Language Technologies for Lifelong
Learning
LTfLL -2008-212578
Project Deliverable Report
D4.1 – Positioning Design
Work Package
4
Task
1, 2
Date of delivery
Contractual: 01-10-2009
Code name
D4.1
Type of deliverable
Report
Security
(distribution level)
Public
Contributors
Authors (Partner)
Gaston Burek, Dale Gerdemann, Adriana
Berlanga, Els Boshuizen, Isobel Braidman,
Alisdair Smithies, Fridolin Wild, Petya
Osenova, Kiril Simov, Gillian Armitt, Stefan
Trausan-Matu
UTU, OUNL, UNIMAN, IPP-BAS, PUB-NCIT
Contact Person
Dale Gerdemann
WP/Task responsible
UTU
EC Project Officer
Ms. M. Csap
Abstract
(for dissemination)
This report explains the issues of positioning learners in a
knowledge domain, to recommend learning materials to follow,
and diagnosing learners’ conceptual development, to provide
formative feedback and recommend remedial actions. It also
discusses how Language Technologies can be used to perform
these two tasks in a (semi) automatic way. To this end, the report
presents an outline of the existing language technologies, and
resources for the analysis of learner texts (task 4.1), and an
overview of the process of acquiring and analysing data for
measuring conceptual development (task 4.2).
Actual: 05-10-2009
Version: 2
Draft
Final
Furthermore, it presents a description of an initial set of
experiments and test results, a plan for extending existing tools
for the positioning task, as well as a description service that will
be validated in the next cycle of the project.
Keywords List
Learner Positioning, Diagnosing Conceptual
Development , ePortfolio, Latent Semantic
Analysis, Knowledge Rich Approaches,
Language Technologies, LTfLL
LTfLL Project Coordination at: Open University of the Netherlands
Valkenburgerweg 177, 6419 AT Heerlen, The Netherlands
Tel: +31 45 576291 – Fax: +31 45 5762800
D4.1 Positioning Design
Table of Contents
Executive summary ................................................................................................................... 3
Work package 4.1: Positioning ............................................................................................. 3
Work package 4.2: Conceptual development ........................................................................ 4
1. Introduction .......................................................................................................................... 6
1.1 Purpose of document....................................................................................................... 6
1.2 Project goals ................................................................................................................... 6
1.3 Tasks of work package 4 and their relation to other work packages .............................. 6
1.4 Positioning and language technologies (WP4.1) ............................................................ 7
1.5 Conceptual development of the learner (WP4.2) ........................................................... 8
1.6 Relationship between WP4.1 and WP4.2 ....................................................................... 9
1.7 Overview of this report ................................................................................................ 9
2. Theoretical background .................................................................................................... 11
2.1 Introduction................................................................................................................... 11
2.2 Theoretical basis of learning ........................................................................................ 11
2.3 Knowledge creation theories ........................................................................................ 12
2.4 Theories which model the gaining of expertise and their relevance to professional
development ........................................................................................................................ 16
2.5 Knowledge restructuring .............................................................................................. 17
2.6 Professional learning .................................................................................................... 18
2.7 Learning processes ....................................................................................................... 19
3. Educational domains included in WP4 ............................................................................. 20
3.1 Medicine as a model ..................................................................................................... 20
3.2 Pedagogic domain: Reflective and critical learning in medicine – role of portfolio . . .20
3.3 The role of ePortfolios in Computer Science education ...............................................22
4. Texts available for analysis ............................................................................................... 23
4.1 Introduction................................................................................................................... 23
4.2 Text material generated in Manchester University medical curriculum ...................... 23
5. Positioning the learner (Task 4.1) ......................................................................................25
5.1 Introduction................................................................................................................... 25
5.2 Knowledge poor based positioning .............................................................................. 25
5.3 Knowledge rich approaches for positioning the learner ............................................... 33
6. Diagnosing conceptual development (task 4.2) ................................................................. 42
6.1 Outputs of WP4.2 ....................................................................................................... 42
6.2 Research problems........................................................................................................ 42
6.3 Introduction to studies .................................................................................................. 43
6.4 Comparison of existing concept mapping tools - which tools to use?.........................44
6.5 Studies on potential reference models using Leximancer and Pathfinder ....................49
6.6 Conclusions and next steps ........................................................................................ 51
6.7 Design of the WP4.2 service ........................................................................................ 52
7. Conclusions ........................................................................................................................ 54
7.1 WP4.1 ........................................................................................................................... 54
7.2 WP4.2 ........................................................................................................................... 54
LTfLL -2008-212578
1
D4.1 Positioning Design
References................................................................................................................................ 56
LTfLL -2008-212578
2
D4.1 Positioning Design
Executive summary
LTfLL work package 4 focuses on two independent but connected issues:
•
•
Determining the learner's position with regard to learning materials to provide the learner
with the 'best' suitable material to achieve their learning goals (WP4.1).
Determining the conceptual development of a learner related to a particular expertise
area, to provide them with formative feedback (WP4.2).
Work package 4.1: Positioning
The WP4.1 scenario is associated with the 'building collaborative knowing' part of the Stahl
cycle that integrates the LTfLL project. Brown and Duguid (2001) argue that communities of
practice develop knowledge and share that knowledge within the community's participants
according to local communication patterns. To develop the positioning system, we will
identify natural language expressions characterising the use of language within specific
communities of practice, using knowledge poor and knowledge rich approaches e.g. LSA,
ontology supported sentiment analysis. Our work in the first phase of the project is
described below.
Knowledge poor approach:
Method: The knowledge poor approach restricts itself to analysing only learners' texts and
modelling only experts' texts. We use techniques of text categorization e.g. LSA with a
traditional bag-of-words model and a novel bag-of-phrases model, where the phrases are
extracted using suffix arrays as in Yamamoto and Church (2001). Phrases are weighted
according to their probability of occurring predominately in high quality expert texts
(representative of an expert community of practice) or low quality non-expert texts.
Results: For comparing and finding prototypical expert texts, phrase-based LSA results
generally provide improvement over the traditional bags-of-words LSA results. The phrase
weighting approach has been successfully used to extract synonymous pairs (e.g. "drug
charts" vs "prescription charts") differing only in standards of usage in communities of
practice.
Conclusions and future work: The results so far have been positive for the data we have
used. Therefore, we would like to generalise by using texts generated within different
communities of practice. As the new texts will present different linguistic features, we plan to
test alternative configurations (higher degree of distinctiveness, etc) for the phrase weighting
and extraction. Using larger training data sets will allow us to afford the analysis of more
distinctive phrases without facing unmanageable levels of sparseness. More distinctive
phrases will allow an improvement in the characterisation of language usage, which will help
both with the qualitative and quantitative feedback to the user.
Knowledge-rich approach:
According to Wenger (2001), communities of practice produce and share knowledge artifacts
that need effective management. Within our knowledge rich approach, the management of
those artifacts is achieved collaboratively and requires reference models for comparison e.g.
ontologies. Our work in the first phase of the project is as follows:
LTfLL -2008-212578
3
D4.1 Positioning Design
Method: Knowledge rich methods rely on analysis of the text by using knowledge sources
outside the text (e.g. linguistic and domain ontologies, lexicon, dictionaries, grammars, etc.)
for reasoning about the semantics (e.g. similarity in text meaning) and supporting sentiment
analysis (Moilanen and Pulman, 2007; Liu, 2008). An annotation grammar is developed to
mark concepts in the text. Ambiguities are resolved by means of discourse segmentation and
lexical chain analysis, and the author's attitude toward the concept is determined through
sentiment analysis. The information obtained through this processing chain is included in the
vector space model of the knowledge poor approach in proper balance to obtain optimal text
classification used for positioning.
Results: The CLaRK system (Simov et al., 2001) has been adapted to accommodate the
processing chain including an annotation grammar, discourse segmentation, lexical chain
analysis and sentiment analysis. As the process of developing the lexical semantic resources
is data oriented and we need to have already available the data sets to be used in the
validation available to be able to built such resources. We have already completed the
analysis of data requirements for validation of WP 4.1 scenarios and the data set is being
built.
Conclusions and further work: We will evaluate our knowledge rich approach for
positioning by creating manually a gold standard corpus and then test our approach by means
of precision and recall metrics. In addition we will compare this approach with the LSAbased approach with the aim to find an optimal combination of both in order to satisfy task
goals.
Work package 4.2: Conceptual development
In WP4.2 we build on the work of Stahl (2006) to provide a tool to support the development
of the individual learner, providing a component of an individual's reflective learning cycle
and corresponding to the 'building personal knowing' in the Stahl cycle.
Method: In the first phase, we compared the outputs from a number of concept mapping
tools providing a means of determining how learners relate basic concepts, to establish which
could meet the requirements identified in the WP4.2 scenario. We investigated the utility of
reference models against which learner texts can be compared, as a basis for feedback.
Results: Although Leximancer and Pathfinder were selected for initial experiments, their
functionality and flexibility was insufficient for the requirements of the project. We will
therefore develop a custom tool based on LSA as a basis for the WP4.2 service. Using these
concept mapping tools, a clear distinction could be made between the ability to integrate
concepts demonstrated by individual learners with that of reference material. Two types of
reference models were investigated: a "pre-defined reference model" based on materials from
the curriculum and an "emerging reference model", drawn from the concepts and inter
relations between them generated by a peer learning group. The pre-defined reference model
was too complex for comparison with a novice learner and may not be suited to early stages
in a curriculum. The emerging reference model was a better indicator of the appropriate level
of abstraction and relationship between concepts attainable by individual learners.
Conclusions and further work: We conclude that the emerging reference model, based on
the relationships between concepts, generated by peer groups of learners, may be more useful
to the novice learner as it approximates to his/her Zone of Proximal Development. Future
studies will investigate the effectiveness of comparisons of the pre-defined and emerging
LTfLL -2008-212578
4
D4.1 Positioning Design
reference models for different groups of learners. Stakeholder feedback will inform the
development of the custom LSA-based service to provide formative feedback.
LTfLL -2008-212578
5
D4.1 Positioning Design
1. Introduction
1.1 Purpose of document
In this report we claim that learner positioning and diagnosis of conceptual development can
be addressed by applying the latest advances in the research domain of Natural Language
Processing (NLP), and particularly Latent Semantic Analysis (LSA). To this end, this report
attempts to answer the following questions:
•
•
•
•
Which educational theories are relevant to develop a solution to tackle these problems?
How will these theories relate to the proposed solutions for these problems?
What are the solutions that will be developed?
What technologies and methods will be used or developed to implement the solutions?
1.2 Project goals
The Language Technologies for Lifelong Learning (LTfLL) project is concerned with adult
learning in the context of the new language processing technologies and of the collaboration
facilities offered by the Web2.0. Adult learners usually bring with them significant prior
experience and will show a degree of autonomy in their learning (Knowles, 1975). Although
this was emphasised in adult learning theory, subsequent research has shown that it is
relatively inefficient and ineffective for adult learners to be wholly self-directed (Dornan and
David, 2000). Their motivation is increased by external sources of support such as feedback,
peer comparison, and mentoring (Sargeant et al., 2006). The success of the new paradigm of
Computer-Supported Collaborative Learning (CSCL, see Stahl, 2006), as exemplified by
Web2.0 tools and social-cultural learning theories (Vygotsky, 1978), adds a collaborative,
social dimension to classical, autonomous learning.
