connectionist model of past tense acquisition

Regular inflection in the mental lexicon:

Section 111: Constructional aspects of the verb

Malefactive by means of GIVE 203

The Norwegian and Russian reflexive-middle-passive

Norwegian reflexive-middle-passive marker -s:

Control and transitivity: A study of the

In June 1998, a small conference took place at Sanner Hotel, Gran,

In their papers, the two keynote speakers provided a broad

2. Section I: Basic issues

The first section starts with Ronald W. Langacker: Topic, subject,

Grammar (Langacker 1987a, 1990, 1991), Goldberg's Construction

3. Section II: Morphological aspects of the verb

This section contains six papers, most of them concerning

relations, and he posits Langacker's and Bybee's theories as

modelling results show that a single mechanism is able to learn to

4. Section III: Constructional aspects of the verb

and Russian, and show how they can be presented in Cognitive

5. Final remarks and acknowledgements

The editors of the Cognitive Linguistics Research series gave very

1. Reviewing is a time-consuming job. We wish to thank the following persons

Language offers ample evidence that the grammatical notions topic,

1. Evidence of special affinities

proposed in cognitive grammar (Langacker 1991:7.3). It also bears

(1) a. Your friend, he will never amount to anything.

Data showing an intimate relationship between subject and topic in

(2) Booth's assassination [of Lincoln];

Another such phenomenon, found in French, Norwegian, and

(3) a. II leve la main. / Han rekker opp handa.

Phenomena specifically relating topic and possessor may be less

(4) a. Nicole, Pierre, eile ne I'aimepas.

(5) a. Ibu anak itu membeli sepatu.

(6) a. Apeci-ka Kim sensayngnim-ul mana-si-ess-ta.

Control of the reflexive marker casin works analogously.

(7) a. Noo=n no-puush konoknish.

These expressions have the following abstract form:

(8) [ noo-up [ no-te' tiiwu-q ] ]

Depending on the language, there is evidence for both the "outer

2. Some basic notions of cognitive grammar

Naturally, I will approach this question from the standpoint of

(9) thing>creature>bird>bird of prey>hawk>red-tailed hawk

The meanings of grammatical elements tend to be quite schematic.

(10) a. A big crack extends from the corner of the window up to

These sentences describe precisely the same, static situation—

(a) Base iris pupil

(") Base admire admirer

O-O O-«O O--O tr Im

Figure 1. Base and profile