See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/318661099

Cognitive validity

Chapter · January 2011

CITATIONS READS

0 438

1 author:

John Field
University of Bedfordshire
29 PUBLICATIONS   1,038 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Rethinking the Listening Test: From Theory to Practice. Publication: Equinox, early 2018 View project

Book published Cambridge University Press 2008 View project

All content following this page was uploaded by John Field on 24 July 2017.

The user has requested enhancement of the downloaded file.

3
Cognitive validity
John Field
University of Reading

This chapter considers the cognitive validity of the Speaking tasks which
feature in the Cambridge ESOL suite. By ‘cognitive validity’ is to be under-
stood the extent to which the tasks in question succeed in eliciting from can-
didates a set of processes which resemble those employed in a real-world
speaking event. A second consideration is how finely the relevant processes
are graded across the levels of the suite in terms of the cognitive demands that
they impose upon the candidate.
Previous volumes in this series have considered the cognitive validity of
Cambridge ESOL tests of writing and reading. This chapter can therefore
draw upon the format and approach already established in Chapters 3 of
Shaw and Weir (2007) and Khalifa and Weir (2009). As with the earlier
analyses, a major goal is to propose a cognitive model of the construct in
question, which can serve as a framework for judging the cognitive valid-
ity of any test of skilled performance. The model is not drawn from testing
theory but from independent insights afforded by empirical research into the
psychology of language use.
It needs to be borne in mind that this is the first volume of the series to
tackle the spoken modality. While there are certain parallels in the process-
ing model presented here – especially with the model of writing proposed by
Shaw and Weir (2007) – there are also marked differences. Perhaps the most
important distinguishing characteristic of the oral skills is that they typically
operate under tight time constraints that are not usually present in writing
and reading. An interactional speaker self-evidently does not have time for
planning and revising in the way that a writer does. Due account will be taken
of this important difference in the discussion that follows.
The chapter falls into three parts. Firstly, there is an explanation of the
general notion of cognitive validity, which retraces some of the points made
in earlier volumes but does so with specific reference to the present exercise.
Next, some background is provided to research findings on the nature of
L1 speaking; and a process account of the skill is proposed, drawing chiefly
upon the model devised by Levelt (1989). The Levelt model is adapted to
provide a five-part framework for an examination of the cognitive valid-
ity of the tests in the Cambridge ESOL suite. The discussion then goes on
to consider two important characteristics of the types of speech elicited by

65
Examining Speaking

the tasks in the text. Clearly, this entails some consideration of task design,
leading us into a grey area at the interface between cognitive and context
validity. But the goal here is to consider these particular features strictly in
terms of any likely additional cognitive demands which they impose upon
the candidate.

The cognitive validation exercise

Cognitive validity: a rationale
An important consideration in establishing the validity of tests that aim to
measure performance is the extent to which the task, test content and pre-
vailing conditions require the taker to replicate the cognitive processes which
would prevail in a natural (i.e. non-test) context. This aspect of validity is
especially critical in the case of tests that assess general competence in any
type of skill across a proficiency continuum – a category into which the
Cambridge ESOL Main Suite of General English tests (KET–CPE) clearly
falls. The concern of these exams is with the test taker’s ability to apply the
four language skills at varying levels of proficiency, rather than with the
simple measurement of linguistic knowledge. Furthermore, their value to
a user of the test scores lies in their predictive power: their ability to indi-
cate how competently a candidate might be expected to perform in actual L2
contexts. Thus it becomes important for test producers to know, and to be
able to demonstrate, how far what happens in the testing situation replicates
cognitive processing in the world beyond the test, so that test users can have
confidence in the meaningfulness and usefulness of the score outcomes from
the test.
Interest in the psychological processes underlying language test perform-
ance has a long history, dating back at least as far as Carroll (1968); but it
received new impetus recently from the work of Cyril Weir. Weir (2005a)
expresses concerns over traditional post hoc approaches to investigating
construct validity, where statistical methods such as factor analysis are
applied to test results in order to establish the nature of the construct that
has been tested. He raises the issue of whether the data derived might to
some extent be compromised by the form and content of the test and by the
assumptions underlying its design. To put it in simple cognitive terms, he
identifies the dangers of relying exclusively on an approach that attempts
to track back from a product or outcome to the process that gave rise to it.
Instead, he argues for what he terms theory-based validity (or more recently
cognitive validity): a complementary approach to test validation which
takes account, at an initial stage, of empirical evidence as to the nature of
the construct that the test aims to assess. Weir (2005a:18) makes his point
powerfully:

66
 Cognitive validity

There is a need for validation at the a priori stage of test development.

The more fully we are able to describe the construct we are attempting to
measure at the a priori stage, the more meaningful might be the statistical
procedures contributing to construct validation that can subsequently
be applied to the results of the test. Statistical data do not in them-
selves generate conceptual labels. We can never escape from the need to
define what is being measured, just as we are obliged to investigate how
adequate a test is in operation.

This additional strand of construct validation attempts to ensure that,

besides benefiting from feedback from piloting and past administrations, test
design also draws in a principled way upon external evidence concerning the
nature of the expertise which is to be targeted. For a more detailed account
of the arguments, see Chapter 3 of Examining Reading (Khalifa and Weir
2009).
This is not a simple matter of ecological validity. The goal is to establish
whether the tasks proposed by a test designer elicit mental processes resem-
bling those which a language user would actually employ when undertaking
similar tasks in the world beyond the test. The processes in question might
relate to the way in which the user assembles or interprets input; or they
might reflect the cognitive demands imposed upon the user by facets of the
task.

The cognitive perspective

The focus in this chapter is thus upon process rather than product, the appli-
cation of the skill rather than the linguistic knowledge that underpins it.
Consequently, the framework that is to be outlined will rely chiefly upon
theory and evidence taken not from testing studies or from general linguistics
but from psycholinguistics and speech science. Within those fields, two areas
of enquiry are particularly relevant to the present exercise. The first con-
sists of research on language production: specifically, on how speech is stored
and assembled by the user. The second concerns the nature of expertise; the
assumption here being that the ability to produce speech fluently and with a
high degree of automaticity is a type of expert behaviour which has certain
elements in common with playing chess or driving a car.
Because second language acquisition (SLA) commentators occasion-
ally misunderstand what cognitive approaches to language skills entail, it
may be helpful to provide a brief explanation of the thinking that under-
lies them. It is mistaken to suggest that, in examining language processing,
cognitive scientists ignore issues of context or deny their relevance. On the
contrary, they concern themselves quite closely with the types of decision-
making which occur when an utterance is adjusted to take account of
listener-relevant factors, such as illocutionary intent or shared knowledge

67
Examining Speaking

(Levelt 1989: Chapter 1). They also take into account the possible impact
upon a speaker’s attentional resources of affective factors such as tired-
ness or anxiety. But their principal focus is on the mental operations which
are engaged by language users under normal circumstances and the way in
which situation-specific information of various kinds can be integrated into
those operations.
It is also incorrect to assert that cognitive psychologists rely upon the
assumption that all language users behave identically. While certain process-
ing routines may provide the easiest and most efficient routes to language
production and reception, some users (including L2 learners) achieve the
same goals by less direct means. It is evident that individuals, whether speak-
ing in L1 or L2, vary enormously in the range of vocabulary they command
and in their powers of expression. It is also evident that L2 speakers respond
in very individual ways to the challenges posed by an inadequate lexical or
grammatical repertoire. Similarly, there is no suggestion that test designers
can afford to ignore the influence of factors arising from individual speaker
differences of age, gender, ethnicity, first language background, etc., as
discussed in the previous chapter on test taker characteristics.
Nevertheless, the premise is adopted that underlying the four language
skills are certain established and shared routines which can be traced by
examining and comparing the performance of expert language users. This
assumption is supported by two lines of argument:

a) The universal argument. All human brains are similarly configured.

They can be assumed, at some level of generality, to share processing
routines which are broadly similar in that they reflect the strengths
and limitations of the organ and the means it adopts for transmitting
information. These routines might be deemed to contribute not simply
to the forms that language takes but also to the ways in which it is
processed in performance.
b) The expertise argument. A marked difference between an adult L1
speaker and an L2 learner lies in the fact that the former has had many
years of experience during which to develop the most rapid and most
effective processing routines for dealing with the vagaries of the target
language – and to develop them without competition from deeply
ingrained routines associated with another language. An understanding
of how such expert users perform should thus assist us in directing the
development of novice users. The novice/expert distinction is not, of
course, an all-or-nothing one. There exists a continuum of expertise
stretching from novice to expert user, which is highly relevant in the
context of language proficiency assessment, where gradations of ability
need to be distinguished and accredited for teaching and learning,
employment or other social purposes. As Chapter 1 explained, a

68
 Cognitive validity

fundamental aim of this volume (and of the companion volumes in

the series) is to improve our understanding of the key criterial
parameters that differentiate one proficiency level from another, and
to provide empirical (as well as intuitive and experiential) evidence in
support of claims that a test or a test taker is or is not at a particular
level.

In line with the notion of ‘cognition’ that has just been outlined, it will be
assumed in considering the speaking skill that variation between tasks due to
situation or genre falls outside the remit of cognitive validity (see Chapter 4
on context validity). Similarly, the discussion will in the main exclude consid-
eration of social-affective factors which might influence the content or deliv-
ery of the speaker’s utterances; these features are the concern of social rather
than cognitive psychology and fall under the speaker-specific aspects covered
in Chapter 2. To the extent that the present chapter considers linguistic dif-
ficulty at all, it does so purely in terms of the cognitive effort which the assem-
bly of a syntactic structure or the retrieval of a piece of vocabulary might
require. More general issues of linguistic accuracy, fluency and complexity
are covered in Chapter 4.

The scope of the chapter

The direction taken in this study conforms closely to Weir’s original (2005a)
concept of theory-based or cognitive validity. Broadly, the approach to vali-
dation as it has been operationalised (Shaw and Weir 2007, Khalifa and Weir
2009) entails constructing an empirically attested model of the target skill
as employed by expert users under non-test conditions; then relating the
processes which feature in the model to the specifications of the test under
examination. The models presented in the previous accounts of writing and
reading were based upon information-processing principles, thus enabling
the researcher to identify specific phases through which a language user nor-
mally proceeds (though the point is well made by Shaw and Weir that the
phases are not necessarily sequential and that not all of them are obligatory).
These phases provide a framework for determining in a systematic way how
the various processes which make up performance in a skill are represented,
explicitly or implicitly, in the test criteria.
While adhering closely to this line of attack, Khalifa and Weir (2009)
extended the scope of cognitive validation by considering not simply the
standard processes involved in employing the language skill in question, but
also the varying cognitive demands placed upon the candidate by the range
of tasks that are specified in the test format. This is especially apposite in the
case of reading, since it enables consideration of the different types of reading
elicited by the tasks and their relative complexity.

69
Examining Speaking

Tests of L2 speaking afford parallels in the form of two important task

variables which determine the cognitive demands placed upon the candidate
and the extent to which a task can be said to replicate a real-life speaking
process. They are:
• the nature of the interaction between speaker and interlocutor
• the extent to which planning is permitted.
Following the presentation of the processing model and its application
to the specifications of the Cambridge ESOL suite, there will therefore be a
brief consideration of the ways in which the suite handles these two aspects
of speaking in its task design. The discussion will adopt a strictly cognitive
perspective, to avoid overlap with the more extensive examination of task
design in Chapter 4.

In sum, the purpose of this chapter is:

i) To consider the nature of cognitive processing involved in first
and second language speaking by reviewing the relevant academic
literature, both theoretical and empirical.
ii) To propose a model or framework against which speaking tests,
including those produced by Cambridge ESOL, could be analysed,
described and evaluated for the purposes of cognitive validation.
iii) To apply the model to a specific set of Cambridge ESOL Speaking
tests at the different Main Suite proficiency levels in order to generate
an a priori descriptive analysis of the cognitive processes underlying
the test takers’ performance.
iv) To consider other cognitive factors which might determine the
relative difficulty of tests at different levels in the Cambridge ESOL
suite. These relate to task variables which can be seen as adding to or
alleviating the cognitive demands made upon the candidate.