LTfLL focuses on adult learning taking place in the work place, in vocational studies and in
Higher Education programs. The aim of LTfLL is to create services that enhance individual
and collaborative development of competence in educational and organizational settings. The
intention is to use language technologies extensively, so that adult learners can be supported
effectively and efficiently.
Learner support can place a heavy load on staff time and resources. Stakeholder analysis has
identified four types of activity that are responsible for this burden: assessment of student
contributions, answering students’ questions, community and group support and monitoring
and assessing the progress of students’ studies (Van Rosmalen et al., 2008). The development
of services in work package 4 addresses these issues.
1.3 Tasks of work package 4 and their relation to other work packages
Work package 4 (WP4) is dedicated to developing the means of positioning the learner,
through monitoring and assessing study progress. Its particular aim is to support lifelong
learners as individuals:
LTfLL -2008-212578
6
D4.1 Positioning Design
•
•
by assessing what they know in order to recommend appropriate subsequent learning
materials (task WP4.1)
by helping learners recognize their understanding of a particular topic, so that they can
develop further as independent learners (task WP4.2).
Work package 4 uses the scenarios developed in WP3 in association with validation activities
(WP7) to guide the design of its services. Working with WP2, the services will be embedded
in the LTfLL Personal Learning Environment and where appropriate, will interoperate with
services in other work packages. The outputs from WP4 will comprise positioning services
based on knowledge-poor and knowledge-rich approaches, in association with user interfaces
delivering the required functionality to learners, tutors and other stakeholders. As well as
establishing how language technologies can best be used to provide the underpinning
positioning and conceptual development services, an important challenge (given the
complexity of data output from the underlying services) will be to establish effective ways to
deliver useful information to end users.
As stated in the LTfLL Description of Work, Medicine and Information Technology have
been chosen to explore the ideas and services to be developed. Both domains include learning
in formal and informal settings and learning, which may lead to certification. In WP4, the
undergraduate Medicine programme provides a good model for self-directed learning in the
work place as well as for more traditional learning in Higher Education. Medicine is also of
interest for its multi-disciplinary nature. Practitioners are expected to adopt a holistic
approach integrating an understanding of disease processes with communications
skills, psychology, sociology and the ethical implications of healthcare, underpinned by an
understanding of how to learn. In contrast, our commercial partner BIT-MEDIA provides
short courses in Information Technology to unemployed adults, where the emphasis is on
assessment of knowledge and skills.
WP4 is concerned with textual evidence of learning, which is an important medium for
communicating knowledge in education and through electronic media. Although some
learners use other media to represent their learning, older and more recent socio-cultural
perspectives see language as a key mediator of learning (Vygotsky, 1978, Wertsch, 1991).
According to socio-cultural conceptualizations, construction of knowledge through
dialectical, dialogical and social processes results in a large reservoir of tacit knowledge (See
Figure 1). In addition, the social cultural interactions have effect not only in shared, tacit
knowledge but also in linguistic patterns of usage (speech genres; Bakhtin, 1986). Analysis
of text is therefore a valid means of achieving the aims of this work package.
1.4 Positioning and language technologies (WP4.1)
The purpose of WP4.1 is to provide the underpinning technology for the project to support
positioning using language technologies. WP4.1 has worked in collaboration with WP3 to
provide a real life scenario, which will be implemented and will demonstrate the use of
advanced positioning technologies in real life situations.
Learners, coming from a variety of backgrounds, have different learning goals and different
prior knowledge. Learners' knowledge gained from previous experiences of learning can be
of a formal nature (certified exams, certificates, etc.), in which case standard
LTfLL -2008-212578
7
D4.1 Positioning Design
admission/exemption procedures may apply, or non-formal learning where such standard
procedures are not available. Positioning in such contexts refers to the identification of a
learner's existing knowledge and to the comparison with knowledge existing within relevant
Communities of Practice (Wenger, 1998). In order to provide relevant text-based knowledge
resources to learners, enabling them to best direct their learning efforts, we will explore
whether language technologies can be used to recommend to learners the most appropriate
text-based knowledge resources in relation to their current position.
A key research problem in order to operationalise the scenario is to establish the best ways in
which to optimise the use of language technologies to achieve meaningful information for the
stakeholders, e.g. learner, tutor. We will study samples of real life texts in the medical and IT
domains in order to establish answers to the following questions:
•
•
To what extent can Language Technologies provide a mechanism to analyse and compare
evidence of previously acquired knowledge that identifies the learner's position in a given
domain?
To what extent can these language use patterns be used as formative feedback for the
learners?
This analysis does not presuppose that language technologies can be used to examine a text
and directly determining from that text what the author knows and does not know. As is
discussed in Section 5, the determination of learning knowledge is indirect in the sense that it
uses evidence based on phrases, terminology and general language use patterns.
This analysis does not presuppose that language technologies can be used as an alternative to
tutor advise but rather as a support to it. That is of particular interest when these patterns are
unconsciously used and would not otherwise be noticed by the tutor without the help of
language technologies.
Our proposed solution is to start with a knowledge poor approach using LSA combined with
a novel bag-phrase approach to compare learner texts with expert texts. Learners will receive
quantitative feedback indicating the distance between these texts. In addtion, learners will
receive qualitative feedback indicating the fit of their language usage in relation to language
used within relevant communities of practice. As second step, we will incorporate a
knowledge rich based analysis to provide similar feedback.
1.5 Conceptual development of the learner (WP4.2)
The purpose of WP4.2 is to provide the underpinning technology for determining the
conceptual development of the learner as the basis for providing feedback to individuals and
facilitators, e.g. tutors. WP4.2 has worked in collaboration with WP3 to provide a real life
scenario in the Medicine domain. The scenario will form the basis for a conceptual
development service which will be tested in real life situations in order to provide iterative
improvements to the conceptual development engine and associated functionality, e.g.
visualisations and reports.
Formative feedback aims to communicate information that will engender accurate, targeted
conceptualizations of a particular topic for the purpose of improving learners’ understanding
LTfLL -2008-212578
8
D4.1 Positioning Design
of it (Shute, 2008). Learners, therefore, need to monitor their progress in their understanding
of a specific area or problem, so that they recognize the limitations of their current level of
expertise. Both these limitations and the degree of progress made will determine the level and
type of formative feedback they need. Monitoring and recognizing a learner’s level of
expertise has to take into account their knowledge, the level of cognitive processing required
by the task, and the associated instructional strategy used (Ertmer & Newby, 1993; Jonassen
et al., 1993b).
The key research problems are to establish the best ways to determine conceptual
development and the most effective means of communicating the results to the stakeholder,
e.g. learner, tutor. We will work closely with stakeholder groups to determine the answer to
the second problem, through validation of WP4.2, in collaboration with WP7. Our overall
questions are:
•
•
•
To what extent can Language Technologies provide a basis of formative feedback service
for individual learners, which takes into account their conceptualization of a topic?
How does this approach with compare other methods and tools that can provide such
formative feedback?
Can it be used to best advantage if it is in combined with them?
Our proposed solution is to build an LSA-based service that provides a comparison of the
LSA analysis of learner texts with reference models, to provide meaningful aggregated
information to stakeholders (learners, tutors).
1.6 Relationship between WP4.1 and WP4.2
In summary, both WP4 tasks rely on information extracted from texts. They differ in that
WP4.2 extracts concept-like clusters from the texts, whereas WP4.1 relies upon surface level
phrases and patterns of usage. Workpackage 4.2 is in this sense “knowledge rich” since the
concepts are knowledge-like units extracted from the texts. For WP4.1 there is also
knowledge rich subtask (Section 5.3) relying upon concepts but within this subtask the
concepts are determined externally by use of ontologies. Thus, the knowledge rich subtask of
WP4.1 provides a bridge between the knowledge poor subtask and WP4.2, as they are both
using notions of “concepts”, which will be defined more precisely in the respective sections.
The two tasks also share a challenge in working out how best to aggregate highly complex
underlying data in ways meaningful to end users to meet their business needs. This will be
addressed through validations of the user interface and associated reports with end users, to
enable fine tuning of the outputs to end . This will be achieved in collaboration with WP7
and according to the results obtained, enhancements to the scenarios may be indicated in
collaboration with WP3.
1.7 Overview of this report
This document is based on work taking place up to December 2008 ('first phase'), plus
remedial work required by European Commission reviewers. Later work will be reported in
deliverable D4.2, due in Month 20. This document should be read in association with the
LTfLL -2008-212578
9
D4.1 Positioning Design
separate deliverable "LTfLL consortium’s approach to integration – additional report"
(September 2009), which explores the integration of WP4 services with those of WPs 5 & 6.
Section 2 of this report discusses the theoretical background to adult learning and how our
work relates to these theories.
Section 3 describes the educational domains in which the work of WP4 is situated.
Section 4 describes the texts from these domains that are available for analysis.
Section 5 describes the knowledge poor and knowledge rich positioning services, the work
during the first phase, the conclusions drawn and the proposed next steps.
Section 6 describes the service for diagnosis of conceptual development, the work during the
first phase, the conclusions drawn and the proposed next steps.
Section 7 summarises progress to date and draws overall conclusions.
LTfLL -2008-212578
10
D4.1 Positioning Design
2. Theoretical background
2.1 Introduction
We will now set this work package in the context of relevant developments in educational
theory, perceptions of how knowledge and understanding are acquired and of learner
development. Adult learning in a work place environment is recognized as a social activity,
so we will discuss learner development from this perspective. We will then consider how this
relates to the development of the individual learner and discuss its application to Medicine.
2.2 Theoretical basis of learning
It is extremely important to state clearly the theoretical perspectives of learning considered in
LTfLL, their limitations and the possibility of integrating them, in order to assure coherence
among all the work-packages in the project. Epistemological questions about education
theory are concerned with the relationship between “the knower” and what can be known,
and the extent to which learning concerns ‘real’ external entities, or social constructed truths.
Learning theories that are concerned with learning as absolute truth that can be understood
through an ‘objective detachment” are characterized by Guba & Lincoln (2005) as ‘naive
realism’, characteristic of a ‘positivist’ system of beliefs. In contrast, a ‘constructivist’ system
of beliefs is focused not on knowledge as an absolute external reality, but on its construction
by the knower, individually and as a member of a social community.
There has been a progressive shift towards a more constructivist epistemology of education in
recent years. Furthermore, Wertsch (1991), continuing Vygotsky's cultural-historical ideas
(Vygotsky, 1978), has discussed the need for socio-cultural perspectives on learning in
addition to traditional psychology, on the grounds that the latter tended to study the individual
“ in vacuo”. Morever, Wertsch (1991) has emphasized that a Vygotsky's ideas are extended by
Bakhtin's dialogism (1981). Bakhtin even considers that every text, not just conversations,
are dialogs in which multiple voices interact. This idea is very important because it can
provide a theoretical foundation for social knowledge construction. Mikhail Bakhtin (1981,
1986) considers that dialogism is not limited to conversation but it is rather a general
phenomenon that occurs even in written "utterances". Always there are several voices that
interact, for example, the writer, the potential reader, the echoes of the voices present in each
word. Moreover, from this multivocality perspective, texts become meaning generation
mechanisms, facilitating understanding and creative thought, as Lotman stated (Wertsch,
1991; Dysthe, 1996). A consequence is that in education, "the interaction of oral and written
discourse increased dialogicality and multivoicesness and therfore provided more chances for
students to learn than did talking or writing alone" (Dysthe, 1996). The dialogic and
multivoicesness features of any utterance, even written, may be unifying factors for the
integration of the modules in the language-centered LTfLL project. Therefore, an integrated
framework is provided for analysing all the textual learning activities as searching
documents, reading, writing summaries or forum posts and chatting.