Psycholinguistic accounts of speech production

Speaking is one of the most complex and demanding of all human mental
operations. Levelt (1989:199) points out that an expert speaker of English
produces utterances at a typical rate of about 150 words per minute, or 2.5
per second. Under pressure, the rate can rise to five per second. A normal
educated adult L1 speaker might have an active vocabulary of about 30,000
words. This means that a fluent speaker makes the right choice from these
30,000 alternatives between two and five times per second, and maintains this
rate of performance without any clear time limit, other than the need to cede
the turn at some point to an interlocutor. Levelt comments: ‘There is prob-
ably no other cognitive process shared by all normal adults whose decision

70
 Cognitive validity

rate is so high.’ All the more reason, then, for those who design tests of speak-
ing to have a detailed understanding of the nature of the skill and of the proc-
esses that contribute to it.
Extensive empirical research (psychological, neurological and phonetic)
into all aspects of the skill of speaking has enabled commentators to achieve
a fair degree of consensus as to the processes engaged when individuals
assemble a spoken utterance in their first language. Early psycholinguistic
research in the 1960s into producing and analysing spoken language con-
cerned itself greatly with syntactic structure and with the extent to which the
rules of grammar (particularly Chomskyan grammar) might have psycho-
logical reality (i.e. represent the processes in which a speaker engages when
producing an utterance). These enquiries indicated that the clause was an
important unit of speech assembly; but proved otherwise generally incon-
clusive. (See Aitchison 2008 for an accessible extended discussion.) It was at
this point that many psycholinguists turned to an evidence-driven approach
to speech production in preference to one that was largely shaped by estab-
lished linguistic theory. One line of enquiry built upon emerging evidence
from phonetics in respect of phenomena such as pausing. It became clear that
brief pauses (generally of 0.2 to 1.0 seconds) were necessary for the forward
planning of speech; and the location of these planning pauses was found to
correspond quite consistently with syntactic boundaries, again implicating
the clause as a unit of assembly. One theory (Beattie 1983) suggested that the
length of planning pauses varied according to whether the planning in ques-
tion related solely to the form of the next utterance or whether it additionally
anticipated the conceptual content of later utterances.
A second line of enquiry focused especially on the errors made by natu-
rally performing speakers, known as slips of the tongue. The notion was that
by examining failures of the speech assembly system, one might gain insights
into the cues that speakers were using in aiming for their targets. Slips of the
tongue provided evidence suggesting that a syntactic frame was prepared by
a speaker in advance of lexical items being slotted into it (He found a wife
for his job, substituted for the target He found a job for his wife.) and that
morphological markings were added at quite a late stage (She come backs
tomorrow)1. These findings provided the basis for an early model of speech
production by Garrett (1980, 1988), which drew also upon research into the
speech impairments associated with aphasia. Important in Garrett’s model
were the assumptions a) that a preliminary structural frame is established
into which the outcomes of a parallel lexical search are inserted; and b) that
there is an initial planning phase where the syntactic framework and the links
to lexis are abstract, followed by a phase where they are realised concretely in
terms of word order and phonological word form.
A problem to which Garrett and others gave much thought was the
extent to which syntactic assembly was so closely intertwined with semantic

71
Examining Speaking

considerations that it could not be treated as a distinct operation. The advent

of brain imaging has shed new light on this vexed question; neuroscientists
(Kutas, Federmeier and Serreno 1999: 366–368) have succeeded in identify-
ing an event-related potential (a pattern of electrical activity in the brain)
that is distinctively associated with syntax and a separate one that is associ-
ated with semantic processing. This suggests that the two processes may also
be distinct in psychological terms.
Slips of the tongue also provided valuable insights into the process of
lexical retrieval – i.e. the way in which a speaker uses a target meaning in
order to locate a particular word in their mental lexicon. Examples such as
white Anglo Saxon prostitute (= Protestant) or I’ve been continuously dis-
tressed (= impressed) by her (Fromkin 1973) indicate that language users
employ more than meaning in their word search; they also draw upon
certain intimations about the form of the word they are seeking (its length,
stress patterns, initial syllable etc.). See Aitchison 2003 (16–23, Chapter 12)
for a discussion.
On the basis of this and other evidence, a number of researchers since
Garrett have proposed models of speech production, the most widely cited
being that of Levelt (1989). Some of the models relate specifically to the way
words are accessed and articulated, including WEAVER (Roelofs 1997) and
the framework proposed by Bock and Griffin (2000), which incorporates the
many variables that determine the speed and ease of lexical retrieval.
Models of speech production, like many others in cognitive psychology,
resemble flow charts. They represent speaking in terms of a series of stages
through which a speaker needs to proceed when assembling an utterance.
Their point of departure is an idea in the mind, and the end-product is a
disturbance of the air caused by the operation of the speaker’s articulators
(tongue, teeth, lips etc.). For the sake of convenience, the stages are often
represented as sequential; but the point should be made at the outset that
one stage does not necessarily wait upon another. On a principle sometimes
referred to as incremental production (Kempen and Hoenkamp 1987, Levelt
1999:88), it is widely accepted that material can be passed on to the next
stage of processing, even though it is incomplete. Let us say that so far only
the first words of a sentence have been encoded phonologically; if they form
a unit (e.g a Noun Phrase such as The man I met . . .) they can be submitted to
the next level for phonetic encoding as a cluster or even an intonation group.
In addition, current thinking is that language skills such as speaking are
based upon highly interactive processes, in which information flows in a
top-down direction as well as a bottom-up one. To give a simple example,
in assembling a sentence such as ‘Any idea of the time?’, a speaker would not
have to proceed word by word and phoneme by phoneme, but could draw
upon stored memories of the way in which the sentence as a whole had been
achieved the last time it was uttered and the time before.

72
 Cognitive validity

Levelt (1989, 1999) makes clear that any model of speech production,
whether in L1 or in L2, needs to incorporate a number of stages. Field
(2004:284) identifies them as:
a) a conceptual stage, where the proposition that is to be expressed first
enters the mind of the speaker
b) a syntactic stage, where the speaker chooses an appropriate frame into
which words are to be inserted, and marks parts of it for plural, verb
agreement etc.
c) a lexical stage, where a meaning-driven search of the speaker’s lexicon
or vocabulary store takes place, supported by cues as to the form of the
word (e.g. its first syllable)
d) a phonological stage, where the abstract information assembled so far is
converted into a speech-like form
e) a phonetic stage, where features such as assimilation are introduced,
which reduce articulatory effort; and where the target utterance is
converted into a set of instructions to the articulators
f) an articulatory stage, in which the message is uttered.
It is important to note that the first three of these stages are abstract
and not in verbal form. It is only at stage (d) that linguistic forms become
involved. A model of speaking also needs to allow for:
• a forward planning mechanism at discourse level, which (for example)
marks out in advance which syllable is to carry sentence stress
• a buffer, in which an articulatory plan for the current utterance can be
held while the utterance is actually being produced
• a monitoring mechanism, which enables a speaker to check an utterance
for accuracy, clarity and appropriacy immediately before it is uttered
and almost immediately afterwards.

A cognitive processing framework for speaking

Levels of analysis
As already mentioned, the most comprehensive account of L1 speech pro-
duction to date is the one provided by Levelt (1989), drawing upon some 15
years of empirical research. There are a number of reasons for proposing this
model as the basis for the cognitive validation of L2 speaking tests. Firstly,
among psycholinguists and speech scientists, it is by far the most widely
cited, and has informed many recent research studies. Secondly, it is a staged
model, with various levels of processing specified in a way that provides a
means of analysing test specifications in a systematic manner.

73
Examining Speaking

Importantly from the point of view of the present exercise, it contributed

to Weir’s initial socio-cognitive framework for validating speaking tests.
Weir’s set of constituent processes (2005a:46) employs a number of Leveltian
terms (conceptualisation, formulation and articulation), as well as the notion
of a target message being transformed from pre-verbal form to phonetic plan
and then to overt speech. The Levelt model also quite closely resembles the
process model for writing outlined earlier in this series by Shaw and Weir
(2007). There is a good reason for this. The Shaw and Weir model was based
upon Kellogg (1996), with additions proposed by Field (2004); and both
Kellogg’s categories and Field’s modifications were strongly influenced by
Levelt’s model of speaking. Consequently, the reader familiar with the Shaw
and Weir volume will note close parallels between the processes for assem-
bling written texts identified there and those that are discussed here in relation
to speaking, though there are some important differences in the terms used.
Levelt’s original 1989 model features four major stages (conceptualisation,
formulation, articulation and self-monitoring). However, an updated version
(1999) divides ‘formulation’ into two, distinguishing between an operation
that provides a general framework for the utterance and one that converts
this abstract plan into phonological form. It also recognises two aspects of
‘articulation’: creating a set of instructions to the articulators (lips, tongue,
larynx etc.) and then carrying out those instructions. The stages that we will
consider here are thus as follows:
• Conceptualisation: generating an idea or set of ideas for expression
• Grammatical encoding: constructing a syntactic frame and locating
the lexical items that will be needed (similar to micro-planning as
conceptualised by Shaw and Weir 2007 in respect of writing)
• Phonological encoding (Levelt’s 1999 term is morpho-phonological
encoding): converting the abstract output of the previous stage into a
string of words which are realised phonologically (the equivalent of
translation in Shaw and Weir 2007)
• Phonetic encoding: adjusting the phonological sequence to make
articulation easier; linking each of the syllables to a set of neural
instructions to the articulators; storing the instructions in a buffer while
the clause is being articulated
• Articulation: producing the utterance (the equivalent of execution in
Shaw and Weir 2007)
• Self-monitoring: focusing attention on the message immediately before
and shortly after it is uttered in order to check for accuracy, clarity and
appropriacy (a rough equivalent to editing in a writing context).
Figure 3.1 presents a version of Levelt’s model of speech assembly, which
has been modified considerably to meet the needs of the present enquiry. It is

74
 Cognitive validity

Figure 3.1 Adapted version of the Levelt model (1989:9), separating levels of
processing from outputs of processing

CONCEPTUALISATION pre-verbal message

GRAMMATICAL ENCODING
constructing a syntactic frame abstract surface structure
forming links to lexical entries

MORPHO-PHONOLOGICAL
ENCODING phonological plan
Conversion to linguistic form

PHONETIC ENCODING
Conversion to instructions to articulators phonetic plan
Cues stored in a speech buffer

ARTICULATION
overt speech
Execution of instructions

SELF-MONITORING self-repair

important to recognise that the model (1989:9, 1999:87) represents two dis-
tinct types of phenomenon: a) the set of processes employed by the speaker
in assembling an utterance; and b) the different forms taken by the message
as it is reshaped by the intending speaker. In the interests of clarity, the two
are displayed here in separate columns. The reader should bear in mind
that what is shown as the output of a given stage also forms the input to the
following one.
The model shown here lacks an important component of the reading
model proposed by Khalifa and Weir (2009), in that it does not show an exec-
utive mechanism or ‘goal setter’ which controls and directs the attention of
the speaker, takes account of context and prevailing circumstances and sup-
ports decision making. Levelt himself (1989:20–22) acknowledges the role of
such a mechanism in conceptualisation and in self-monitoring but stresses
the fact that elsewhere speaking is heavily dependent upon processes that are
highly automatic and thus not subject to central control.
Nevertheless, to bring the model into line with the Khalifa and Weir
account, it is useful to make clear the information sources upon which the
speaker draws (whether automatically or with a degree of intentionality)
when assembling an utterance. The relative richness or poverty of those
sources is clearly an important factor in shaping the performance of a second

75
Examining Speaking

language listener. Describing L1 speech assembly, Levelt (1999) identi-

fies three main information sources: a mental lexicon, in which the speak-
er’s vocabulary repertoire is stored; a syllabary which stores an experiential
record of the articulatory gestures which enable the speaker to utter all the
syllables of the language; and the individual’s store of world knowledge. To
this, a more detailed account would add some kind of syntactic store, though
one can remain neutral as to whether it consists of the type of grammar rule
favoured by linguists, of a set of stored examples or of procedural rules which
enable the rapid assembly of phrases. It also seems important to clearly sepa-
rate conceptual knowledge of the world, the situation and the addressee from
the speaker’s discourse representation, or record of what has been said so far
in the conversation.
Clearly, the quality and availability of all these information sources affect
the individual’s ability to construct accurate and appropriate L2 utterances.
A performance deficit in a test of L2 speaking might arise from:
• linguistic sources: gaps in the mental lexicon, imprecise or incomplete
representations in the syllabary, inability to encode a syntactic pattern
into a form of words
• knowledge sources: cultural gaps in world knowledge or pragmatic
knowledge
• failures of comprehension or recall which leave gaps in the discourse
representation.
A relatively detailed model of speaker information sources is shown
in Figure 3.2. Those that can be considered part of linguistic knowledge
are marked with a grey tone; those not so marked relate broadly to world
knowledge, experience or recall of the conversation so far. The distinction
in the figure between current topic and discourse representation merits a brief
comment. Levelt (1999) refers to the role of current topic in generating a new
utterance in interactional speech or in licensing a new topic that is thematically
linked. It also plays a part in the decision to employ an anaphor (for example
a pro-form such as she, it, this or did so) to refer to an entity or idea currently
foregrounded in the minds of both speaker and listener. The discourse rep-
resentation carried forward by both speaker and listener has an altogether
wider role. It constitutes a record of all that has been said so far in the con-
versation, enabling either party to refer back or to check a new comment for
consistency with what has already been said or for general relevance.