For LTfLL, the two dominant theoretical perspectives are cognitive and social. A cognitive
orientation focuses on perception, memory and meaning; it assumes the memory is an active
processor of information, and knowledge, as a commodity plays an important role in learning.
A social orientation assumes that learning is a social activity, which occurs in interaction with
LTfLL -2008-212578
11
D4.1 Positioning Design
others. It takes account of both the learner and the environment, where learners are not just
products of their experiences, but pro-active producers of the environment in which they
operate. Socio-cultural learning theory, of particular relevance to the workplace learning
environment of this project, originated in the work of the Russian scholar Vygotsky and had
its roots firmly in Marxist notions of collaboration (Wertsch, 1991). Its emphasis is on
learning as an essentially collaborative activity, where not just the processes but also the
products of learning reside in ‘activity systems’ or ‘communities of practice’(Lave and
Wenger, 1991). Socio-cultural theorists are concerned with learners’ social engagement in
communal activity and the identities, language, and cultural artefacts of the social groups in
which they learn.
2.3 Knowledge creation theories
Knowledge creation theories focus on how individuals and groups develop knowledge that is
new to them. They stress that knowledge is not transmitted untouched and unchanged from
one – knowledgeable – person to another person who is unknowing. In contrast, they
emphasize that knowledge is constructed in a dialectical and social process, and that not only
explicitly stated knowledge and information is a source or result of this process but that there
is also a much bigger reservoir of tacit knowledge (Figure 1).
Figure 1: Explicit and implicit knowledge (Brown & Adler 2008)
Contemporary trends in educational and organizational contexts share this view of how
knowledge is created. As described in the integrated report, in the area of collaborative
learning, Stahl (2006), following a social epistemological perspective (Brown & Duguid,
1991; Lave & Wenger, 1991) models the learning process as a mutual construction of
individual and social knowledge building.
LTfLL -2008-212578
12
D4.1 Positioning Design
Figure 2 shows Stahl's cycles of knowledge building. The diagram depicts the interaction
between the cycles of personal and collaborative knowing building. The lower left corner
shows the cycle of personal understanding, which can start with a tacit pre-understanding
influenced by personal knowing. The right part of the diagram depicts how the social process
of interaction with people and with our shared culture influences the individual’s
understanding. Although in the diagram personal understanding and social knowledge
building are separated, it is only a matter of representation, they can only be separated
artificially. Our motivation for introducing the Stahl cycle is that it fits both WP4.1 and
WP4.2 approaches. Moreover, as it was described in the integrated report, WP5 and WP6 also
fit this model.
Figure 2: Cycles of knowledge building (Stahl 2006)
Cycle of personal understanding. As indicated in Figure 2, learning may begin with tacit
pre-understanding, which may change if we clarify the implications of that understanding and
resolve conflicts in our perceptions or fill gaps — by reinterpreting the basis for our
knowledge — in order to arrive at a new understanding. This typically involves some
feedback: from our experience with artefacts such as our tools and symbolic representations.
It is noteworthy that this parallels the constructivist view of “self” development, described in
Kelly’s Personal Construct Theory (Kelly, 1955), where man is viewed as “scientist”
constantly testing and retesting his experiences of the environment in order to construct a
view of himself as an individual. This is discussed further in relation to the development of
learner identity in medicine. This new comprehension is embedded to become our new tacit
understanding and to provide the starting point for and further learning. If we cannot resolve
the problematic character of our personal understanding alone, then we might need to enter
into an explicitly social process and create new meanings collaboratively. To do so, we
typically articulate our initial belief in words and express ourselves in public statements,
LTfLL -2008-212578
13
D4.1 Positioning Design
entering into the cycle of social knowledge building. Conversely, our problems may arise
thorough social interaction and need to be solved by the individual.
Cycle of social knowledge building. (Building collaborative knowing) In this cycle the
interchange of arguments from different perspectives, may converge on a shared
understanding if differences in interpretation and terminology are clarified through
negotiation, thereby culminating in an accepted knowledge. This process can therefore be
viewed as a continuum between personal understanding and generally accepted knowledge.
The knowledge poor subtask of WP4.1, with its focus on surface language, is represented
mostly by the right hand side of the cycle since language is the mean of social
communication. It is understood here that experts in a given field develop a speech genre
(Bakhtin, 1986), which includes terminology and phrasal usage. Thus, a learners degree of
expertise can be indirectly measured by textual distance to expert usage. Moreover, feedback
concerning phrases and usage can facilitate communication, and ease the learner's integration
into the community of practice, leading in turn to increased social learning.
WP4.2 is represented in both cycles. In the left hand side of the cycle, it provides a cognitive
artifact (i.e., individual visualization of learner’s textual inputs) that can help learners to
understand and resolve conflicts or filling in gaps of their knowledge. If this is not possible,
learners enter into the cycle of social knowledge building. In this cycle, WP4.2 provides a
‘cultural artifact’ (i.e. an amalgamated visualization of all peers textual inputs), which can be
seen the joint understanding of that group (at that moment in time) that can help to foster
shared understanding. This is of particular importance to WP 4.2 as a reference point for an
individual learner can be the concepts and the relations between those concepts that a group
of people (e.g. peers, participants, co-workers, etc.) used most often.
It is important to note that from a cognitive viewpoint, the Stahl's diagram does not represent
the skills and sub-processes required for learner development in this context, for example
personal skills, like summarizing discussion, understanding, texts, critical thinking and
logical structuring of arguments; social interaction skills such as turn-taking, repair of
misunderstandings, rhetorical persuasion and interactive arguing. Of particular significance,
this also includes activities for providing feedback in both cycles, to support personal
understanding and social knowledge building.
The interplay between the individual and group learning has also been described by, Nonaka,
Toyama, and Konno (2000) who identify four connected and interacting processes of
knowledge conversion, together the “SECI”-process:
•
•
•
•
Socialization - the permeation of tacit knowledge through and between groups through
shared experiences
Externalization – the articulation of implicit knowledge through distinguishing
phenomena and episodes. Conceptualization and mental modelling are the basis for two
processes: combination and internalization (see below)
Combination – where explicit knowledge is critically analysed and combined with other
explicit knowledge or restructured into more complex new knowledge
Internalization: the process in which new explicit knowledge is embodied, connected to
new contexts and made useful and productive; the implicit knowledge that is then
accumulated can start a new SECI-cycle (Figure 3)
LTfLL -2008-212578
14
D4.1 Positioning Design
This can take place at different levels of sophistication, depending on how people create and
employ a context for implicit and explicit communication, the quality of the input in the
process, etc.
Figure 3: SECI spiral and subprocesses (Nonaka, Toyama & Konno 2000)
Many educational practices start by providing students with explicit knowledge, and only
after this has reached what is considered a critical mass, are they allowed to acquire implicit,
experiential, applied knowledge. Ertmer and Newby (1993) and Jonassen et al. (1993b),
however, do not advocate to a single theory of learning, but emphasize that the instructional
strategy and the content addressed depend on the level of the learners. They claim, therefore,
that behavioural strategies can facilitate mastery of the content of a profession (knowing
what); cognitive strategies are useful for procedural knowledge (knowing how); and
constructivist strategies are appropriated to dealing with ill-defined problems as summarized
in Figure 4
LTfLL -2008-212578
15
D4.1 Positioning Design
Figure 4: The Continuum of Knowledge Acquisition Model (Jonassen et al., 1993b)
Ertmer and Newby (1993) believe that instructional strategies depend on the level of learners’
task knowledge and the level of cognitive processing required by the task. From our point of
view, this implies as well that while monitoring and providing formative feedback (Section
6), the level of expertise of the learner should be considered.
2.4 Theories which model the gaining of expertise and their relevance to professional
development
Theories which explain and predict how learners develop expertise in a specific domain
involve growth paths, an understinding of the essence of expertise, strategies used by learners
and teachers/tutors ("instruction strategies) and physical and/or cognitive changes that occur
during this development. One approach is described by Ericsson (1996, 2004) in the theory of
Deliberate Practice (DP), in which the key concept is that practice in a particular field must
be informed by a good analysis of the present state of mastery, which targeted at the
improvement of specific points, with the help of well-chosen teachers, models, and other
support persons. Ericsson also showed that different stages of development require different
teaching strategies and that the best predictor of the final level of expertise reached was the
accumulated time spent in deliberate practice. The key criterion for defining expertise is
superior performance. This is theoretical approach is extended to professional domains
(Ericsson, 2004). Others(Mieg 2009, Sternberg and Frensch, 1992), however, include the
function of experts within their community of practice in their definition, termed the
"attributional definition". In contrast, the Model of Domain Learning, MDL (Alexander,
2003), describes development of expertise in three increasingly advanced levels (stages):
Acclimation, Competency, and Proficiency. Within these stages, three interrelated dimensions
are proposed that change with level of expertise: (i) domain knowledge that undergoes both
quantitative and qualitative changes, (ii) learning strategies that are related to depth of
knowledge of a domain, and (iii) interest that varies along axes of generality, and contextdependence.
Models for acquiring professional expertise are positioned in between the MDL and DP
theories but are also significantly different from them in some key aspects (Boshuizen and
Schmidt 2008). They focus on the learners’ commitment to specific domains and on the
learning environments, both in an academic and workplace situation. They consider
requirements for acquiring competences, integrating knowledge, skill and developing
professional attitudes. This is further discussed in relation to Medicine. DP may not be
completely applicable in these situations, even if good or excellent performance can be
discriminated from those which are mediocre, it can be difficult to reach a consistent elite
level, due to the breadth of the domain (different but interrelated specialities in medicine), or
continuous changes in the environment or in the subject of the profession (e.g., investment,
meteorology). In studies of professional expertise development, the definition of expertise
levels is less strict, and experience and reputation are used as proxies. Although the models
are tightly linked to the professions, they are able to make general predictions and, like MDL
and DP, focus on knowledge structure and instruction strategies. This further emphasises the
inter relationship between the basis of expertise development and the understanding of
knowledge building and acquisition, discussed previously.
LTfLL -2008-212578
16
D4.1 Positioning Design
Expertise theories, more or less adhere to the definition that experts do a better job than nonexperts (or that professionals do a better job than non-professionals). Some of them also state
that experts are able to move their field further, by being well acquainted with key issues and
being actively engaged in problem solving (Alexander 2003) Although Ericsson (1999) also
states that high expertise and innovation are intricately related, prolonged time practising in a
profession may not equip the professional to cope with a changing environment, and the
individual may be prone to poor practice and skill-obsolescence (Weggeman, 2000; Thijssen
& Van der Heijden, 2003), Expertise should, therefore also include flexibility and
employability (Van der Heijden, 2000) and exhibit routine and adaptive expertise (Hatano &
Inagaki, 1986) and emphasizes the need for training for transfer at every level of learning
(Nokes & Ohlssen, 2003; Salomon & Perkins, 1989). This is an essential component of
lifelong learning and is an important basis for WP 4.1.
2.5 Knowledge restructuring
In MDL (Alexander, 2003) the process of gaining proficiency involves an extension of
knowledge and connecting between concepts. This process of knowledge restructuring has
been intensively investigated in medical education and is especially important to reevaluating how basic sciences knowledge is introduced and used. Thus, knowledge
encapsulation, namely forming macro concepts (encapsulations) that subsume biomedical
concepts under clinically relevant headings (e.g. in terms of diagnosis, patient management or
treatment), plays a prominent role (Schmidt et al., 1990). Biomedical knowledge provides
structure to isolated clinical case concepts (Woods, 2007; Woods, et al., 2005). Woods et al.