The validation framework in an L2 context

The stages of the adapted Levelt model will now be explained in some detail,
with particular attention given to the factors which may shape the perform-
ance of L2 speakers under task conditions like those in the Speaking tests of

76
 Cognitive validity

Figure 3.2 Information sources feeding into the phases of the processing system

Speaker’s general goals

World knowledge
Knowledge of listener
Knowledge of situation
Recall of discourse so far CONCEPTUALISATION

Rhetoric
Discourse patterns

Recall of ongoing topic

Syntax
Lexical knowledge (lemma*)
Pragmatic knowledge GRAMMATICAL
Knowledge of formulaic chunks ENCODING
Combinatorial possibilities
(syntactic / collocational)

Lexical knowledge (lexeme*) PHONOLOGICAL

Phonological knowledge ENCODING

Syllabary: Knowledge of
PHONETIC ENCODING
articulatory settings

[ARTICULATION]

Speaker’s general goals

Target utterance stored in buffer SELF-MONITORING
Recall of discourse so far

* In the 1999 update of his model, Levelt distinguishes between three components in a lexical
entry in the mind storing information about a word. There is a semantic component which
enables a match to be made between a meaning and the target word; a lemma containing
syntactic information about the word (its word class and combinatorial possibilities); and a
lexeme containing information about the word’s phonological form and morphology.

the Cambridge ESOL suite. Throughout the discussion, we need to bear in

mind that, though language users assemble speech with astonishing facility
given the complexity of the skill, certain limits still apply to their perform-
ance. Most obviously, there are physiological constraints which make it

77
Examining Speaking

impossible for the articulators to achieve a speech rate of more than about
eight syllables per second (Miller 1951). But there are also psychological
constraints: the human processor (responsible not just for language produc-
tion and reception but for handling all kinds of mental task) is characterised
by its limited capacity. What this entails is that an increase in the demands
created by one aspect of a task will limit the performer’s ability to deliver in
other areas. Thus, it is very difficult for a language user to speak and write at
the same time or for an individual to memorise a set of facts while repeating
nonsense words.
The limitations upon what can be held in the mind short-term have two
important implications for any discussion of the role of the second language
speaker:
a) For a low-proficiency L2 speaker, the process of retrieving words
and syntactic patterns is much more effortful than it is with a native
speaker – potentially limiting performance in other areas of the speaking
process, such as the ability to hold long-term plans in the mind.
b) The complexity of the task that is set can affect performance. The more
difficult the task, the more attention the speaker gives to handling it and
the less will be available for the delivery of speech.
So, in any consideration of the processes in which test takers engage, full
account needs to be taken of these cognitive demands upon them.

Conceptualisation
Levelt (1989:107) envisages conceptualisation as entailing two types of
operation:
• macro-planning, in which a set of speech acts is anticipated
• micro-planning, at a more local level relating to the role and form of the
upcoming utterance.
These resemble the subdivisions proposed by Shaw and Weir (2007) in
their framework for the cognitive validation of L2 writing. However, macro-
planning in speaking is much more constrained than it is in writing. The
speaker is under pressure to respond promptly in most speaking contexts,
thus limiting the time available for planning and structuring content. In
addition, there are working memory constraints on how much longer-term
material the L2 speaker can store while at the same time dealing with current
production demands (unless, of course, the speaker has the support of pen
and paper to record their intentions). Micro-planning is much more local-
ised. It positions the intended utterance in relation to the discourse as a whole
by taking account of knowledge shared with listener and current topic. It also
adds indications of language-specific features such as tense or interrogation
to the proposition that is to be expressed.

78
 Cognitive validity

Levelt’s account of how the message is generated (1989: Chapter 4) views

spoken interaction as a process of joint construction, in which a speaker has
to relate each new utterance to a shared discourse framework. Amongst the
factors which Levelt identifies as affecting both macro- and micro-planning
are: awareness of the ongoing topic, thematisation of new information, rec-
ognition of information shared and not shared with the listener, accommo-
dation to the point-of-view and even the form of words of the interlocutor
and certain basic principles which determine how information is ordered.

Grammatical encoding
As already suggested, Levelt’s original formulation phase can be treated as
falling into two parts. The first entails the construction of a surface struc-
ture, an abstract framework for the sentence to be uttered, based upon a syn-
tactic pattern. The second converts the framework and the associated lexis
to phonological form; this involves retrieving the appropriate forms from
memory.
Levelt views surface structure as built around the major components of
the idea that the speaker wishes to express. Thus, the proposition I put two
pounds in the meter would, in an English speaker, trigger the valency pattern
associated with the word PUT:
[Agent + PUT + thing put + destination].
Syntactic complexity is clearly a factor in the cognitive difficulty of pro-
ducing an utterance. Raters readily assume that inability on the part of an
L2 speaker to form a particular structure derives from a lack of linguistic
knowledge; but it may equally well derive from the demands of assembling
the structure and of retaining it in the mind while the utterance is being
produced.
In tests of speaking, however, linguistic content is often expressed not
in terms of syntactic complexity of form but in terms of the language func-
tions which test takers are required to perform. This makes it simpler to
envisage the transition from the test taker’s initial idea to a template for an
utterance. It can be treated as a matter of mapping from the function that
the test taker wishes to perform to the pattern that best expresses that func-
tion. Discussion of pragmatic language falls mainly within Chapter 4, which
concerns itself with linguistic criteria; but it is important to recognise that
any staging of the functions to be performed is also a staging of cognitive
demands. The issue at stake is not primarily the difficulty of the language
that has to be retrieved, but how easy it is for the test taker to perform
the mapping exercise. Contributory factors might include the frequency and
transparency of the function and the complexity of the form of words that
expresses it.

79
Examining Speaking

Phonological encoding
For the second language speaker, the most critical phase of forming an
utterance is the one at which they retrieve phonological forms from memory
in order to give concrete form to what has been planned. Whereas the
process of retrieval in one’s first language is generally rapid, automatic and
accurate, it is likely to be much slower in a second language, especially at
lower levels of proficiency. The search is likely to be more effortful, requiring
higher levels of attention; and the speaker is likely to be less confident of the
outcome.
A widely favoured view of second language acquisition represents it as the
acquisition of a type of expertise, and traces parallels, sometimes sustainable,
sometimes not, with the acquisition of the ability to drive or to play chess
(Anderson 1983). According to this analysis, a speaker’s ability to retrieve L2
word forms from memory develops as a result of increasing familiarity with
the operation. The speaker begins with a retrieval process which is laboured
and heavily controlled in terms of the attention it demands (compare the
careful step-by-step way in which a driver first learns to change gear). By dint
of continued use of the process, what were once separate steps become com-
bined, and the move from stimulus to output becomes increasingly automa-
tised (Schneider and Schiffrin 1977) until the speaker can achieve it without
the conscious allocation of attention. The development is known as proce-
duralisation. Motivated learners often find a means of assisting it by rehears-
ing forms of the spoken language in their minds in anticipation of speech
encounters that may occur. For an application of the notions of automaticity
and control to second language performance see DeKeyser (2001), Robinson
(2003) and Kormos (2006:38–51)2.
Evidence of increasing proceduralisation in L2 speakers is often marked
by a shift towards producing language in chunks (Wray 2002). What begins
as an utterance assembled piece by piece (I + do + not + know + why) later
becomes articulated as a single unit (dunnowhy). It also appears to be stored
in the mind in this form, enabling more rapid retrieval. By taking this path,
L2 speakers align their behaviour with a native speaker’s (and indeed with
their own behaviour in L1). The speed with which native speakers assemble
grammatically correct sentences can only be explained if we recognise that
the operation relies heavily upon stitching together well-established and
often-used combinations of words. Strong evidence for this comes from the
productions of sports commentators, forced to plan their utterances under
extreme time pressures (Kuiper 1996, quoted in Wray 2002):
They’re off and racing now + threading its way through + round the turn
they come
and from the productions of sound-bite politicians (Frost interview with
Tony Blair, quoted in Fairclough 2000:112):

80
 Cognitive validity

. . . I think it’s sensible + if for example in areas like erm + the constitution or
indeed in respect of erm education it may be + or any of the issues which matter
to the country + you can work with another political party because there are
lots of things we have in common with the Liberal Democrats why not do it.

It should also be noted that the chunking process promotes grammatical

and lexical accuracy. Rather than having to make a series of decisions, any
one of which could misfire, the speaker produces a single unit, ready made
and internally accurate and consistent. In the case of the L2 speaker, we
should thus expect increased chunking to be accompanied by an increase in
syntactic and collocational accuracy.
There is a small problem with this analysis so far as the designers of inter-
national tests are concerned, which should not pass unnoticed. The devel-
opment of speaking skills, as described here, assumes an EFL (English as a
Foreign Language) context, by far the most common situation worldwide.
The learner is taken to begin with word-level information (plus a few stand-
ard and highly conventionalised formulae such as How do you do?) which
then becomes integrated into larger structures. However, learners acquir-
ing a language naturalistically in an ESL (English as a Second Language)
situation often proceed in the opposite direction: first acquiring chunks and
then deconstructing them. Raters may thus need to be more wary of treating
chunking as evidence of increasing oral proficiency in the case of an ESL pop-
ulation. See Fillmore (1979) for data relating to naturalistically acquired L2.
The phenomenon of chunking forms part of a larger theory of second lan-
guage acquisition which is gaining wide acceptance and which has particu-
lar relevance to the speaking skill. There is evidence that the human mind
is capable of storing far more information than we previously supposed. It
has been suggested that our ability to use a language may be the product of
multiple stored traces of utterances that we have heard over the years (for an
introduction, see Dąbrowska 2005: Chapter 2). The principle can be applied
at all levels: explaining our ability to distinguish variant forms of phonemes
and words (Bybee 2000) as well as the reductions to which syntactic patterns
are sometimes subject in the mouths of speakers. On the basis of this analysis,
our stored linguistic knowledge is serendipitous, a product of any exposure
we may have had to the language in question. The communicative compe-
tence underlying the ability to express oneself orally in L2 is an emergentist
phenomenon (N. Ellis 2003), i.e. the result of an accumulation of individual
encounters, both receptive and productive. The issue of frequency (whether
of lexical items or syntactic patterns) comes very much to the fore (Bybee
and Hopper 2001), since the more often a learner comes across a particular
feature, the more detailed is the representation of it that is stored.
Whether one accepts the emergentist argument or not, it is reasonable to
expect a graded suite of tests of L2 speaking to take account of the speaker’s

81
Examining Speaking

move from slow and intentional retrieval at lower levels to a high degree of
automatisation at higher levels, and from individually assembled strings of
words to the production of formulaic sequences. These developments make
an important contribution towards a listener’s impression of fluency.
Fluency is a notoriously slippery concept, and attempts to define it have
caused much controversy over the years. As Luoma (2004:88) points out, the
term can be given a wide range of definitions, from narrow specifications relat-
ing to hesitation and speech rate to broader ones that are ‘virtually synony-
mous with “speaking proficiency’’. Furthermore, it is not a simple question
of identifying features that are physically present in a speaker’s productions;
there is also the issue of how those features are perceived by a listener. With
this in mind, Lennon (1990) suggests that ratings of fluency in speaking
tests differ from scores based upon quantifiable facets such as accuracy and
appropriacy because they draw purely upon performance phenomena.
However, the assertion perhaps misses the point. Judgements of fluency
are heavily influenced by the ease with which a speaker retrieves and assem-
bles word forms. Lennon himself describes fluency as ‘an impression on the
listener’s part that the psycholinguistic processes of speech planning and
speech production are functioning easily and effectively’ (1990:391). Schmidt
(1992) goes further and explicitly links fluency with automaticity. It is cer-
tainly possible to identify a number of surface features in the speech of a
test candidate which provide indicators of proceduralisation and thus of the
speaker’s progress towards more automatic retrieval processes. They include
the chunking of words, the distribution of pausing and the average length of
stretches of uninterrupted speech.
Increased chunking of word strings by L2 learners manifests itself in a
number of ways. Firstly, as already noted, there are likely to be gains in accu-
racy because a chunk is produced in a way that is pre-constituted. There are
also likely to be improvements in aspects of delivery such as rhythm because
the chunks are stored as phonological wholes; this leads to the impression of
a more native-like command of the mechanics of producing the L2.
Chunking enables several words to be produced as a unit, almost as if
they were a single lexical item. Its effects are thus observable in the speaker’s
length of run – often taken to be the mean number of syllables uttered between
pauses. Raupach (1987, cited in Towell and Hawkins 1994:222) compared
French-acquiring German schoolchildren before and after a period of resi-
dence in France and found increases of up to four syllables in their mean
length of run.
One cannot detach consideration of length of run from consideration of
how frequently the speaker pauses. The more pauses there are, the shorter the
runs are and the more fragmented the discourse is likely to appear, whether
in L1 or L2. But the issue here is not just how often a speaker pauses but also
how long and where. All speakers, however fluent, need to pause at syntactic