(2007) demonstrated that learning causal explanations for features of clinical conditions
resulted in ability to make a quicker and more accurate diagnosis in complex situations.
Illness script formation (Schmidt et al., 1990) is another knowledge restructuring process
reviewed in detail by Charlin et al., 2007. This process takes place under the influence of
practical application of knowledge and is summarized in Table 1. Even in early stages of
learning, application of knowledge is linked to the conditions in which this knowledge is
acquired. This non-analytic aspect of knowledge application is rapid and not under conscious
control. Depending on the context of learning and later application the effects can be
advantageous or disadvantages (see Norman et al., 2006, pp 344-347, for an overview).
Enabling learners to understand how they are restructuring their knowledge, in a manner that
is appropriate both to their level of expertise and to the social context of their learning, is an
essential feature of the services that Work package 4 aims to provide.
Table 1: Knowledge restructuring, clinical reasoning and levels of expertise (Boshuizen & Schmidt, 2008;
reprinted with permission from Elsevier)
Expertise
level
Knowledge
representation
Novice
Networks
LTfLL -2008-212578
Knowledge
acquisition
and
(re)structuring
Clinical
reasoning
Control
required
in clinical
reasoning
Demand
on
cognitive
capacity
Clinical
reasoning
in action
Knowledge
accretion and
validation
Long chains
of detailed
reasoning
steps
through pre-
Active
monitoring
of each
reasoning
step
High
Difficulty to
combine
data
collection
and
17
D4.1 Positioning Design
Expertise
level
Knowledge
representation
Knowledge
acquisition
and
(re)structuring
Clinical
reasoning
Control
required
in clinical
reasoning
Demand
on
cognitive
capacity
encapsulated
networks
Clinical
reasoning
in action
evaluation
and clinical
reasoning
Intermediate
Networks
Encapsulation
Reasoning
through
encapsulated
network
Active
monitoring
of each
reasoning
step
Medium
…
Expert
Illness scripts
Illness script
formation
Illness script
activation
and
instantiation
Monitoring
of the level
of script
instantiation
Low
Adjust data
collection to
time
available and
to
verification/
falsification
level of
hypotheses
2.6 Professional learning
In the context of professional learning and development, Schön (1987) approached
application of knowledge to professional problems and situations from the perspective of the
practices in which learners were trained and in which they work. A key element is reflectionin-action, which results from monitoring one’s own action and is at least partly conscious. It
involves recognizing unexpected phenomena, and the consequences of “knowing-in-action”
i.e. the active use of knowledge and understanding. Reflection-in-action can provoke
questioning of in-action knowledge and may lead to a revised understanding and experiments
with problem-solving strategies, etc. Although Schön’s theory takes this in-action aspect for
granted, practices and professions have different time structures, which may constrain the
possibilities for reflection-in-action (Eraut, 1994), for example flying an aircraft, or operating
on a patient. Eraut suggests high-speed practices require instant recognition and response, and
routinised, unreflective action at the time but would require subsequent reflection post action,
for example in the medical audit process. Low speed practices allow deliberative analysis and
decisions, with actions following a period of deliberation. The mode of cognition in
intermediate situations would consist of rapid interpretation and decisions, and action
monitored by reflection. Working in low-speed workplaces can go hand in hand with
learning. Medium and high-speed work requires both preparation of learners, to help them
become aware of what can be expected, and reflection-after-the-fact. The ability of a learner
to understand his/her learning process and to respond to feedback provided by peers, tutors or
experts is therefore essential to reflective practice
LTfLL -2008-212578
18
D4.1 Positioning Design
2.7 Learning processes
The learning processes taking place in education either in formal or non-formal learning
contexts, both undergraduate and post graduate, are the same as those described for expertise
development: knowledge accretion, restructuring and script development. Which of them
presents most difficulty to students depends on the domain involved. Knowledge of these
difficulties is part of Pedagogical Domain Knowledge , PDK (Shulman, 1986). Those in
mathematics are different from those encountered in medicine or law. Conceptual change and
the resistance to it is a problem shared by many domains (Vosniadou & Ortony, 1989).
Approaches have been formulated which may help learners reach a change in their
understanding of key concepts (Chinn & Brewer 1993), for example integrating several
different perspectives of a specific problem (Spiro et al., 1992) or the development of specific
problem-solving scripts (van Merriënboer et al., 2002). The teacher/tutor may be regarded as
encapsulating PDK. They develop their approach to learner’s difficulties in their daily
interaction with them within their domain of study and use this knowledge to adapt
instruction. A good teacher adapts instruction and feedback to the learner’s Zone of Proximal
Development (Vygotsky, 1978) and can be considered part of the teacher/tutor’s PDK.
Hatties’ (1996) meta-analysis of the effectiveness of teaching strategies showed that the most
powerful strategies are:
•
•
•
Provide feed forward and feedback information that tells students how they are doing,
where they are going, how they are going and whereto next; help and stimulate learning
from feedback.
Provide real challenges and communicate learning intentions and success criteria.
Provide opportunities for modelling, both life (by expert and coping peers) and in the
form of worked examples.
LTfLL -2008-212578
19
D4.1 Positioning Design
3. Educational domains included in WP4
3.1 Medicine as a model
Modern medicine requires clinicians to be independent life long learners, with a clear view of
the limits of their expertise and competencies. The continuing advancement of research in
medical and human sciences necessitates constant adaptation and revision of practice in the
workplace to ensure patients’ care and treatment are both safe and effective. The ability to
sustain lifelong learning by doctors is best achieved by a self – awareness of how the
individual learns. Instilling the characteristics and behaviours of independent lifelong
learning in doctors is therefore major goal of modern medical education. Learning in the
clinical workplace is typified by interactions and exchanges with fellow learners, senior
clinicians, other healthcare professionals (nurses, pharmacists etc.) and, most importantly
patients. They exemplify models of Communities of Practice (Lave and Wenger, 1991). In
undergraduate medical education, approaches to learning are used which aim to accustom
students to this environment, for example Problem Based Learning (O’Neill et al, 2002).
Bleakley and Bligh (2008) describe modern medical education as patient focussed,
identifying the learning of a young trainee doctor as a clear example of situated learning,
focussing on information gained from patients by the learner’s interaction with them. All
these interactions represent learning experiences, on which the learner reflects and from
which they construct their own knowledge, similar to that described in Stahl’s model, which
we have discussed previously. The introduction of the learner at a junior level into this
environment also raised issues of their personal and professional development, even at an
undergraduate level. Their perceptions of self identity as a learner, which we discussed
previously in relation to Kelly’s personal construct theory (Kelly, 1955) are significant
influences on the importance that students attach to lifelong learning and reflective learning
(Lown et al 2009).
3.2 Pedagogic domain: Reflective and critical learning in medicine – role of portfolio
The ability to think critically is key to enabling an individual to understand limitations of
knowledge, competencies, skills, attitudes and behaviour throughout medicine but especially
for medical students and newly qualified doctors. We have already discussed the basis for this
in Schon’s theory of reflection-in action and the requirement for post action reflection. An
example is shown in Figure 5.
LTfLL -2008-212578
20
D4.1 Positioning Design
Figure 5: Example of reflection post action in Medicine
The awareness of such limitation is an essential prerequisite for obtaining feedback from
tutors, seniors with expertise and acting upon it to formulate a learning plan for future
development and is an important basis for WP4.
The evidence for that reflective learning has been implemented is now routinely maintained
in learner’s portfolios. These are widely incorporated into many professions; nursing,
teaching, psychotherapy and counselling, law, engineering are all examples of careers which
require mandatory (e)portfolios. The content of such portfolios are defined by the appropriate
regulatory bodies for each profession and are required for scrutiny and evidence during
regular professional appraisals. Its role has therefore developed from its earlier function as an
indicator of an individual’s learning and achievements to one in which the evidence that it
contains clearly demonstrates the knowledge and competences required to progress within
professions. The current view of portfolios (both electronic and otherwise) has changed since
the DoW was written, and can no longer be regarded as a collection of essay type texts
generated by the learner, but are more flexible and variable in content to accomodate the
requirments of modern professional learniing and development. Thus, the Centre for
Recording Achievement (http://www.recordingachievement.org/) states that eportfolios
should include criteria/standards/outcomes for learners (e.g. assessment); enable learners
to create and store plans which can be shared with others tutor/mentor/coach, possibly
peers), which can be re accessed, and provide a means of recording experiences and
achievements as they happen, which can be selected if they are relevant for a review by a
senior. Learning plans and their re-assessment are relevant to the reflective learning cycle
discussed above.
In the UK, all university students are expected to maintain a record of their personal and
professional development, but also for medical students, the GMC stipulates that personal
and professional development ePortfolios are an essential component of the undergraduate
curriculum. They contain
•
Evidence of engagement in activities that promote professional development (i.e. that is
signed by the appropriate authorities, is dated and is verifiable)
LTfLL -2008-212578
21
D4.1 Positioning Design
•
•
•
•
•
•
•
Evidence that the student is acquiring key clinical skills and the level of competency
reached in those skills
Critical self analysis of the student’s own strengths and limitations and how they are
building on the former and addressing the latter
Evidence that the student is responding to the feedback given to him/her by those
assessing development in skills and clinical practice
Demonstration of how the individual student is developing their own learning strategies
and clear indications that they can formulate their own learning plans, especially at key
stages in the curriculum
Evidence that they are developing an understanding of disease processes, which they
apply to their clinical learning situations
Clear indications that students understand and develop competencies in key features of
modern medical practice e.g, team working, communications, governance of medical
practice and evidence based medicine
Critical thinking and reflective learning, which underpins many of the above features of
the ePortfolio.
In summary, the “ideal” ePortfolio for the medical student is one which genuinely helps and
supports the student’s own personal and professional development. It demonstrates his/her
development as a learner, both in terms of understanding and acquisition of skills and
competencies, based on ever improving critical and reflective thinking and learning, thereby
preparing the student for their life as a medical practitioner. In workpackage 4, the proposed
automatic analysis (i.e. knowledge poor and knowledge rich based approaches), technology
has the capability to inform and guide students through composite tasks of the “ideal”
ePortfolio. The provision of feedback, supported by the services developed in WP4.2, and the
learner's responses to this, would form important components of such portfolios. Conversely,
some of the text material included in ePortfolios, for example participation on reflective
online discussions, can be used as material for analysis with LSA. (See Section 4.2.2 below).
3.3 The role of ePortfolios in Computer Science education
One objective of the project is to apply the techniques and services, which we will develop, to
more than one domain and language. We decided to choose computer science as the second
domain for the following reasons: a) as one outcome of the LT4EL project
(www.let.uu.nl/lt4el/), we have access to a large number or learning objects from the domain
of computer science and in various languages; in addition we have access to lexica which
have been derived from these learning objects and to an ontology connecting all the terms
through language-independent concepts; b) some of the partners have experience in teaching
computer science.
The situation with ePortfolios for this domain is far from ideal. In contrast to e.g. medicine,
the use of ePortfolios is not an established practice in science and engineering (cf.