82
 Cognitive validity

boundaries – often clause boundaries (Levelt 1989:256–60) – for the pur-

poses of planning what comes next. In the case of a less fluent speaker (e.g. an
L2 learner of low proficiency), the process of constructing an utterance will
be slow and effortful in terms of retrieval (locating the phonological form of
a lexical item), of morphological marking (adding the correct inflections) and
of assembly. One can thus anticipate longer pauses at planning boundaries.
But such a speaker may also:
a) plan the forthcoming syntactic unit only partially because of the heavy
cognitive demands which the process entails, and therefore have to
resume planning while actually producing it
b) revise the phonetic plan while in the process of uttering because of limited
linguistic resources or because of the perception that it has misfired.
The results are seen in breaks in connected speech which occur not at syn-
tactic boundaries but within syntactic units, and which might best be termed
hesitations (Cruttenden 1986:37–38) rather than pauses. The more frequently
this kind of interruption occurs, the less likely a speaker is to be rated as
fluent, whether in L1 or L2. Laver (1994:537) comments as follows:

When the speaking turn is broken into individual utterances by the

insertion of silent or filled pauses at the junctures between the phonemic
clauses, it can be regarded as non-continuous but fluent, in that the linguis-
tic material of the phonemic clauses is uninterrupted. When the silent or
filled pause falls internally within a phonemic clause, breaking its coher-
ent intonational structure, then the speaking turn can be regarded as
non-continuous and interrupted or hesitant. If the clause internal pause
is silent, the effect of inserting the pause into the speaking turn will be to
multiply the number of separate utterances making up the speaking turn.

Once again, chunking assists the speaker – here because it simplifies the
planning process and reduces the need to revise plans while speaking. The
results are seen in shorter planning pauses and a much lower incidence of
hesitation.
To summarise, discussion of ‘retrieval’ (i.e. converting the abstract output
of the previous stage into linguistic form) has led us to identify a number of
physical characteristics which are associated with progress in the acquisition
of L2 speaking skills and which together form possible indicators of fluency.
They are:
• use of pre-assembled chunks – leading to syntactic accuracy and native-
like rhythmic properties
• length of run
• duration of planning pauses at syntactic boundaries
• frequency of hesitation pauses.

83
Examining Speaking

For a more extensive specification, see Fulcher’s checklist of phenomena

contributing to dysfluency (1996), which draws upon empirical evidence; see
also work by Hasselgreen (2005).

Phonetic encoding, articulation

Levelt points out (1989:413) that: ‘fluent articulation is probably man’s
most complex motor skill. It involves the co-ordinated use of approximately
100 muscles, such that speech sounds are produced at a rate of about 15
per second.’ In the case of the second language speaker, there are further
complications in the pre-existence of:
a) a set of phonological representations in the mind which serve to define
the phoneme values of L1, plus
b) a set of highly automatic processes, attuned to the articulatory settings
of L1 and the movements which link one to another.
The accommodation of these two elements to the unfamiliar values of the
target language forms the basis of any cognitive account of how L2 articula-
tion is acquired. One should not lose sight of the fact that poor L2 pronuncia-
tion does not result solely from an inability to form the target sounds but also
from an inadequate representation of those sounds in the mind and/or the
inability to communicate the appropriate signals to the articulators.
It is also relevant that the need to focus upon adjusting the articulators to
unfamiliar settings imposes additional cognitive demands upon the speaker.
Such demands can potentially detract from the speaker’s ability to store and
hold in the mind a detailed phonetic plan for the utterance being produced
(Levelt 1989:414–422), the result being short and dislocated productions.
This is particularly to be expected in circumstances where the speaker feels
that ‘correct’ pronunciation is a requirement and that considerable resources
of attention need to be directed towards ensuring that articulation is precise.
There is thus considerable pressure upon test designers to ensure that,
while account is taken of pronunciation in tests of speaking, it does not
feature so prominently that it inhibits the candidates from producing longer
and phonologically cohesive pieces of discourse. Test specifications need to
strike a fine balance, focusing upon intelligibility (Smith and Nelson 1985)
rather than upon narrow considerations of accuracy. Gradation of difficulty
can be positive – reflecting an increasing approximation to L2 segmental
and suprasegmental norms – or it can be negative, representing the gradual
elimination of traces of L1 phonology.
The reliance upon the rather subjective notion of intelligibility might be
questioned in some quarters, not least on the grounds that the criterion typi-
cally entails intelligibility to a native speaker. However, this dependence is
significantly reduced in speaking tests that involve peer interaction (such as

84
 Cognitive validity

the paired Cambridge ESOL Speaking tests), where the intelligibility of the
test taker to another non-native speaker is also put to the test.
A different concern about the assessment of intelligibility focuses on the
unrepresentative nature of the assessor, usually somebody whose own pho-
nological representations are finely honed by dint of long exposure to a wide
variety of L2 accents (see Kenworthy 1987: Chapter 2 for an interesting dis-
cussion of this issue). It can be suggested that this very expertise potentially
diminishes the predictive power of the tests since intelligibility to the assessor
does not guarantee intelligibility to the wider L1 public. However, practical
considerations have to prevail here: the alternative, of using ‘lay’ assessors,
is not a viable solution, given the potential implications for reliability. The
alternative is to raise awareness of the issue among examiners of speaking
so that they are encouraged to imagine themselves in the position of a lis-
tener with less experience of L2 varieties, as is the case with Cambridge ESOL
examiners (see Chapters 4 and 5).

Self-monitoring
In the final stage of the speaking process, a speaker assesses how precisely
and effectively each utterance realises the plans that were laid down during its
assembly. Self-monitoring might compare the rhetorical impact of what was
said against the goals of the speaker at the conceptualisation stage. It might
compare the syntactic structure that was actually produced against the frame
selected during formulation. Or it might compare the realisation of a particu-
lar word against the correct form of the word that is stored in memory. For a
skilled speaker, a major concern is whether the utterance is unambiguous and
whether it conveys clearly to the listener the speaker’s pragmatic intentions.
Self-monitoring thus potentially takes place at many different levels of a
message. However, Levelt (1989:463) concludes that it is extremely unlikely
that an L1 speaker can attend to all the levels in the brief time span available
– especially given that the speaker is also intent on completing the utterance
under delivery and on planning the next one. Certain levels might be priori-
tised (in much the same way that teachers of writing sometimes choose to
focus only on errors of spelling or of sentence construction). Levelt suggests
that the exact levels that are monitored may reflect the demands of the pre-
vailing context and that the degree of monitoring may fluctuate during the
course of an extended utterance.
These comments are illuminating when considering the self-monitoring of
L2 speakers. As noted several times in this account, the effort of assembling
speech in a second language makes additional cognitive demands, which
limit the speaker’s performance compared to that in L1. From this, it seems
reasonable to conclude that second language speakers are even more prone
to limit their self-monitoring to specific target areas. A number of researchers

85
Examining Speaking

(e.g. Lennon 1984, Poulisse 1993, van Hest 1996) have suggested that they
pay more attention to errors of lexical appropriacy than to errors of gram-
matical accuracy, though Kormos (2006:131) remains unconvinced on the
basis of her research with Hungarian learners. It is clearly dangerous to gen-
eralise because factors such as task type and instructional tradition may lead
to variations in the amount of attention allocated to monitoring and in the
importance attributed to accuracy. But it would seem likely that lower pro-
ficiency learners focus attention on linguistic features rather than pragmatic
ones, comparing one or more of their syntax, lexis and pronunciation with
what they perceive to be L2 norms. A mark of increased competence as an L2
speaker would thus be a gradual increment in the extent to which the speaker
heeds the effectiveness with which the message has been conveyed.
A further consideration in both L1 and L2 is the way in which the speaker
handles the recognition that problems of transmission have occurred. A
competent speaker needs to be able to self-repair efficiently and promptly
following certain implicitly recognised norms (Levelt 1989:478–499), which
may or may not be language-specific.

A note on timing
Psycholinguistic accounts of speaking do not always make it clear that the
processes which contribute to planning3 an utterance can occur at three
different points:
a) before the speech event
b) while an interlocutor is speaking
c) immediately before or during the speaker’s turn.
This has consequences for the way one interprets the model of speak-
ing just outlined. It also serves to distinguish two types of speech event: one
where extensive pre-planning is not possible and one where it is.
The ability to plan at point (a) is restricted in most genres of speech event.
An exception can be found in pre-planned monologues such as lectures, where
the speaker has the opportunity not only of choosing the propositions to be
addressed but also of ordering them and making the connections between
them plain. There is even a possibility of encoding the ideas phonologically
and rehearsing the form of words to be used by turning them over in the mind.
It is only in this type of speech event that speaking comes close to the degree of
planning that is possible in skilled writing (Shaw and Weir 2007).
Most genres of speech event are dialogic, and call for immediate responses
to points made by an interlocutor. Planning therefore has to be reactive and
to take place while the event is ongoing. Ample evidence that a degree of
active planning is undertaken while an interlocutor is speaking (point (b))
can be found in the way that a new turn in natural speech often overlaps with

86
 Cognitive validity

the preceding one. But much planning also takes place at point (c), while the
speaker is actually engaged in their own turn. As already noted, it occurs during
short planning pauses, usually located at the end of a syntactic structure such
as a clause. During these pauses, speakers face a complex task. They have to
assemble the next utterance by means of grammatical, phonological and pho-
netic encoding. But they also have to plan ahead conceptually, taking adequate
account of where the present turn is leading or how the listener has reacted. An
early study of planning pauses (Beattie 1983) identified what appears to be a
regular pattern in which a phase of short pauses for linguistic planning gives
way regularly to a phase of longer pauses where a degree of forward concep-
tual planning also takes place. More recent research by Roberts and Kirsner
(2000) appears to confirm the existence of this temporal cycle.

Cognitive processing: Cambridge ESOL practice

We have represented speech production, whether in L1 or L2, in terms of an
adapted and simplified Levelt (1989, 1999) model that falls into six major
phases of processing: conceptualisation, grammatical encoding, phonological
encoding, phonetic encoding, articulation and self-monitoring. This cognitive
model has the potential to provide us with a validation framework for ana-
lysing and describing the operations that take place in tests of speaking, and
will now be applied specifically to a set of Cambridge ESOL tests at different
proficiency levels.
The evidence presented here is drawn from multiple sources, including:
the set of sample speaking tests for KET, PET, FCE, CAE and CPE, which
appear in Appendix A of this book; the Cambridge ESOL Instructions to
Speaking Examiners 2011 (referred to throughout as ISE); the specifica-
tions in the relevant Handbooks for Teachers; the Main Suite Speaking Test
Assessment Scales; and the Cambridge Common Scale for Speaking (see
page 26 in Chapter 1). The review thus attempts to link construct specifica-
tions with the other two points of the test-content triangle in the form of task
specifications and assessment specifications.

Conceptualisation
As already noted, Levelt (1989:107) envisages conceptualisation as entailing
two types of operation:

• macro-planning, in which a set of speech acts is anticipated

• micro-planning at a more local level relating to the role and form of the
upcoming utterance.
In terms of L2 testing criteria, one might consider ‘conceptualisation’
under two main headings:

87
Examining Speaking

a) Provision of ideas – the complexity of the ideas which test takers have to
express and the extent to which the ideas are supplied to them
b) Integrating utterances into a discourse framework – the extent to which
test takers are assessed on their ability to relate utterances to the wider
discourse (including their awareness of information shared with the
interlocutor).
A further factor which plays an important part in assisting conceptualisa-
tion is whether the speaker is given time to pre-plan what to say (in terms of
general ideas, of the links between those ideas or of the actual form of words
to be used) or not. The question of pre-planning is partly a matter of task
design and is therefore discussed later when considering task demands.