Bhattacharya et al., 2006). Within the consortium there are no partners actively using
ePortfolios in Computer Science as part of their teaching. We will therefore generate
eportfolio materials. In any case we are convinced that the practice of using ePortfolios will
become more widespread, and therefore we see an increasing number of users of our service,
which uses ePortfolios in order to position learners and monitoring their conceptual
development.
.
LTfLL -2008-212578
22
D4.1 Positioning Design
4. Texts available for analysis
4.1 Introduction
Learner text that have been made available by the University of Manchester, for testing the
tools and services that are being developed. The learner population is the undergraduate
students of the University of Manchester Medical School. For WP4.1, text material generated
in the curriculum by online discussion will be used, whereas for W4.2, conceptual
development, a variety of text material will be used including learning diaries (blogs) created
by individual learners specifically relating to problem based learning cases and clinical
problems encountered in the work place environment and transcripts of "think aloud"
protocol, described in section 6.5.1.1.
4.2 Text material generated in Manchester University medical curriculum
4.2.1 Theoretical Basis of Online Discussions
For undergraduate students in the University of Manchester Medical School, development
of independent learning strategies and self directed study are necessary for progress in the
curriculum, in which problem-based learning is an essential component. This is also
reflected by the use of online learning, which is becoming more important in undergraduate
medical education. It involves the learner as an individual but also can re recreate some
aspects of Communities of Practice by using structured group discussions, to promote
reflective learning and critical thinking. The contribution of individual group members to
these discussions are therefore evidence of their interactions with others group members and
of their own cognitive development (Braidman et al., 2008) The asynchronous nature of these
discussions, allows time for reflection on workplace experiences, so that participants respond
reflectively and critically in the context of their own experiences (Newman et al., 1997).
Learner and group development in these situations is explained by a Community of Inquiry
model (Garrison et al., 2000), based on Wenger’s Community of Practice concept, and
reminiscent of Stahl’s model, discussed previously. It explains development of such online
groups as reflective learners through interplay between cognitive development, social
interaction and influence of tutor/facilitator
4.2.2 Online Discussions in the University of Manchester Medical School
Asynchronous online discussions were introduced into Years 3 and 4 of the medical
curriculum, when students are in a clinical workplace learning environment, with the purpose
of enabling them to discuss issues concerned with personal and professional development.
Large student numbers and their geographic dispersal required the introduction of
electronically based discussion fora.. Students work in peer facilitated groups to discuss
specific contemporary issues throughout the curriculum, with a nominated student taking on a
facilitator role. There are 63 discussion groups in each of years 3 and 4, consisting of 8
students, including the student facilitator. All groups discuss the same topic, which chages
each semester. Each text flow usually consists of at least 8 – 10 postings. Although some of
these are of just one or two lines, mostly they consist of at least 10 lines of text. All
discussion groups participate i.e. the entire student population of Year 3 (452 students)
LTfLL -2008-212578
23
D4.1 Positioning Design
participated and each discussion topic stimulates over 2,000 postings. Participation in these
discussions is an essential component of the students' eportfolio
These discussion fora provide ideal datasets for LSA, both for comparative analysis and for
corpus based analysis. For learner positioning, this provides some interesting possibilities
across the groups, to suggest concepts that other student groups may explore, for example. It
may also be possible to determine the progress of individuals within a particular group. An
example is provided by discussions on the safe prescribing of medicines. This topic was
initiated by suggestions to the students of the type of knowledge- based material on which to
base their discussions, including some detailed compendia of drugs and treatment e.g. the
British National Formulary. This may provide a sound corpus on which student discussions
can be analysed and feedback provided to ensure good coverage of the subject area.
4.2.3 Data sets for LSA based positioning
The specific texts found most suitable as data sets for LSA were the online discussions
related to the safe prescribing of medicines. These are based on a series of clearly defined
learning objectives, which formed the basis of six main topic areas, which could be used for
text analysis:
•
•
•
•
•
•
The role of medical students in patient safety
The role of the ward pharmacist in preventing serious medication errors
The critical points in medicines management where serious errors can occur
The swiss cheese model of accident causation
The minimum core knowledge, skills and attitudes required to prescribe safely
The legal consequences of negligent or reckless prescribing / administration of drugs.
Individual students' contributions have been selectively sampled based on a set of criteria that
will provide a good quantity of content. Each contribution is presented in a format that
includes the student identifier, the topic area that is covered and a grade (i.e. Excellent, Good,
Fair, Poor) indicating the effectiveness with which the student has covered the domain topic.
4.2.4 Data generated by “think aloud” analysis of PBL cases
As discussed before, there is a major need to provide structured formative feedback to the
students as individuals, of their conceptual development, resulting from their learning in
their PBL sessions. An appropriate means of achieving this is to record a student’s brief
analysis of a recent PBL case and its relevance to the clinical condition concerned. The
transcribed text output is suitable for language analysis that will allow a grading on which
feedback to the student can be based. This is referred to as a "think aloud" protocol. The
feedback will focus on their understanding of the subject area and their ability to relate it to
the clinical problem originally posed. This method of text generation is especially appropriate
to the medical domain as clinicians routinely provide verbal reports of clinical problems to
colleagues, which often include a range of healthcare professionals.
LTfLL -2008-212578
24
D4.1 Positioning Design
5. Positioning the learner (Task 4.1)
5.1 Introduction
As life long learners have neither common learning goals nor common educational
backgrounds, as it is the case in traditional learning settings, life long learning educational
providers need to rely on available written materials produced by individual learners to
identify their degree of expertise within areas of knowledge that are relevant to study
programs that are offered. Lifelong learners seeking to further develop their competences will
consider the different educational programs that best fit their needs. Accreditation or
Recognition of Prior Learning (APL/RPL) provides a procedural solution for identifying the
prior knowledge of learners in formal and informal education (Merrifield et al., 2000). This
procedure is time consuming and maybe inaccurate therefore a (semi-)automatic procedure
that addresses either or both of these two problems is worth considering. In this case,
evidence of prior knowledge is provided by the learner in the form of texts, which need to be
compared to public or domain knowledge. In this section, we describe knowledge-poor and
knowledge-rich approaches to analysing learner texts. The WP4 description of work
describes LSA and related methods (e.g. probabilistic LSA) that use text as the only source of
knowledge as knowledge poor methods. Additionally, it describes the use of ontologies and
lexical knowledge resources for the analysis of texts as knowledge rich methods. Knowledge
poor based approaches can be considered as a base line for the knowledge rich approaches.
5.2 Knowledge poor based positioning
In this section, we introduce our knowledge poor approach to positioning. The approach that
we will present is based on comparing words and phrases extracted from a learner's text(s)
with terms and phrases extracted from model, expert texts. Overlap of these usage patterns
will then be interpreted to mean that the learner has mastered parts of the expert domain. We
will argue that this seemingly indirect approach is, in fact, reasonable. The more direct
approach would be to analyze the learner's texts to determine what the learner knows and
does not know. But this direct approach presupposes a degree of natural language
understanding which is not available with current language technologies, and probably never
will be available. Therefore, an indirect approach to assessing acquired knowledge is the only
path available. We will argue that, since knowledge is, to a large extent, acquired socially,
becoming an expert in a domain involves socializing with members of a community of
practice, which in turn involves adoption of the speech genre of this CoP. Thus if a learner
uses terms and phrases typical of this CoP, it is evidence that the learner may have socialized
with the CoP, and has thus, most probably, also acquired some of the expertise of the CoP.
This line of reasoning is quite indirect, and clearly requires validation. We believe, however,
that this is the only reasonable way to make sense of a knowledge poor based approach to
positioning.
An added benefit of the approach is that it involves identification of words and phrases that
are typical and atypical of a CoP. These phrases can be used to provide qualitative feedback
to the user of the knowledge poor positioning service. If a learner uses phrases which are
atypical of the CoP, the user can be informed of this inappropriate usage and more
appropriate, synonymous phrases can be suggested. If the user then adopts these more
LTfLL -2008-212578
25
D4.1 Positioning Design
appropriate phrases into his or her linguistic usage, this will lead to facilitated communication
with members of the CoP, which will, in turn, lead to increased social learning.
5.2.1 Integration of the positioning scenario within the LTfLL service platform
Although current scenarions for the positioning service implementation are relevant only to
the Information Technology domain, the basic principles of our services to be validated by
WUW and Bitmedia remains the same. Thus, this document focusses on the Medical domain.
Services providing support in determining the learner's degree of expertise or position by
means of LSA in the IT domain will be are being implemented by WUW. The
implementation of those scenarios is being guided by the philosophy of the relevant 4.1
scenarios written by Bitmedia and UTU. Bitmedia is a stake holder in the validation of the
mentioned service and is currently working with UTU and WUW in developing the data sets
requiered for that end (annotation of texts and training materials in the IT domain). A second
service for positioning the learner will be implemented in a later stage using the knowledge
rich approach.
5.2.2 Desiderata for (semi-)automatic positioning
We are interested in applying the techniques of text categorization for the purpose of
positioning life long learners. Quite simply, we can rate learner texts, likely to be short and
generated in informal educational settings, by using a vector-space comparison to goldstandard, expert texts. Then if the similarity is high enough, the learner will proficiency out of
the course. This approach is straightforward, but in it's naive version, it is unlikely to be
successful. The problems concern accuracy, suitability and justification of the categorization.
Accuracy is clearly important. False positives can result in learner frustration in courses
beyond his or her level, and similarly false negatives can result in learner boredom.
Suitability refers to model texts against which the learner texts are compared. If these are not
suitable (or prototypical) models of expert language usage in the domain, then the positioning
will not be valid. Thirdly, justification refers to the reasoning given by the service for the
positioning decision. It is well known that decisions of expert systems are more accepted
when the system gives reasons for its decisions. All three of these issues must be addressed
by our system.
5.2.3 LSA based positioning
5.2.3.1 General introduction
In recent years, Latent Semantic Analysis (LSA; Landauer and Dumais, 1997) has been
proposed as a suitable language technology for the automatic positioning of learners (Van
Bruggen et al. 2004). Although, LSA has been successfully used in the context of language
technologies enhanced learning (e.g. automatic assessment of student essays), learner
positioning presents new challenges that expose the limitation of such an approach. In
particular, learners produce text repositories containing few samples of text, many of them of
small size and using language that is rich in non domain-specific expressions e.g. email
messages, chat conversations, forum online discussions, blog postings, etc. In addition, those
texts are generated in contexts where learners feel encouraged to hide their poor usage of
language by articulating redundant expressions and making extensive use keywords. In
addition, LSA is also limited in that is not capable of recognizing directionality in causal
relationships. Burek et al. (2007) presents a solution to this problem by means of a triple
based LSA that calculates a set of similarity measures between the semantically related
constituents of the sentence structure (i.e. subject, verb and object).
LTfLL -2008-212578
26
D4.1 Positioning Design
5.2.3.2 Accuracy
The categorization obviously must be accurate. A false positive, indicating learner
proficiency in a particular domain, could be dangerous as it could lead to work place
incompetence. A false negative, on the other hand, could lead to boredom, as the learner is
forced to take courses on topics that he or she has already mastered. The problem of accuracy
is compounded by the fact that texts (selected from text collections are often short. To deal
with this problem, our approach attempts to lose as little information from the text as
possible. Traditional approaches to categorization lose information by case normalization,
stemming and ignoring word order. The idea of the traditional approach is to deal with the
data sparseness problem by collapsing textual features into equivalence classes, losing
information in the process.
In our approach, we attempt to balance the problem of data sparseness with the goal of not
losing information. This balance is obtained in two ways. First, we use LSA as a technique
for dimensionality reduction. It is well known that LSA can be used to discover weighted
clusters of words, which are loosely understood to be ''concepts''. Since these clusters can
contain derivationally related terms, the need for stemming (and also case normalization) is
reduced. Second, our more innovative contribution is to flexibly use n-grams of different size
(phrases) and the above mentioned relational triples as opposed to strictly unigrams in the
traditional bag-of-words model. Our approach to extracting such n-grams is to use an
extension of the suffix array approach of (Yamamoto and Church, 2001).
5.2.3.3 Suitability
Suppose that learner texts could be accurately classified as similar or not similar to the gold
standard text (or set of texts). Then the question arises as to whether or not the gold standard
text is a suitable prototype for a good learner text. One approach to choosing a gold standard
text would be to use a published journal article in the field. But such a text is unlikely to be
similar to learner texts either in tone or in content. It is well known that effective teachers use
scaffolding to present material within the zone of proximal development of the learner.
So perhaps a better gold standard would be a text written by an expert e.g. textbook, blog
entry, etc. or other learning material, written at the level of the student. This is certainly an
improvement, but on the other hand, it is still rather unreasonable to expect learners' texts to
closely match the tone of a textbook, unless of course the learners are copying from the text.
In fact, the texts that we have consist of online discussions of medical students on several
topics related to safe prescribing. These texts have been categorized as to subtopic and graded
for quality (excellent, good, fair, poor) by annotators at the University of Manchester. The
texts contain serious conversations, with very little off-topic wandering. But the tone of the
texts is chatty, and not at all similar to textbook writing. So rather than to use an external gold
standard, we have opted for an internal gold standard. The prototypical “excellent” text is
simply one that was rated as “excellent” by the annotators. But not all “excellent” texts are
equally good as prototypes. Clearly, for any text t, the remaining texts can be ranked in order
of similarity to t. If t is a good prototype, then this ranking should have other “excellent” texts
as most similar to t and “poor” should be least similar. So we need to choose as prototypes,
those texts that induce the best ranking. For purposes of comparing such rankings, we have
experimented with the nonparametric permutation test.
LTfLL -2008-212578
27
D4.1 Positioning Design
5.2.3.4 Justification
As stated above, it is well known that users are more inclined to trust an expert system when
the system can give some reasons for its judgement. Our system is designed to give both
quantitative and qualitative feedback. The quantitative information indicates simply a
distance measure of the learner text from the expert text. The qualitative information, on the
other hand, uses the phrases that are extracted using the suffix array algorithms. These
phrases are weighted according to the probability of occuring predominately in expert or nonexpert texts. When the learner uses non-expert terminology, this can be reported to the learner
as useful qualitative feedback. When the learner sees that the system is able to perceive
subtle, perhaps unconscious, patterns of language usage, this will help to increase the
learner's confidence in the quantitative judgement of the system.
The use of phrases, however, presupposes a fairly large body of training texts to overcome
the data sparseness problem. Our phrase weighting approach is, however, flexible in the sense
that if no distinctive pattern of use is detected for longer phrases, then the system will fall
back to shorter phrases or single words.
5.2.4 Our progress to date
5.2.4.1 Phrases
Traditionally, text categorization by means of LSA has relied on a bag-of-words model. It
seems, in some sense, obvious that a model based on phrases should be better. But it turns out
that this is not necessarily the case. Recently, Bekkerman and Allan (2004) reviewed the
literature on text categorization and found no general improvement when unigram models
were replaced with bigram models. The problem is that using bigrams contributes heavily to
the data sparseness problem. Bekkerman and Allan have, however, compared two rather
extreme positions. Our idea is to extract phrases of any length from the the training corpus, as
long as the phrases are distinctive (occurring predominately in particular categories of
documents). It may well be that the most distinctive phrases are generally phrases of length
one (concurring with the bag-of-words model), but if there are phrases of other lengths that
are more distinctive, then there seems to be no reason not to use these phrases. To give an
idea of the approach, consider the word ''side''. In the medical discussions in our corpus, this
word almost always occurs as part of the phrase “side effect(s)”. In a few cases, ''side'' occurs
in a unique context or as part of another phrase, such as ''flip side''. In this case, the distinctive
phrase is apparently ''side effect'', and the other occurrences are just noise. These noise
phrases are not only unhelpful for text categorization, they are are also unhelpful for
generating explanations that would be useful for learners and examiners. The example above
raises some interesting counting issues. But first we need to specify more precisely what it
means for a phrase to be distinctive.
5.2.4.1.1 Distinctiveness
In general, phrases that are evenly distributed across document categories are not very
distinctive, whereas phrases that tend to cluster in one particular category are distinctive. This
general principle must be applied carefully, however, since with small numbers, clustering
may occur due to chance. A common measure of distinctiveness used for weighting in vector
space models is tf-idf (term frequency multiplied by inverse document frequency) discussed
by Salton (1988). It is unclear, however, that this is the best measure for picking out which
phrases to consider and which phrases to ignore. It is problematic, for example, that idf
simply prefers terms that cluster in a small number of documents, regardless of the
LTfLL -2008-212578
28
D4.1 Positioning Design
classifications. Given the ordinal classification of Manchester text set as “excellent”, “good”,
“fair” and “poor”, we are not interested, for example, in terms that cluster in the “excellent”
and “poor” texts. So a distinctive term should be one that occurs predominately in “excellent”
and “good” texts or predominately in “fair” and “poor” texts. Consider, for example, the
bullet point, with occurrence vector [31,5,0,0] [1]. The interpretation is that there are 31
occurrences in ''excellent'' documents, 5 occurrences in ''good'' documents and no occurrences
in either of the poorer texts, this term appears be very distinctive of better texts. But if we
count instead the number of different documents the bullet point occurs [4,1,0,0], we see a
very different picture. The bullet point does occur in higher rated texts, but it is very bursty
and is therefore not very useful for categorization. There are various approaches in the
literature for dealing with burstiness. Since this is not our primary concern here, we deal with
the problem by counting the number of texts containing a term rather than the total number of
occurrences of the term. Thus, for the bullet point, we use the vector [4,1,0,0].
To rate a term such as the bullet point, we need some measure of goodness for the vector
[4,1,0,0]. There is clearly no objective measure that can be used here. As a fairly reasonable
score, we simply assign 1 point for every “excellent” text, 0.8 points for every “good” text
and 0.2 points for every “fair” text. So, the bullet point receives a score of 4.8. This appears
to be a good score, but what is the probability that a randomly chosen term appearing in 5
texts would have a higher or equally high score? We can answer this question by using a
simulation. Random vectors are generated according to the known proportion of
“excellent”, “good”, “fair” and “poor” texts. Then, for a high score such as 4.8, the idea is to
count the proportion of randomly generated vectors have an equally high or higher score. And
for a low score, the opposite idea is to count the proportion of randomly generated scores that
are equal or lower.
5.2.4.1.2 Phrase extraction
In principle, the distinctness measure given above can be used with phrases of any length. If
longer phrases can be found that are more distinct than single words, then there is no reason
not to use the longer phrase. The problem is that the simulation-based distinctness test is very
expensive, and it is certainly not possible to run this test for n-grams of every length in a text.
The solution to this problem comes from Yamamoto and Church (2001), who show suffix
arrays can be used to put the large number of n-grams into a much smaller number of
equivalence classes. Using suffix arrays, it is very easy to pick out just the phrases that are
repeated n times for some n, and it is very easy to extend phrases to the right: if “mumbo
jumbo” repeatedly occurs together as a phrase, then it makes no sense to count “mumbo” by
itself. Yamamoto and Church's suffix array program will put these two phrases into an
equivalence class, so that that statistics can be calculated for the class as a whole rather than
individually for all the members of the class. Since the time of Yamamoto and Church's
paper, suffix arrays have been an active area of research, primarily in bioinformatics. One of
the weaknesses of the suffix array approach used by Yamamoto and Church is that extensions
to the left are difficult to discover. So it is difficult to discover, for example, that “jumbo”
always combines to the left to form the phrase “mumbo jumbo”. Simply stated, the problem
is that suffixes are extensions of phrases to the right, so it is hard to look to the left. This
problem was solved, however, by Abouelhoda et al. (2004), who added a Burrows–Wheeler
transform table to their extended suffix array data structure, giving this this data structure
properties of suffix trees. One weakness of Abouelhoda et al.'s approach, however, is that it
does not adapt well to large alphabets. This is, of course, a serious weakness for use in text
processing, where one wants at least to work with some subset of Unicode, or even worse, to
LTfLL -2008-212578
29
D4.1 Positioning Design
treat each tokenized word as an alphabet symbol. Fortunately, the restriction to small alphabet
size has recently been eliminated in the approach of Kim et al. (2008), who deal with the
large alphabet by using binary trees, which are linearly encoded using the child table of
Abouelhoda et al. along with a longest common prefix table (lcp).
Using extended suffix arrays makes it possible to count different kinds of occurrences of
phrases in different ways. To begin with, we are only interested in counting phrases that
repeat. In the text S = “to be or not to be”, the occurrence of the phrase “to be” at [1, 2] is
said to be a repeat since the same sequence of tokens occurs at S[5,6] [2]. The difference is
that Abouelhoda et al apply the terms to a pair of occurrences, whereas we apply the terms to
a single occurrence. An occurrence of a phrase S[i,j] is left maximal is left maximal if the
longer phrase S[i-1,j] is not a repeat. Thus, for example, the phrase to at S[1,1] is left
maximal since the phrase at S[0,1]is not a repeat [3]. Similarly, an occurrence of a phrase at
S[i,j] is right maximal if S[i,j+1] is not a repeat. If an occurrence of a phrase is both left and
right maximal, then the occurrence is said to be maximal. Note that the occurrence of the
phrase or not at S[3,4] is maximal, though it is not a repeat. Since non-repeats are rarely of
interest, we generally assume that we are talking about repeats unless otherwise stated. A
phrase is also said to be maximal in a text if there exists a maximal occurrence of the phrase
in the text. For example, in the text “mining engineering”, tokenized by characters, the phrase
“in” is maximal since there are maximal occurrences at S[2,3] and S[11,12]. But the longer
phrase “ing” is also maximal since it occurs maximally at S[4,6] and S[16,18]. So the
occurrence of “in” at S[16,17] is a non-maximal occurrence of a maximal phrase. A maximal
repeated phrase that is not a subsequence of a longer maximal repeated phrase is said to be
supermaximal. Thus the phrase “ing” is supermaximal in this text.
Generally, we are only interested in counting occurrences of maximal phrases since a phrase
that never occurs maximally is unlikely to be of interest. But what kind of occurrences should
we count? Should we count all occurrences, or only the left maximal, right maximal or
maximal occurrences? The answer is that we don't need to decide ahead of time. We can
simply test each of these four cases for distinctness, and chose the most distinct case. Take,
for example the word “side”, which is a maximal phrase in our texts. Should we count all
instances of this phrase? Or should we perhaps restrict the count to right maximal
occurrences so as to avoid counting those instances that are extended to the right to create the
longer phrase “side effect”? Or maybe left maximal occurrences to avoid the longer phrase
“flip side”? Or perhaps we should restrict in both directions to avoid either kind of extension.
Since it is not generally possible to predict which is best, the reasonable approach is to try all
possibilities to see what works best.