Provision of ideas
Retrieving information and generating ideas impose heavy cognitive
demands upon a speaker. If the need for conceptualisation is reduced, then
the task becomes less onerous, allowing more working memory to be allo-
cated to retrieving the relevant linguistic forms. This is one means by which a
designer can adjust test requirements to make allowance for the more effort-
ful processing demands faced by an L2 speaker with limited knowledge and
experience of the target language. Another important consideration in decid-
ing how much support to provide is the need to ensure that a test does not
too heavily reward the candidate’s imagination rather than their language
proficiency.
In the Cambridge ESOL specifications, one can identify two broad deter-
minants of cognitive difficulty. The first lies in the availability of the informa-
tion demanded of the test taker. An emphasis in the early stages upon personal
and everyday information assists test takers because it asks for the retrieval of
information that is conceptually simple and easily accessed. (This is an interest-
ing side benefit of what would appear to be the testers’ main concern: namely,
to grade the difficulty of the language that is needed to achieve the task.)
The task content gradually moves towards more abstract discussion at
CAE and CPE level. The relevant specifications from the Cambridge ESOL
Common Scale for Speaking are as follows:

KET: ‘able to convey basic meaning in very familiar or highly predict-

able situations’
PET: ‘able to handle communication in most familiar situations’
FCE: ‘able to handle communication in familiar situations’
CAE: ‘able to handle communication in most situations’
CPE: ‘able to handle communication in most situations, including
unfamiliar or unexpected ones’

88
 Cognitive validity

These availability criteria are (like some others in the Common Scale and
Can Do statements) not very sharply differentiated. But the way in which
they are operationalised in actual Cambridge ESOL tests is well illustrated
in the interview stage of the sample material at the end of this volume. The
questions in PET Part 1 (p. 315) relate to candidates’ names, occupations
and home towns – familiar terrain indeed, and material that is quite easily
pre-rehearsed. At FCE level, the questioning in Part 1 (p. 318) still features
familiar topics (home, local town, school, jobs) but is considerably more
open-ended. In Part 1 of the CAE sample (p. 321), some standard questions
are retained; but there are also wider topics such as future plans, travel and
holidays and personal tastes and habits. The specifications ‘unfamiliar’ and
‘unexpected’ contribute to a substantial hike in cognitive difficulty in Part 1
of the CPE materials (p. 324) where wide-ranging questions cover housing,
the importance of study routines and sports facilities and elicit personal views
on social change, communications, internet shopping and space tourism.
(For a more detailed discussion of the gradation of topic content across the
proficiency levels, see Chapter 4.)
A second consideration is how much support is provided by the test rubric
in the form of ideas that the candidates might wish to express. An obvious
way of grading difficulty in a suite of speaking tests is to gradually reduce this
support as the level of the exam increases. However, that seems not to have
been the policy adopted in the Speaking tests of the Cambridge ESOL suite:
quite detailed written or visual support for conceptualisation is provided in
all five Speaking tests from KET to CPE.
The benefit of this is that it ensures comparability between the perform-
ances of candidates at a given level since the concepts and the areas of lexis
upon which they draw are similar. Importantly, it also avoids the danger of
weighting assessment at the higher levels too heavily in favour of the test
takers’ imagination rather than their language.
Nevertheless, the support given is carefully calibrated in terms of how spe-
cific it is and how complex the cues are. The cognitive demands are ratcheted
up gradually by moving:
• from set questions for interlocutors to looser and more open ones
• from the precise demands of prompt cards to visual cues that require
description and then on to visual cues which need to be compared and
evaluated
• from a single prompt (usually in visual form) to multiple ones
• from single-modality prompts to multi-modality ones which combine
oral rubrics with both visual and written stimuli.
There is thus a cline which begins with simple written prompts which serve
to constrain the form and content of the productions in KET (in the sample
materials, posters advertise an air museum and a bookshop). It moves on

89
Examining Speaking

to visual prompts of gradually increasing number and complexity; and the

role of spoken rubrics is expanded from FCE onwards (in the sample FCE
materials, candidates are asked not to describe what they see but to bring an
interpretation to bear). Written cues in the form of questions continue to be
used in FCE and CAE to support visuals. For example, in the CAE sample
materials on p. 322, candidates are asked orally why a group of people in
a photo have come together and how important their relationships might
be; but these open-ended questions also appear in written form above the
photos. In CPE, the written support is minimal (in the sample, it consists of
only four words Poster campaign – insurance protection) and the candidate
has to rely heavily on their listening proficiency.
Table 3.1 provides a more detailed outline of the types of information
required of the test taker at each level of the suite, set against the type of
support assisting conceptualisation.

Integrating utterances into a discourse framework

Amongst the factors which Levelt identifies as affecting both macro- and micro-
planning are: awareness of the ongoing topic, thematisation of new infor-
mation, recognition of information shared and not shared with the listener,
accommodation to the point-of-view and even form of words of the interlocu-
tor and certain basic principles which determine how information is ordered.
The Cambridge ESOL specifications take indirect account of factors such
as these in the advice to interlocutors on how to handle highly formulaic and
pre-rehearsed language that does not sufficiently recognise the interlocutor as
a participant in a dialogue (ISE:15, 26, 32, 38). More importantly, the ability to
relate utterances to a wider discourse representation is explicitly taken to be a
determinant of successful performance. A ‘discourse management’ criterion is
included in the assessment scales for all levels except for KET.

Grammatical encoding
As noted earlier, the linguistic content of speaking tests is often specified
not in terms of grammatical structure but in terms of the language functions
which test takers are required to perform. This makes it easier to describe the
transition from the test taker’s initial idea to a rough template for an utter-
ance. It can be treated as a question of mapping from the function that the test
taker wishes to perform to the pattern that best expresses that function.
The issue under discussion when considering cognitive validity is not the
complexity of the language that has to be retrieved (discussed in Chapter 4)
but how easily the test taker is able to perform the mapping exercise. There
are two ways in which the demands of mapping can be reduced in order to
lighten the cognitive load upon lower level test takers with limited linguistic
resources. One lies in restricting the number of functions that a test taker is

90
 Cognitive validity

Table 3.1 Cambridge ESOL suite: information elicited and information

provided

Test Elicited Provided

KET ‘factual personal’ information Oral. Interviewer questions.

everyday non-personal Visual. Prompts with cue words ‘to stimulate

information questions and answers . . . related to daily life’
[See Part 2 of sample KET material, p. 314]
PET ‘factual, personal’ information Oral. Interviewer questions.
[See Part 1 of sample PET materials, p.315]
‘own views and opinions about Visual. Prompts ‘designed to generate ideas’
an imaginary situation’ [See Part 2 of sample PET materials, p. 316]
describing photo Visual. Prompt: one colour photo.
[See Part 3 of sample PET materials, p. 317]
discussion on topic Oral. Prompts by interlocutor.
[See Part 4 of sample PET materials, p. 317]
FCE give ‘basic personal’ Oral. Interviewer questions.
information [See Part 1 of sample FCE materials, p.318]
compare a pair of photos Visual. Photos with a written question.
[See Part 2 of sample FCE materials, p.319]
decision-making task Visual. Photos / pictures with two questions.
(no correct ans.) [See Part 3 of sample FCE materials, p.320]
discussion on topic Oral. Interlocutor led questions.
[See Part 4 of sample FCE materials, p.320]
CAE personal information, opinions Oral. Interviewer questions.
[See Part 1 of sample CAE materials, p.321]
comment on / react to visuals Visual. Photos / pictures plus two written
questions.
[See Part 2 of sample CAE materials, p.322]
decision-making task Visual. Photos / pictures plus two written
questions.
 [See Part 3 of sample CAE materials, p.323]
express, exchange and justify Oral. Interlocutor led questions.
opinions [See Part 4 of sample CAE materials, p.323]
CPE personal information, opinions Oral. Interviewer questions.
[See Part 1 of sample CPE materials, p.324]
react to visuals/decision-making Visual. One or more photos.
task [See Part 2 of sample CPE materials, p.324/5]
impart information; express, Visual. Written question Oral. Interlocutor led
exchange and justify opinions questions. Oral. Interlocutor prompts.
[See Part 3 of sample CPE materials, p.326]

Source: Instructions to Speaking Examiners, 2011

expected to perform (particularly in relation to a single task). The need to

employ a multiplicity of functions is self-evidently more demanding upon the
processor than the need to employ only one or two. The other takes account
of how accessible a given function is likely to be. Functions that are familiar,
frequent and concrete will clearly be mapped more rapidly and reliably than
those which are not.

91
Examining Speaking

Table 3.2 indicates which new functions are added to the repertoire at each
level of the suite. Rather than following the specifications in the Speaking
Test Features, which indicate a wide range of possible functions at each level,
it is based upon a close reading of the interactive task types with a view to
establishing what are the necessary functional demands which they impose.

Table 3.2 Incremental functional requirements of Cambridge Speaking tests

(based on demands of tasks)

Exam Functional demands

KET factual presentation
requesting facts
PET referring to past – present – future
suggesting
agreeing and disagreeing
giving cause – reason – example
FCE comparing and contrasting
giving an opinion
negotiating
speculating
decision making
expressing modality
CAE expressing and justifying opinions
hypothesising
summarisation
CPE evaluation

Clearly, assessors take adequate account of the language resources avail-

able to a test taker at each level for expressing these functions. There is also a
well-designed gradient in terms of the demands imposed upon the test taker:
moving from the presentation of simple facts to the expression of functions
which are increasingly multiple and abstract. Examining a sample of the
visually stimulated tasks (some long-turn, some candidate–candidate) in the
materials at the end of this volume, there is a clear progression:
• KET Part 2: reporting factual information from posters relating to
prices, times, numbers, size etc. on p. 314
• PET Part 3: describing what is happening in a photo on p. 316
• FCE Part 2: speculating about what is happening in a photo; then
expressing a personal viewpoint connected to the photo (in the materials
on p. 319, making a choice and expressing a personal interest)
• CAE Part 3: on p. 323: discussing the difficulties of different types of
fund raising
• CPE Part 2: speculating on why photos might have been taken, then
exploring their usefulness to a poster campaign on p. 324–5.

92
 Cognitive validity

It is noticeable that there is a hike in cognitive difficulty at FCE level: in

particular, it is here that the test taker is required to handle issues of modality
and to give reasons for choices, which can prove both conceptually and lin-
guistically difficult. It represents a considerable advance on what is required
of a candidate at PET level.

Phonological encoding
The discussion of the cognitive framework on pages 80–84 identified a
number of characteristics which are associated with proceduralisation in
the acquisition of L2 speaking skills and which together form indicators of
fluency. They are:
• use of pre-assembled chunks – leading to syntactic accuracy and native-
like rhythmic properties
• length of run
• duration of planning pauses at syntactic boundaries
• frequency of hesitation pauses.
Of these characteristics, the one most consistently represented in the spec-
ifications and instructions to assessors for the Cambridge ESOL Speaking
tests is hesitation. At the lowest level of the suite, an attempt is made to allow
for the greater cognitive demands of planning when one’s linguistic and pho-
nological resources are limited. KET examiners are advised: ‘Candidates at
this level may need some thinking time before they respond. Be sensitive to
this and do not rush candidates, but do not allow pauses to extend unnatu-
rally’ (ISE:15). In the Common Scale for Speaking, there is an attempt to
define hesitation according to the factors that might be responsible for it:

KET: ‘Responses are limited to short phrases or isolated words with

frequent hesitation and pauses.’
PET: ‘Produces utterances which tend to be very short – words or
phrases – with frequent hesitations and pauses.’
FCE: ‘Has problems searching for language resources to express ideas
and concepts resulting in pauses and hesitation.’
CAE: ‘Maintains a flow of language with only natural hesitation result-
ing from considerations of appropriacy or expression.’
CPE: ‘Pronunciation is easily understood and prosodic features are
used effectively; many features, including pausing and hesitation,
are “native-like”.’

The descriptors could perhaps be more specific, with a sharper distinc-

tion made between extended pausing due to planning difficulties, hesitation

93
Examining Speaking

due to delayed planning or restructuring during the utterance, and hesitation

associated with lexical retrieval.
Hesitation as a factor is also mentioned consistently at levels above
KET in the assessment criteria for Discourse Management. Here, inter-
estingly, a recent revision of the criteria has resulted in this indicator of
fluency being closely linked to another – namely, length of run. The panel
below shows the performance descriptors at the middle of the marks range
for each test.

PET: ‘Produces responses which are extended beyond short phrases,

despite hesitation.’ (ISE:48)
FCE: ‘Produces extended stretches of language despite some hesita-
tion.’ (ISE:52)
CAE: ‘Produces extended stretches of language with very little hesita-
tion.’ (ISE:56)
CPE: ‘Produces extended stretches of language with ease and with very
little hesitation.’ (ISE:50)

Chunking does not feature prominently as a criterion. The problem is

perhaps partly that formulaic language can be both a negative and a positive
indicator so far as an assessor is concerned. On the one hand, it can indi-
cate dependence upon a limited range of highly conventionalised formulae,
some of them rote learned; on the other (as previously discussed) it can show
that the test taker has progressed as a speaker to a point where they have
established a repertoire of word strings which can be produced with minimal
effort.
The ESOL criteria include reference to the former, more negative aspect.
At KET level, the Common Scale recognises the likelihood of ‘rehearsed or
formulaic phrases with limited generative capacity’, and at PET, it accepts
that complex language may only occur in ‘well-rehearsed utterances’. At all
levels except (curiously) PET, assessors are given a warning against rote-
prepared utterances: ‘It is expected that some candidates’ responses will
sound rehearsed in Part 1 of the text as talking about oneself is the most
familiar of topics . . . (L)engthy and obviously prepared speeches should be
tactfully intercepted and deflected’ (ISE:15, 26, 32, 38).
However, in their current form the criteria do not make specific mention
of the positive role played by chunking in the development of fluency. Given
the level of disagreement there has been historically about what constitutes
fluency, it is entirely understandable that the criteria avoid reference to the
construct4. The treacherous term does not appear in the Common Scale and
only features, rather anomalously, at one level (CAE) in the Handbook Can
Do statements. Nevertheless, with our developing understanding of the L2

94
 Cognitive validity

speaking process, and in particular the impact upon it of the incremental

storage and retrieval of formulaic items, it might in future be advisable to
focus the attention of assessors more systematically upon observable features
such as degree of chunking, and location of pausing – especially if empiri-
cal research continues to demonstrate correlations with listener judgements
of L2 fluency. As discussion within SLA swings in favour of an emergentist
view of how second language speaking skills are acquired, the importance of
chunking, in particular, will need to be recognised.
A further marker of increasing competence as a speaker lies in the candi-
date’s ability to use the more conventionalised type of chunk (do you know
what I mean? you know, kind of) in order to gain planning time, in the way
that a native speaker might. The insertion of these fillers at natural points
makes speech appear less fragmented and contributes to the impression of
fluency (Hasselgreen 2005:118). Again, this criterion might be a useful addi-
tion to the Common Scale, as well as to the analytic and global scales used by
Oral Examiners.

Phonetic encoding, articulation

In line with the criteria discussed on pages 84–5, the specifications for the
Cambridge ESOL suite strike a fine balance: giving credit for pronunciation
in tests of Speaking, while at the same time not giving this facet such promi-
nence that it inhibits the test taker’s performance or imposes constraints
upon their ability to produce longer and more phonologically cohesive pieces
of discourse.
The focus is appropriately upon intelligibility rather than upon narrow
considerations of ‘accuracy’. The first sentence in each Pronunciation section
of the ISE could not be more explicit: making clear that the term pronun-
ciation in this context ‘refers to the intelligibility of the candidate’s speech’.
The construct of intelligibility is defined in the examiners’ glossary as follows
(ISE:66): ‘The candidate may have a pronounced accent but the speech can
generally be understood by a non-EFL/ESOL specialist.’ This addresses the
reservation expressed earlier about the more sophisticated listening skills of
an EFL professional who has been exposed to a range of varieties.
However, despite the foregrounding of intelligibility as a criterion, the
analytical assessment criteria for raters are disappointingly vague in specify-
ing distinctions between levels. The criteria for KET and PET thoughtfully
balance intelligibility against the candidate’s problems at the articulation
level due to limited control of phonological features (ISE:44). But at the
higher levels, the descriptor says no more than ‘is intelligible’.
The more general pronunciation descriptors at the different levels of the
suite are listed in Table 3.3 (taken from the Common Scale for Speaking on
page 26).

95
Examining Speaking

Table 3.3 Pronunciation specifications for the Cambridge ESOL suite

Exam Common Scale specifications

KET Pronunciation is heavily influenced by L1 features and may at times be
difficult to understand
PET Pronunciation is generally intelligible, but L1 features may put a strain on the
listener
FCE Although pronunciation is easily understood, L1 features may be intrusive
CAE L1 accent may be evident but does not affect the clarity of the message
CPE Pronunciation is easily understood and prosodic features are used effectively

The grading relies chiefly upon a gradual reduction of elements which

reflect the influence of the stored phoneme values of L1 and the automa-
tised articulatory processes that are associated with them. The criteria are
expressed in very general terms. It could be argued that it is difficult to be
more specific because the precise relationship between representation and
production can be expected to vary considerably from individual to individ-
ual and is not transparent to the observer.

Self-monitoring
Like articulation, self-monitoring and self-repair are aspects of speaker per-
formance which are difficult to capture in the form of test specifications. The
closest the Common Scale criteria come is in focusing upon the candidate’s
ability to deal with breakdowns of communication as and when they occur
– specifically, the extent to which the candidate relies upon support from the
interlocutor in addressing such problems.

KET: ‘Requires prompting and assistance by an interlocutor to prevent

communication from breaking down.’
PET: ‘Has some ability to compensate for communication difficulties
using repair strategies but may require prompting and assistance
by an interlocutor.’
FCE: ‘Does not require major assistance or prompting by an
interlocutor.’

It should be noted that the term ‘repair’ as conceptualised here combines

two distinct situations: one relating to a failure of understanding as a listener
and the other relating to a failure of communication as a speaker. Another
reservation is that the criteria on monitoring and repair are limited to the
lower level of the suite. A more direct approach would be to provide spe-
cific reference to the extent to which candidates at different levels show them-
selves capable of self-monitoring and display an ability to redress a subset of

96
 Cognitive validity

their errors that threaten understanding. Higher levels of competence could

then be characterised by evidence of monitoring for pragmatic effectiveness
as well as for linguistic accuracy.

Cognitive task demands

The cognitive demands of speaking tasks
Two important features distinguish speech from writing (Bygate 1987). First
of all, most speech takes place under time pressure. In normal circumstances
(exceptions are rehearsed and ritualised situations and speech events that are
scripted or assisted by prompts), the speaker has to formulate utterances with
a considerable degree of spontaneity. Constructing a spoken sentence clearly
requires an element of planning; but time does not allow for reflection, for a
range of options to be considered or for initial decisions to be revised. It may not
even allow for planning to extend further ahead than a few phrases. Similarly,
time pressures constrain the execution of a sentence in that there has to be a
rapid transition from constructing an utterance to articulating the sounds
which compose it; convention requires a response to be articulated with a
minimum delay (even a pause of 0.5 seconds is recognisably a hesitation). Time
pressures also limit the extent to which a speaker can self-monitor: unlike a
writer, a speaker has little time for checking what has been said for grammatical
accuracy, pragmatic appropriacy and congruence with the speaker’s intentions.
Secondly, most forms of speaking are reciprocal. In such encounters, the
parts played by the participants switch regularly between speaking and listen-
ing. Speakers have to identify with the point of view of their partner in the
exchange, just as writers do; but the role of a listening partner is a dynamic one
in the way that the role of a reader is not. A conversational partner might at
any point exercise the right to change the topic or to disagree with a speaker,
changing the direction of the discourse in a way that is beyond the speaker’s
control. In addition, a response formulated in a conversational context has
to conform to the topic and might even incorporate some of the language of
the most recent turn by the other party – a turn which may still be in progress
while the response to it is being assembled. The speaker is thus required to do
more than simply answer a question relevantly: they need to acknowledge the
presence of the interlocutor by picking up lexis and even syntactic patterns
from the preceding turn. This imposes considerable constraints in a test of
speaking. But it also affords benefits: in tasks requiring only brief responses
(e.g. a question followed by single sentence answer), the interlocutor might
well provide a linguistic model that the respondent can adopt (‘What is the
kind of music that you like best?’ – ‘The kind of music that I like best is . . .’).
The characteristics of spoken interaction just reviewed suggest that the
cognitive demands imposed by a speaking task are not simply a matter of the

97
Examining Speaking

duration of the task or the familiarity of the content and language that has
to be retrieved. The task design determines the amount of time available to
the candidate for the assembly of an utterance. It also specifies the relation-
ship between the examiner and the candidate or between two candidates in
a pair. Here, one consideration is the length of turn required of each party.
The shorter the examiner’s turn, the greater the pressure on the candidate to
formulate and respond. The shorter the turn expected of the candidate, the
more likely they are to be able to construct a grammatically correct response
and/or to base a response upon linguistic forms employed by the interlocu-
tor. An additional factor in an examiner–candidate relationship (and indeed
in a candidate–candidate one) is the predictability of the content. A task
where the topic changes frequently makes much greater demands upon
the speaker-as-listener than one where the line of conversation is relatively
predictable.
Two of the task variables just identified will be considered in the sections
that follow. They are firstly the nature of the interaction that the task requires
and with it the length of turn demanded of the candidate; and secondly the
amount of time available for the assembly of spoken utterances. Each con-
tributes importantly to the cognitive load that is imposed upon a candidate
by a speaking task.

Patterns of interaction
There are several possible interaction formats for a test of speaking (Davies,
Brown, Elder, Hill, Lumley and McNamara 1999:182, Luoma 2004:35–45,
Fulcher 2003:55–57). It can be one-way, with the test taker responding to
a computer screen or to a voice on a CD or DVD in what is usually termed
a SOPI (Simulated Oral Proficiency Interview). This interaction format
is sometimes referred to as indirect and sometimes as semi-direct (Fulcher
2003:190). In this case, the course of the conversation is entirely pre-
determined, and the candidate has no impact upon the direction it takes.
This approach may meet certain practical needs of mass testing (Stansfield
1990) but is difficult to defend in terms of cognitive validity for the following
reasons:

a) The interviewer is limited to certain stereotyped questions and has no

opportunity to respond to the replies of the test taker or to elaborate
where the test taker shows signs of poor comprehension.
b) If the SOPI takes place in audio mode, no facial expressions or
gestures are present to provide the test taker with the normal feedback
indicating comprehension or support. Even in visual mode, these
paralinguistic cues are necessarily depersonalised. Similarly, back-
channelling is impossible.