One counterintuitive feature of our approach is that it also makes sense to count 0-grams. A
left maximal occurrence of a 0-gram, for example, must have a hapax legomena to its left,
and a maximal occurrence of a 0-gram must have hapax legomena on both sides. These
sequences of two hapax legomena may well be distinctive, since they often are an indication
of a named entity or a foreign phrases. Counting all occurrences of the empty sequence is, of
course, equivalent to counting the text length, which may well also be a distinctive feature.
5.2.4.2 Experimental setting for LSA based positioning
The experiments described in this section demonstrate the use of language technologies
involved in the implementation of the LSA based approach for positioning. The experiments
compare training set results obtained with the traditional bag of words LSA configuration
LTfLL -2008-212578
30
D4.1 Positioning Design
against the alternative configuration that uses maximal phrases as the unit of analysis. The
alternative LSA configuration starts with a vector space model that (instead of using words
counts) uses counts of distinctive phrases that occur at least once as a maximal phrase within
the text collection under analysis.
5.2.4.2.1 Data
Our ongoing work in word co-occurrence models for learner positioning extends the existent
LSA based approaches and is aimed at analysing and then scoring texts posted on an online
medical student discussion forum, where University of Manchester students discuss issues
related to one of 6 subtopics of the general topic of safe prescribing. We built a training set
consisting in 504 postings that were annotated by experts with four grades (i.e. 109 poor, 200
fair, 142 good, 50 excellent) and one of six topics (i.e. 42 of topic 'a' , 50 of 'b', 130 of 'c', 22
of 'd' , 247 of 'e' and 13 of 'f'). Each grade is based on the individual posting's textual
contribution to a series of expected learning outcomes. Highly scored postings can then be
used as evidence of learner proficiency in the corresponding topic.
5.2.4.2.2 Building the bag of words and phrase based vector spaces
As already explained, to identify and extract the maximal phrases we analyse suffix arrays
using an extended version of the Yamamoto and Church algorithm to generate all n-grams
from a text and avoiding the combinatorial explosion by grouping these n-grams into
equivalence classes. Each phrase was counted in one of 4 ways: all instances, leftmaximal, right-maximal and maximal. To avoid an unmanageable level of sparseness we
include in the analysis all instances of all phrases that occurs at least one time as
maximal. Phrases are sorted by their scores absolute values. We then built a 19730 phrases to
504 chat texts matrix that contains the frequency of occurrence of each phrase in each texts.
We then weighted the matrix using the tf-idf weighting scheme. We then generate three LSA
semantics spaces by reducing the SVD resulting matrix singular values to 50, 100, and 200
respectively. In addition we created another set of 3 bag of words based LSA vector spaces
using the same weighting scheme and respective number singular values. In this case using a
6320 tokens to 504 chat texts matrix. The number of token used is the results of choosing the
tokens that occurs at least two times within the chat texts collection.
5.2.4.2.3 K Nearest Neighbour based classification
The k Nearest Neighbours algorithm (kNN; Cover and Hart, 1967) is a learning algorithm
that classifies texts on the basis of a measure of distance (e.g. cosine) between them. The
algorithm classifies each text by looking at k of its nearest neighbours and then assigning it to
the most common category represented by those neighbours. If no class is associated to a
majority of neighbours, the text is assigned to the category represented by texts with higher
cosine similarity. We arbitrarily used a low k value (i.e. k=5) as we expect that noise from the
semantic space will be reduced by means of LSA. A common criticism of kNN is that, since
it doesn't make any generalizations from the data, it is prone to overfitting. We assume that
this criticism should not apply completely to vector spaces generated by means of LSA as the
SVD dimensional reduction smoothes and therefore reduces the effect of over fitting that is
usually present in kNN based classification.
5.2.4.2.4 Training set kNN results
Experimental results shown in Figure 6 demonstrate that for some topics and grades using
maximal phrases as units of analysis can improve the performance of LSA. The x axis
corresponds with the different LSA and k-NN configurations implemented for training
LTfLL -2008-212578
31
D4.1 Positioning Design
purposes e.g. D50 indicates that 50 LSA dimensions where used, w or p indicates if phrases
or words where used and the last value corresponds to the number of neighbours used in the
kNN algorithm. The y axis indicates the fraction of correctly classified documents for each of
the classes (grades or topics). In fact, the best results for two of the four grades (e.g. excellent
and poor) were yielded by vector spaces build from phrases. Different results are produced by
kNN classification for topics where spaces built from bags of words produced the best results
clearly for at least four of the six topics representative of their class.
Figure
6: Training set results for kNN topics and grades categorisation
For particular grades and topics, the phrase based LSA (i.e. using vector spaces built from
phrases occurring at least one time as maximal) appears to improve over LSA results that
have been obtained with the traditional bags of words approach. These results are
encouraging and therefore we plan to test alternative semantic space configurations in
particular using more distinctive phrases (e.g. all maximal, left maximal and right maximal).
We expect that as we collect a larger text sample to build the training set, we will be able to
afford the use of those phrases without facing unmanageable levels of sparseness detrimental
to results already obtained.
5.2.5 Conclusion
Our results at this point are rather tentative. They seem to show that the bag-of-phrases
approach may work better than the bag-of-words approach. But there are still many
LTfLL -2008-212578
32
D4.1 Positioning Design
parameters that can be experimented with. Therefore more definitive results may be
forthcoming. In particular, different approaches for measuring the distinctivness need to be
further investigated, both for accurate text classification and for qualitative feedback for the
user. Moreover both the quantitative and qualitative feedback need to be carefully validated
to confirm that this is an effective service for practical use.
5.3 Knowledge rich approaches for positioning the learner
The learners within some community of practice can position themselves with the help of
various knowledge-rich resources, viewed as target competence providers. These resources
might be shared among community members and/or recommended by a community member,
or even developed by them. For example, it might be a domain ontology, a terminology
lexicon, curriculum and learner’s profile, annotated with concepts from the ontology. As
stated in (Wenger 2001:39): “Communities of practice do produce and share documents and
other knowledge artifacts, which can be put in electronic form, and which they need to
manage effectively.”
Concerning the Stahl cycle, this subtask can be placed predominantly in the space of
cognitive artifacts (upper left part of the table), because it uses knowledge rich resources and
methods to do the positioning. Thus, on the one hand, it complements knowledge poor
subtask, which is more collaboratively oriented. On the other hand, it can be viewed as a
transition to WP4.2, because it uses some kind of reference model for comparison. In our
case, this is a domain ontology which is adpated to the needs of our particular stakeholders.
As Bowker and Star (1999) have shown, ontologies are not absolute but are rather dependent
on the world view of particular comnunities. As such world views change, ontologies need to
adapt. One currently popular approach for achieving a continuously adapting ontologies is to
use folksonomies, as is being investigated in WP6.2. But since folksonomies are
democratically determined by users of many different backgrounds and education levels, they
are less appropriate for the positioning task.
In order to avoid terminological misunderstandings among tasks, we will give a definition for
the term ‘concept’ in our positioning task. It is as follows: formalization of a class of objects,
specifying relations with other objects and their properties. The concepts are defined in a
formal ontology. Additionally, the ontology contains definitions of relations and some
instances.
The term general ontology might mean either 1. upper ontology, or 2. linguistic ontology, or
3. how to handle prototype conceptualizations in various environments (groups, cultures,
etc.). Thus, in order to achieve completeness, our answers cover these three possibilities:
1. Upper ontology is the upper part of any ontology. The relation is that the domain
ontology inherits information from this upper part. The upper part also supports the
reasoning and consistency of the domain ontology.
2. Linguistic ontology is any lexicon, organized with respect to some taxonomical
structure. These types (e.g. wordnets) are considered as thesauri by (Guarino 2000).
Their relation to the light ontologies, at which we aim, are as follows: they provide
the lexicalizations of the concepts, presented in the ontology. Also, they support the
connection between the domain ontology and the upper one, i.e. they cover a very
LTfLL -2008-212578
33
D4.1 Positioning Design
important part – the middle one. This part is a bit more abstract that the domain
concepts, but less abstract than the upper ones. In this way, it is more convenient to
work with, when connecting lexica to ontology, and when extending ontologies.
3. Prototype conceptualizations can vary among various groups of interests, age,
nationality, etc. However, we aim at covering the most important concepts in a certain
domain. Partly, we handle the interpretation problem with the lexicons, mapped to the
ontology. Via a filtering mechanism, the stakeholder can choose whether to use the
expert terminology, or the more common lexicalisations.
Knowledge rich approaches are considered in two ways: (1) a separate module for learner
positioning; and (2) a module which provides information for the LSA-based methods to use
concepts instead of terms for positioning. Knowledge-rich approaches provide a more robust
way for explication of learner and curriculum competence, because concepts within
ontologies integrate all the term lexicalisations in some domain. Thus, this method is
supplementary to LSA, which recognizes mentioning of concepts in texts that are not in the
ontology. In our work on the positioning of the learner we rely on the ideas reported in (Kalz
et al., 2007). They discuss the notion of learning networks, considered within the task as a
community of practice. According to this notion, the learner’s competence can be
automatically compared to a set of concept evidences of the target competence. Our goal is to
achieve an ontology-based positioning where the learner competence is represented by a
learner competence ontology and a curriculum competence ontology. However, reliable
competence ontologies are still missing. Thus, in our work we will rely on domain ontologies
which reflect the knowledge part of the learner’s competence. The ontological analyses of the
learner’s profile and the textual description of the relevant curriculum will be an
approximation of learner’s competence and curriculum competence. The domain ontology
and/or curriculum might be provided and/or recommended by an expert within the
community of practice. Learner profiles (CVs, interests, suggestions, opinions, comments)
are shared within the community. Thus, the expert(s) (designated as such within the group)
might help the others with the annotation of profiles and comparison with the curriculum via
an ontology. This help might be of different types: the expert has enough expertise to do it
himself, or he can contact the appropriate people outside to do that for the community. The
curriculum might also be developed by the community. In order to introduce some first steps
in the evaluation of the learner’s knowledge degree, we will evaluate the usage of this
knowledge represented within the profile. This evaluation will be done via the techniques of
the sentiment analysis. After analysing the curriculum description and the profile, we will use
several approaches to compare the extracted conceptual information. In the rest of this
section, we discuss the envisaged analyses in the process of profile analysis.
Knowledge rich methods rely on analysis of the text by using knowledge sources outside of
the text, such as lexicons, ontologies, grammars. For this reason, this task is placed in the
cognitive part within the Stahl learning model. In the case of profile analysis, the result from
it is used as an evidence of the learner competence and knowledge in the domain. Within the
framework of work package 4 we consider several types of text analysis (also their
applications to learning tasks, and more specifically the positioning of learner with respect to
a curriculum description). These text analyses include: (1) ontology-based semantic
annotation, (2) discourse segmentation, (3) lexical chains approach to disambiguation of
concept annotation and (4) sentiment analysis for evaluation of the concept usage in the text.
These facilities are envisaged to serve communities of practices in a specific domain of
LTfLL -2008-212578
34
D4.1 Positioning Design
interest. The combination of the above mentioned analyses has to explicate the conceptual
content of the profile which is to be used for positioning of the learner. The ontology-based
semantic annotation relates the text of a profile to the conceptual information in the ontology.
The discourse segmentation facilitates the creation of lexical chains and the sentiment
analysis. The lexical chains identification supports the disambiguation of the ambiguous
terms and phrases within the text. The sentiment analysis determines the attitude of the
learner to the concepts explicated within the profile. At the end, the result of the whole set of
analyses presents a classification of the concept usages in the text as known and unknown.