98
 Cognitive validity

c) The test taker is placed under a time pressure even more extreme than
that which obtains in real life, since a response has to be completed
before the next utterance on the recording.
d) The test taker is limited to a single role, that of respondent.
e) The test taker is unable to demonstrate the ability to seek repair or
clarification in cases of uncertainty.
One can say that, in terms of the real-life cognitive processes engaged, this
type of test might well provide indications as to L2 listening skills but cannot
be said to measure the ability to participate actively in a conversation. As
Fulcher puts it (2003:193) after reviewing the evidence: ‘Given our current
state of knowledge, we can only conclude that, while scores on an indirect
test can be used to predict scores on a direct test, the indirect test is testing
something different from the direct test.’
An alternative use of recording requires the test taker to produce a mono-
logue on a specific question or topic. Again, it cannot be claimed that this
approach replicates the conditions and demands of conversational speech:
there is no interaction and the test taker usually has to be allowed time to
prepare what to say. It does, of course, enable an assessment of the test
taker’s oral presentation skills; but, even so, an important element is absent
in the form of auditors who provide signals of understanding. Furthermore,
the formality of a recording tends to inhibit the test taker from stopping to
rephrase points that they may have made inadequately (Luoma 2004:45).
The part played in normal speaking by retrospective self-monitoring and
repair is likely to be severely reduced. In short, the test does not conform to
the view that most speech events entail a process of co-construction between
interlocutors (McNamara 1997b, Swain 2001).
A second approach is to conduct the speaking test by telephone. Here,
interviewer–test taker interaction is indeed provided for. The interviewer
has the possibility of developing a topic more freely, reacting to initiatives
by the test taker and providing support and back-channelling. However,
this type of test models a single, very specific form of spoken exchange. It
is qualitatively different from a face-to-face engagement in several ways
that directly affect phonological processing. Most obviously, there is the
absence of visual context and of the paralinguistic cues provided by facial
expression and gesture. In processing terms, the consequence is that into-
nation assumes a much more important function for both speaker and
listener. A second drawback is that the physical signal is different since a
phone line employs a reduced frequency band, with the result that acous-
tic cues to certain phonemes are absent because they occur at frequen-
cies above 4000 Hz. An experienced phone user learns to adjust for this
(though the /f/-/s/ distinction in particular remains a problem). However,
the point remains that the low-level acoustic-phonetic processing of a

99
Examining Speaking

phone message differs markedly from that of other types of speech event.
It is unsurprising that many speakers find the process of handling a phone
conversation in a second language, whether as initiator or respondent, a
daunting experience.
An approach which requires the physical presence of an interviewer would
thus seem to be preferable in terms of cognitive validity. However, much also
depends upon the role that the interviewer is required to take. One can trace
a continuum in terms of the demands imposed upon the test taker. At one
end, the interviewer’s questions might elicit responses which are formulaic or
which echo the linguistic patterns employed by the questioner. This is neces-
sary for the purposes of controlling from task difficulty at lower levels; but
it detracts from cognitive validity because relatively closed responses of this
kind impose little requirement upon the speaker to engage in the process of
a novo utterance construction. At the other end, the interviewer’s role might
be limited to back-channelling and follow-up questions once a discussion has
been initiated. This not only places a much greater onus upon the candidate,
but aligns the processing that takes place much more closely with that of
normal spoken discourse.
There have also been sociolinguistic concerns (Ross and Berwick 1992,
Young and Milanovic 1992) that the power imbalance between interviewer
and test taker may affect the nature of the communication that takes place (at
its extreme, raising the prospect of a lockstep question and answer pattern).
These concerns have led test designers to introduce a further interaction
format in which test takers communicate with each other in pairs (Taylor
2000b). The approach has been the subject of some criticism (for a review of
the issues, see Fulcher 2003:186–190, Luoma 2004:36–39), which is discussed
in detail in Chapter 4. Clearly, assessing communication between two non-
native speakers can be said to possess strong ecological validity in light of
the widespread use of English as a language of international communication.
(See, for example, the discussion by Canagarajah 2006). On the other hand,
a cognitive analysis cannot overlook the fact that the format increases the
demands of the test, since the candidate as listener potentially has to deal
with two voices (interviewer and fellow test taker), two varieties of the target
language (one native and one non-native), and even potentially a three-way
interaction.
This last comment draws attention again to the critical role that listening
skills play in what are ostensibly tests of speaking. Clearly, it is impossible to
filter out listening from an interactive communication task, and Chapter 1
noted how the ALTE Can Do statements (Council of Europe 2001) combined
the Listening and Speaking descriptors onto a single scale for Interaction.
Nevertheless, the extent to which achievement in a speaking test is reliant on
the listening skill is an issue in an examination where listening is also tested in
a separate paper. The only way of shifting the balance towards speaking is to

100
 Cognitive validity

provide a further section of the test in which the candidate speaks for a more
extended period – but does so to a live audience in the form of an examiner
and/or another candidate, who can provide the encouragement and feedback
which would normally be available.
This discussion of patterns of interaction in speaking tests has thus iden-
tified three possible and indeed potentially desirable formats: interviewer–
candidate (I–C), candidate–candidate (C–C) and solo candidate (C). A
fourth scenario might allow for a three-way exchange (I–C–C) in which the
interviewer engages in discussion with two test takers.
As noted, these formats vary in the cognitive demands they make of test
takers. At first glance, it might appear that the most demanding is C. The
reasons lie partly in the difficulties of speaking at length, even in one’s first
language, but also (in an L2 situation) in the reduced opportunity for repair
or support by the interviewer. Fulcher comments (2003:19): ‘If we accept the
view that conversation is co-constructed between participants talking in spe-
cific contexts, our construct definition may have to take into account such
aspects of talk as the degree of interlocutor support.’
However, the fact is that each of the four patterns of interaction poses its
own cognitive challenges:
• I–C may well require a relatively rapid response, particularly when the
interviewer’s turns are short.
• C–C contains a major element of unpredictability (particularly in cases
where the partner test taker has listening comprehension difficulties
and is inclined to go off topic). It also contains an important variable
in the form of the test takers’ familiarity with each other’s L2 variety.
A further complication lies in the extent to which a test taker can
accommodate to and echo the language of their partner. Whereas
the language of the interlocutor can presumably be trusted as a
potential source of linguistic information and emulated accordingly,
fine judgements have to be made as to the extent to which a fellow
learner’s language is to be trusted. This kind of decision clearly imposes
additional cognitive demands.
• So far as I–C–C is concerned, a three-way conversation clearly demands
more complex processing than a two-way exchange. The test taker
not only needs to process input in two different voices; but also needs
to keep track of the points-of-view and foregrounded topics being
expressed by two different individuals.
In addition, the cognitive demands made of the test taker may vary
considerably within these various formats. Two important factors briefly
touched upon have been: how much the test taker has to contribute to the
discourse, and the extent to which the test taker has to engage in a novo utter-
ance construction.

101
Examining Speaking

Planning time
The amount of time speakers have in order to prepare what to say has an
important impact upon several of the phases of processing. This is true
whether they are performing in a first or a second language.
Pre-planning time clearly assists conceptualisation. The speaker has
greater opportunity to generate ideas that are relevant to the topic to be dis-
cussed. They have also has greater opportunity to organise them and to mark
how they are linked conceptually. Pre-planning time also assists grammatical
encoding and retrieval: increasing the likelihood of utterances that are care-
fully formed syntactically and of precision in the choice of lexis. One might
also expect a greater degree of fluency in that retrieval of many of the appro-
priate lexical and syntactic forms can take place in advance of task perform-
ance. Indeed, there are opportunities for rehearsing fully formed utterances
and committing them to long-term memory before delivering them. These
utterance templates go on to assist self-monitoring, in that they provide the
speaker with a concrete target against which to match actual performance. In
short, pre-planning time supports not only the search for ideas to express and
the definition of goals, but also the organisation of information, the precision
of the language used and the awareness of performance errors.
That said, it has to be recognised that most speech events are interactive.
They take place under time pressure and require utterances to be assembled
spontaneously. Only a limited number of contexts allow the speaker the
luxury of planning what to say. They include formal monologue situations
such as making speeches or giving academic presentations; but also situations
where a speaker knows in advance that they will be called upon to express an
extended opinion, report an event, tell a story or outline a set of proposals or
requirements. Applying strict ecological criteria, one might argue that it is
only in conjunction with these types of speaking that it is entirely appropri-
ate for a task to incorporate pre-planning time. This is not just an academic
point. It is clear that the cognitive processing associated with a planned task
is markedly different from that associated with an unplanned. One can argue
that it results in a different type of discourse. A listener might reasonably
expect a greater degree of coherence and cohesion, less hesitation and rep-
etition, fewer false starts and reformulations; and this in turn might lead to
setting the bar for fluency at a higher level than with an extempore speaker.
The implications for the assessment of planned monologues will be obvious.
The effects of pre-planning an L2 speaking task have been quite widely
discussed in the literature on task-based learning (TBL), as have the effects of
repeating a task. Most of the studies that have explored the impact upon per-
formance (e.g. Bygate 2001, Crookes 1989, Foster and Skehan 1996, Mehnert
1998, Ortega 1999) concur in concluding that preparation or repetition leads
to an increase in fluency and complexity. Bygate (2001) reports unclear

102
 Cognitive validity

evidence on improvements in accuracy in the case of repetition; but accuracy

must certainly be increased by pre-planning and by the opportunity of men-
tally rehearsing the appropriate forms of words before performing a task.
An important variable in a testing context is the length of time available
for pre-planning a monologue turn. A simple analysis might assume a close
correlation between the time allowed and the well-formedness of the candi-
date’s productions. It is not quite as clear-cut as that, however. There must
be a cut-off point (determined by how much language can be pre-rehearsed
in the special circumstances of a test), after which additional time is unhelp-
ful and may even lead to second thoughts and a blurring of conceptual and
linguistic targets. Of course, across an entire suite of tests, this considera-
tion is counter-balanced by the growing competence of the candidates. As
they acquire greater knowledge of L2 and (above all) greater automaticity
of lexical retrieval and speech assembly, they are able to make increasingly
effective use of any time allowed.

Cognitive task demands: Cambridge ESOL practice

Having considered the way in which task demands can be increased or
diminished by patterns of interaction and by the opportunity to plan, we
now examine the way in which these cognitive variables are represented in
Cambridge ESOL practice.

Interaction
Live interaction, termed direct testing, is very much a feature of the approach
to the assessment of speaking favoured by Cambridge ESOL (see Chapter 1 for
discussion of the long tradition of direct speaking assessment by this board).
The suite of Speaking tests embraces all four of the possible formats that have
been identified. The tests provide for an interviewer (referred to as an interlocu-
tor) whose function is to initiate interaction and to keep it going. But they also
include phases of test taker–test taker interaction, thus mitigating the possible
‘power’ effects previously mentioned as associated with the interviewer’s role.
Most tests also make provision for solo presentation, in which (as noted) lis-
tening skills play a minimal part. To support the interlocutor, who may be per-
sonally engaged in the exchanges that take place, a second examiner, known as
an assessor, monitors performance. Saville and Hargreaves (1999) argue that
the use of a range of interactional formats confirms to general principles of test
design, in which tasks are varied so as to elicit different types of language.
Table 3.4 summarises the speaking tasks at different levels of the suite.
It is clear that a range of interaction types is covered, and that they foster
language processing which draws upon both interactive and presentational
skills. An initial ‘interview’ section in all tests features the I–C relationship,
while C–C appears in the specifications as ‘Two-way collaboration’, C as

103
Examining Speaking

Table 3.4 Interactional patterns in Cambridge ESOL Speaking tests

Exam Task mins Interaction

KET Q&A based on an interlocutor frame 5–6 I–C
Q&A based on prompt material 3–4 C–C
PET Q&A based on an interlocutor frame 2–3 I–C
Imaginary situation with prompt 2–3 C–C
Describe photo 3 C
Conversation on topic of photo 3 C–C
FCE Q&A based on an interlocutor frame 3 I–C
Compare photos; comment 4 C
Decision-making task with prompts 3 C–C
Three-way discussion 4 I–C–C
CAE Q&A based on an interlocutor frame 3 I–C
Long turn based on visual prompt; comment 4 C
Decision-making task and report 3 C–C
Three-way discussion 5 I–C–C
CPE Q&A based on an interlocutor frame 3 I–C
Decision-making task 4 C–C
Long turn based on written prompt; comment 4 C
Discussion on long turns 8 I–C–C

Source: Instructions to Speaking Examiners, 2011.

Note: Timings are for a complete task involving two candidates. Timings for FCE, CAE and
CPE separate Part 3 and Part 4. Actual speaking times for long turns per candidate are: PET:
up to 1 minute; FCE: 1 minute (plus 20 secs. peer feedback); CAE: 1 minute (plus 30 secs.
peer feedback); CPE: 2 minutes (plus up to 1 minute peer feedback).

‘Long turn’ and I–C–C’ as ‘Three-way discussion’. Test takers are required
to respond to task demands involving negotiation as well as to more formal
questions, both open and closed.
In grading the cognitive demands imposed upon candidates, the test devel-
opers have relied partly upon the duration of the tasks. The overall contact
time for a pair of candidates increases gradually from KET (maximum 10
minutes) to CPE (19 minutes). These figures are only broadly indicative
of the amount of speaking an individual candidate has to do, since there is
inevitably variation in the relative contributions made. A more important
criterion appears to be the way in which the test is distributed between the
various types of interaction. Figure 3.3 provides evidence of how this feature
has been calibrated.
At KET and PET, the Part 1 I–C interview depends quite heavily upon
simple routinised Q&A exchanges. Here are some examples from KET part 1:
Where do you live / come from?
Do you work or are you a student?
Do you like (studying English)?

104
 Cognitive validity

Figure 3.3 Interaction formats in Speaking tests: timings for tasks for two
candidates

Note: the PET figure combines two C–C interactions

9
I–C
8 C–C
C
7
I–C–C
6
Minutes

0
KET PET FCE CAE CPE
Examination

The questions are relatively predictable, allowing candidates to rehearse

their answers. The answers are simple and consist of one or two sentences.
However, the cognitive demands imposed by the I–C format increase
markedly in FCE Part 1. The sample materials on p. 318 include questions
which are more open-ended and more difficult to predict. The questions
can even invite the candidate to consider hypothetical constructs (see FCE
Handbook for Teachers, p.83):
Could you tell us something about your family?
How much time do you usually spend at home?
Is there anything you would like to change about the area where you grew up?
Questions of this type demand quite extended responses from the
candidate; and there is clearly a stronger requirement to engage with the
interlocutor rather than simply participating in a Q&A routine.
As for grading by format, the complex three-way interaction does not
appear until FCE (4 minutes) but then reappears at CAE (max. 5 mins) and
CPE (max. 8 mins). The length of the ‘long turn’ when the candidate per-
forms solo increases from PET (1 minute but a shorter period tolerated) to
CPE (2 mins). The test designers thus treat I–C as the least demanding of the
formats (especially when it is based on a simple Q&A exchange), and reserve
C and I–C–C in particular for higher levels. At the same time, they attempt
to redress the possible sociolinguistic constraints upon interlocutor–test

105
Examining Speaking

taker communication by providing for a C-C phase even at the lowest level
in KET – though obviously here in a very controlled form, with interlocutor
prompting.
Overall, then, there has clearly been a principled approach to the grading
of the task demands arising from different interaction types.

Planning time
As noted earlier, all the tests in the ESOL suite, with the exception of KET,
include one task which requires extended individual performance. In cogni-
tive terms, a long turn of this kind places heavier conceptualisation demands
upon the test taker, who has to generate more ideas than when responding
briefly to the comments of others and has to organise them meaningfully. It
might be said to constitute a different type of discourse by virtue of the part
played by forward planning (Brown, Anderson, Shillcock and Yule 1984:16–
18). From the assessor’s point-of-view, coherence and cohesion would then
be expected to be a greater consideration and hesitation, self-repair and
loosely connected ideas might be regarded unfavourably.
The Instructions to Speaking Examiners (2011) make it clear that certain
assessment criteria are more applicable to long turn tasks that are principally
monologues. Particular attention is given to discourse features. From PET
level up to CPE, these are:
• sustaining a long turn
• coherence and clarity of message
• organisation of language and ideas
• accuracy and appropriacy of linguistic resources.
(ISE:47, 51, 55, 59)
Due account has thus been taken of the part played by ‘macro-planning’
in extended discourse. However, there is an anomaly here. The candidate is
allowed little or no pre-planning time in order to reflect upon the prompts
provided and to undertake the kind of forward thinking that would feature in
some types of longer turn. To be sure, the turn remains a relatively short one at
most levels (indeed, the PET instructions allow the candidate to speak for less
than a minute). But if candidates are to be judged on criteria such as coher-
ence, cohesion and organisation, then the longer 2-minute slot at CPE level
perhaps requires a brief period of preparation to permit macro-planning. It is
true, as noted earlier, that a degree of forward planning can occur alongside
speech assembly, once a monologue is underway. But one wonders whether
in this type of task a speaker needs time to formulate in advance a set of ‘sub-
goals’ or ‘speech act intentions’ (Levelt 1989:109). The requirement that test
takers speak extempore but are assessed in part by criteria based upon pre-
planned speech would seem to break one of Weir’s test performance condi-
tions (1993:39), namely ‘processing under normal time constraints’.

106
 Cognitive validity

However, the case is not as clear-cut as this line of argument might

suggest. It would seem from the specifications5 that the perceived function
of the monologue tasks is not to test the candidate’s ability to undertake a
formal lecture-style presentation of the kind that would require pre-planning
in a real-life context. Instead, it is to assess the extent to which the candi-
date is capable of engaging in a longer turn of the kind that might occur if
they were called upon impromptu to tell a story, argue a point or explain a
procedure. On these grounds, one might argue that cognitive validity is not
compromised, though there should perhaps be some degree of latitude in
applying the discourse-level criteria6.

Summary: Cognitive processing across

Cambridge ESOL levels
The main purpose of the present review was to match the specifications of
the Cambridge ESOL suite of exams against an external model of the cogni-
tive processes which the speaking skill requires of a native user. The model
chosen was the most comprehensive one available: that of Levelt (1989). The
Levelt model incorporates a number of components which were mentioned
in the preamble as essential: a mechanism for forward-planning, a means of
storing plans for forthcoming utterances while they are being articulated and
a system for monitoring one’s own productions to see if they accord with
one’s intentions.
The cognitive validity of the Cambridge ESOL suite was examined with
reference to the stages of processing identified by Levelt (1999).
• Conceptualisation. The ESOL suite was found to provide detailed
task input to the test taker in order to reduce the demands of
conceptualisation. The effect is to lighten the cognitive load upon
test takers, and also to reduce any potential bias towards rewarding
imagination rather than linguistic performance.
• Grammatical encoding. The Cambridge ESOL suite specifies linguistic
content in the form of language functions to be performed by test
takers. It thus operationalises this phase of the Levelt model as a
mapping process between the target functions and the syntactic forms
which correspond to them. Two possible principles were identified for
the grading of the functions in terms of cognitive demands: the first
related to the semantic complexity of the function to be expressed, the
second to the number of functions elicited by a particular task. The
Cambridge ESOL suite was found to have borne both principles in mind
in a systematic way when staging the difficulty of its Speaking tests.
A marked increase in cognitive demands was noted at the FCE level,
which may be intentional.

107
Examining Speaking

• Phonological encoding. There is growing support for a view of second

language acquisition as a process of ‘proceduralisation’, in which the
retrieval of linguistic forms begins as a slow and attention-demanding
process but becomes increasingly automatic. Because this is a gradual
and internalised development, it is difficult to represent reliably within
a set of test specifications; however, a major driving force behind it
is the chunking of words into formulaic strings – affecting both the
way in which they are stored in the mind and the ease with which they
are retrieved. Helpfully for the test designer, chunking is associated
with certain observable developments in an individual’s productions,
which give rise to the impression of increased fluency. They include:
a reduction in planning time at clause boundaries, a reduction in
hesitation, an increase in length of run, an increase in grammatical and
collocational accuracy within the chunk and a progression towards
a more native-like rhythm. Of these, hesitation and pausing are
represented at all levels of the Cambridge ESOL specifications, with
clearly marked gradations between the levels. The use of formulaic
language features only negatively, as an indicator of inability on the
part of low-level test takers to generate novel utterances. The proposal
was made that future specifications might incorporate evidence of
chunking as a marker of increasing proceduralisation as the test taker
moves up the scale. Test designers may also need at some point to adjust
their thinking in the light of increasing support for emergentist views
of language acquisition. It is not clear at present how this development
might impact upon test specifications.
• Phonetic encoding, articulation. Potential problems of articulation were
represented as deriving from two sources: inadequate phonological
representations in the mind and the inability to adjust to unfamiliar
articulatory settings. The L2 speaker may also face a tension between
the need to hold a phonetic plan in the mind and the need to focus
attention upon precise articulation. In these circumstances, it is
important that test designers do not unduly emphasise the importance
of accuracy in pronunciation. The Cambridge ESOL suite deals with
this issue sensitively by adopting intelligibility as its principal criterion.
• Self-monitoring. A competent speaker monitors their own productions
for accuracy and appropriacy; and is capable of introducing self-repairs
both promptly and following certain norms. This aspect of the speaking
process is represented in the ESOL specifications in terms of the level
of support needed from the interlocutor and the test taker’s ability to
achieve repair. Two types of repair are combined in the descriptors, and
it might be advisable to define the term more precisely. It might also
be possible to grade self-monitoring more specifically by reference to
the levels of attention which test takers find themselves able to allocate

108
 Cognitive validity

to, respectively, the linguistic and semantic-pragmatic features of the

utterances.
This validation exercise also considered the types of interaction which
are possible in a speaking test, and the extent to which they can be said to
conduce to the cognitive processes that might apply in non-test conditions.
The formats employed by the Cambridge ESOL suite were found to possess
greater cognitive validity than possible alternatives – providing a clear cost–
benefit justification for the rather complex practical procedures which they
entail. It was noted that the suite features a range of different interaction
types; and thus attempts to represent the variety of speaker–listener relation-
ships which occur in real-life speech events, and the processes that each type
requires of the speaker.
A particular strength of the suite lies in the various ingenious ways in
which it fosters test taker–test taker interaction but continues to control the
range of language that is used. Clearly, it is not possible in test conditions
to create a context identical to that of a natural speech event. However, the
absence of a power relationship between test takers plus the problem-driven
form of the tasks provided in this type of exchange ensure that the encoun-
ters which take place elicit processes which are as close as one can perhaps
achieve to those of real life.
A second issue relating to the relative cognitive difficulty of the tasks was
the amount of pre-planning time permitted. It was noted that pre-planning is
not provided for – a decision which (in terms of cognitive validity) is entirely
sound in the case of tasks that are designed to measure spontaneous spoken
interaction. It is more open to question in relation to the monologue tasks,
in that the absence of pre-planning time means that they do not replicate
the cognitive processes which often accompany the preparation of a formal
presentation. However, a great deal depends upon the perceived purpose
of those tasks. The Cambridge ESOL test designers might argue that they
are intended to provide an indication of a candidate’s ability to produce an
extended turn, not of their ability to engage in a markedly different type of
speech assembly.
It is clear that, in grading the specifications for the five levels of the suite,
designers have given careful thought to the relative cognitive difficulty both
of the tasks and of the interaction formats. Task demands are increased only
gradually; and the more demanding types of interaction (particularly three-
way discussion) are reserved for higher levels of the suite. Overall one con-
cludes that the Cambridge ESOL specifications do indeed correspond closely
to what we know of the cognitive processes involved in the production of
speech. The few reservations that have been expressed represent omissions;
none constitute provisions that run counter to the findings of speech science.
It is also apparent that the cognitive requirements have been sufficiently
finely graded in relation to the different levels of the suite. Full consideration

109
Examining Speaking

has been given both to task demands and to the types of processing that
can be deemed to be representative of performance at different stages of
proficiency.

The focus of the present chapter has chiefly been on the cognitive demands of
the speaking process, whether in L1 or in L2. The criteria for assessing cogni-
tive validity that have been identified concern the target behaviour of the test
candidate and potentially provide a framework for the cognitive validation
of any test of second language speaking.
A secondary focus has been on two types of task variable present in tests
of speaking: namely the nature of the interaction involved and the availabil-
ity of pre-planning time. Here, the emphasis has been strictly upon the way
in which these variables increase or diminish the cognitive demands upon
a speaker. Clearly, other aspects of task design (including task setting and
linguistic demands) are beyond the remit of this chapter, and will now be
considered in Chapter 4.

Acknowledgements
The author is enormously indebted to Lynda Taylor for her sensitive editing
and her helpful suggestions and feedback. He is also especially grateful to
Kathy Gude, Chair of the Cambridge ESOL Speaking Panel, for percep-
tive comments and advice about the criteria and the instructions given to
examiners.

Notes
1 The examples are from Aitchison (2008:254–5); for a discussion of syntactic
speech errors, see Fromkin 1988.
2 It is curious that a procedural component of this kind does not feature
in some standard models of test performance. The Bachman and Palmer
(1996) model, for example, provides for task-related metacognitive strategies
used by test takers; but does not explicitly recognise the very different type
of highly learned cognitive procedure that enables a candidate to apply
linguistic and pragmatic knowledge with minimal working memory demands.
3 Some caution is needed with the term planning, which is used in a narrow
sense when discussing conceptualisation (macro- vs micro planning) but is also
used more generally to refer to the process of assembling an upcoming piece
of speech.
4 Chair of the Cambridge ESOL Speaking Panel, Kathy Gude, comments
(personal communication) ‘Perhaps it is [because] our assessment of ‘fluency’
is too subjective and not as easily quantified as some of the other assessment
criteria’.
5 Describing the long turn section, the CAE Handbook for Teachers
(2008b:76) asserts: This part tests the candidates’ ability to produce an
extended piece of discourse. On the other hand, the profile also foregrounds

110
 Cognitive validity

the kind of criteria that might be associated with a pre-planned presentation:

Candidates have the opportunity to show their ability to organise their thoughts
and ideas. . .
6 Interestingly, a minute is allowed for selecting and preparing the topic ahead
of the long turn in the Speaking papers of the specialised Cambridge tests:
Business English Certificates (BEC), International Certificate in Financial
English (ICFE) and International Legal English Certificate (ILEC). Indeed,
in BEC, candidates are provided with pen and paper to assist them. Here
one can reasonably claim that cognitive validity is not threatened since a pre-
planning component is characteristic of the types of more formal real-world
presentation that occur in business, financial and legal contexts.

111

View publication stats