This classification will be used for the comparison to a conceptual representation of a
curriculum via an ontology. Here are some recent references to relevant works. Galley and
McKeown (2003) present the idea of the lexical chains, which trace the cohesion within the
texts. The automatic establishment of lexical chains is outlined in Mihalcea (2007). Wolf and
Gibson (2006) propose 11 coherence relations, which successfully segment the discourse
within the texts. According to Schauer (2000), 15 to 20 percent of coherence relations are
signalled by some kind of conjunction but not all of them are unambiguous.
The various levels of sentiment analysis scope are described in Moilanen and Pulman (2007)
and Liu (2008), among others. It is often emphasized that adding knowledge rich features
improves the results in sentiment analysis. For example – Moilanen and Pulman (2007),
Kennedy and Inkpen (2006), Kim and Hovy (2006).
As mentioned above, knowledge rich approaches are usually connected with the availability
and the usage of knowledge rich data bases, such as ontologies and lexicons. The ontologies
reflect the conceptualizations in some domain of interest. For example, the DAML ontology
library, SWOOGLE, or the LT4eL ontology. These ontologies have to be connected to an
upper ontology in order to cover in better way the general knowledge. For example, DOLCE,
SUMO, SIMPLE. The most famous knowledge rich lexicons are the so-called wordnets
(WordNet, EuroWordNet, BalkaNet, SIMPLE). Such resources are exploited for semantic
annotation of documents and/or for semantic retrieval. For better semantic annotation and its
usage in positioning of the learning task, we consider discourse segmentation and sentiment
analysis methods as relevant. Once available, the resources might be changed with respect to
community’s needs by the community itself or by the providers. The sentiment analysis might
be done within the community, using only its members’ opinions, or in general – using other
people’s opinion outside community.
Within the LT4eL project, an ontology-to-text relation was defined (Simov and Osenova
2007; Simov and Osenova, 2008). We briefly present this relation here. We assume that the
ontology is the repository of the lexical meaning of the language. Thus, we have started with
a concept in the ontology and we searched for lexical items and non-lexical phrases that
convey the content of the concept. There are two possible problems here: (1) there is no
lexical item for some of the concepts in the ontology, and (2) there are lexical items in the
language without a concept representing the meaning of the lexical item in the ontology. The
first problem is overcome by allowing in the lexicon also non-lexical (fully compositional)
phrases to be represented. The second problem is solved by extension of the ontology. The
lexical items are then mapped to grammars. We call them concept annotation grammars.
These grammars relate the lexicon to the text. Such a mapping is necessary, as many lexical
items and phrases from the lexicons allow for multiple realizations in the text. Thus, they
require some additional linguistic knowledge in order to disambiguate between different
meanings of some lexical item or phrase. Figure 7 depicts the elements of the model.
LTfLL -2008-212578
35
D4.1 Positioning Design
Figure 7: Ontology–To-Text relation to be exploited in the knowledge rich approach to learner positioning
We have been using the relations between the different elements for the task of ontologybased search. The connection from ontology via lexicon to grammars is relied upon for the
concept annotation of the text. In this way we have established a connection between the
ontology and the texts. The relation between the lexicon and the ontology is used for
definition of user queries with respect to the appropriate segments within the documents.
Another direction of the knowledge rich methods is discourse analysis. As a benchmark, the
work of Wolf and Gibson (2006) can be considered. The authors present 10 coherence
relations. The advantage of their work is that they succeed in collapsing the large number of
possible relations into a small set of operational relations.
Sentiment (opinion) analysis can be specified broadly as a kind of analysis that aims to
determine automatically the attitude (sentiment, tone, polarity) of a speaker/writer with
respect to a certain topic. Usually this kind of analysis is opposed to the standard fact-based
analysis and at the same time it is rendered as a classification task. It is commonplace to
evaluate the sentiment of an opinion on a two-value scale: negative or positive. When free of
subjectivity, the text is regarded as neutral.
Much work is done for detecting negative or positive judgment, but sentiment analysis is not
only about the more general polarity of an opinion, it is also about identifying the opinion
LTfLL -2008-212578
36
D4.1 Positioning Design
holder, the object (the topic) that have been evaluated, the type of propositional attitude
expressed (belief – think, believe, assume, emotion – hate, adore, etc.), and the strength of the
polarity. Of course, more features can be added.
The type of texts subject to this kind of information retrieval are numerous, just to give some
examples: customer reviews on all kinds of products, brands, reviews on cultural events,
opinion polls. Here are some freely downloadable corpora with different domain-specific
texts: (1) MPQA Opinion Corpus, containing 535 news articles collected during the 2002
NRRC Workshop on Multi-Perspective Question Answering (Wiebe, 2002); (2) polarity
dataset, containing 1000 positive and 1000 negative movie reviews (Pang and Lee, 2004).
Sentiment analysis is concerned with two levels of granularity: sentence and document level,
the second estimated as too coarse for most applications (Liu, 2008). According to Moilanen
and Pulman (2007) though, while sentiment classifiers work well with a large input (e.g. a
750-word movie review), the results for sentential and subsentiential units – clauses or noun
phrases, are not satisfying. Taking into account linguistic features, such as valence shifters
(for example negation) intensifiers, gradable adjectives, patterns, semantic role labelling and
syntactic structure, adding a level of compositionality do improve the analysis in terms of
accuracy (Moilanen and Pulman, 2007; Kennedy and Inkpen, 2006; Kim and Hovy, 2006,
amongst others).
The prevailing majority of techniques use some form of machine learning, supervised and
semi-supervised. The most common and basic approach to sentiment classification is
keyword-based, starting from a list of sentiment indicators or clues prepared manually, semiautomatically (relying on WordNet of FrameNet) or acquired by machine-learning (Rimon,
2005). Other types of elements used in different algorithms are: Semantic Orientation (SO) –
Pointwise Mutual Information (PMI), support vector machines, maximum entropy, naïve
Bayes, latent semantic analysis (LSA) and so on (for a detailed overview see Liu, 2008).
In this knowledge rich approach for learner positioning we will rely on the reported works.
We will integrate the above technologies into a common processing module in order to
explicate the conceptual content of the profile and the curriculum description. The explication
of the conceptual content will be done via annotation of the text part of the profile and the
curriculum description. This annotation could be used as input for different tasks. Firstly, the
concept annotation will be used to find the position of the learner with respect to the
curriculum and to select appropriate learning materials to cover the gaps discovered by the
method (see below for more details). Secondly, the concept annotation within the text could
be used as an input for LSA methods. Concepts could substitute the terms within the vector
space.
To evaluate the performance of this knowledge rich approach for positioning we plan to use
comparable data sets and validation method as the ones proposed for the LSA based
approach.
5.3.1 Extending knowledge rich approach tools
As mentioned above, additionally to LSA based functionalities, the service will provide
knowledge rich profile positioning functionalities. For that purpose we plan to extend the
CLaRK system (Simov et al., 2001) originally implemented with the aim of minimizing
LTfLL -2008-212578
37
D4.1 Positioning Design
human work during the process of corpora creation. CLaRK will provide service’s
functionalities for calling external programs when they are necessary for some specific task.
CLaRK is implemented in Java and the necessary functional interface will be provided.
The knowledge rich methods which are envisaged to be implemented by the profile analysis
and positioning service are as follows: (1) semantic annotation of the profile, (2) lexical
chains approach to disambiguation of concept annotation (3) discourse annotation of these
texts; and (4) sentiment analysis of the discourse segments as well as the mentioned concepts
with respect to the levels of learner’s concept competence. They will be combined in a
common procedure. The result of the knowledge rich analysis of an profile will be a concept
evidence of the learner’s competence expressed in the profile. The elements of the concept
evidence of the learner’s competence will be a set of concept descriptions extracted from the
profile with links back to the text of the profile. In this way the concept evidence of the
learner’s competence can be automatically compared to a set of concept evidences of the
target competence (learning network in the terms of Kalz et al. 2007). Those will be selected
that are not covered by the current learner’s competence. For the comparison of the concept
evidences we will use the standard vector metrics from Information Retrieval community.
The links to the profile will support the assessors of the student competence to find out the
reason for the inclusion of a concept description in the concept evidence of the learner’s
competence. The content analysis which is meant to be implemented for this task will allow
us to use the methodology for positioning of learners presented in Kalz et al. (2007). Concept
descriptions used for the semantic annotation and for the representation of concept evidences
are taken from the domain ontology. Recall that here we consider only an approximation of
the learner’s competence based on the concepts from a domain ontology and their usage in
the profile. Much more work will be necessary in order to support a full representation of the
learner’s competence. The same applies to the target competence encoded in the curriculum
description.
The semantic annotation and the discourse annotation will be used also in work package 6
(WP6). The difference will be in the domain of application and the specific type of text which
will be analysed here, namely the profile document.
5.3.1.1 Semantic annotation
In order to use the LT4eL model for the analysis of the profile we will implement the
ontology-to-text relation for the new domain (medicine) with a new vocabulary. We will
extend the previous implementation with new disambiguation functionality which will be
based on lexical chains (Galley and McKeown, 2003), using semantic annotation of general
words in the text (in addition to the domain specific terms) and discourse annotation. For the
semantic annotation of the general words we will use OntoWordNet (Gangemi et al., 2003),
which is already aligned to the same upper ontology which will be used in the construction of
the domain ontology. The output of this new functionality will be a semantically annotated
text of the profile. Each domain term will be annotated with a concept from the domain
ontology and each general word will be annotated with concepts from the upper ontology.
5.3.1.2 Discourse annotation
Similarly to the task within WP6, our main goals in developing an additional layer of
discourse segmentation and relations annotation are: (1) to investigate the possibility for
refining concept recognition and sense disambiguation for targeted words (lexical terms) via
coherence relations (discourse relations, rhetorical relations) markup; (2) the discourse
LTfLL -2008-212578
38
D4.1 Positioning Design
annotation will be used for the sentiment analysis in order to evaluate the learner’s attitude to
the concepts mentioned within the profile. The input for this functionality will be the results
from the previous functionality. The discourse annotation process consists of several steps,
which may be iterated.
1. Creation of coherence relations taxonomy. We will start with the set of relations and the
coding scheme defined by Wolf and Gibson (2006). Their taxonomy is based on the Hobbs
list (Hobbs, 1985), and is more coarse-grained than others that include up to 400 types of
relations. This fact makes it really applicable for our task. It consists of eleven types of
coherence relations:
Temporal sequence: When one discourse segment describes an event that takes place
before another event, expressed in another discourse segment.
Cause-effect: When one discourse segment describes the cause, and another – the
effect for a given event.
Condition: When one discourse segment describes a possible event that will occur
only if another event, described in another discourse segment, also occurs.
Elaboration: When one discourse segment elaborates, i.e. gives more detailed
information about another discourse segment.
Example: When a discourse segment provides examples for another discourse
segment.
Similarity: When the event, expressed in one discourse segment, is similar to an
event, expressed in another discourse segment.
Contrast: When the event, expressed in one discourse segment, contrasts an event,
expressed in another discourse segment.
Generalization: When one discourse segment states a generalization for the content
of another discourse segment.
Violated expectation: When there is an absence of a causal relation between two
discourse segments.
Attribution: When one discourse segment states the source for the content of another
discourse segment. It is usually used in constructions, such as: John said that…
Same-segment: Same-segment is a structural type of relation, because it holds
between disconnected parts of one discourse segment (subject NP separated from its
predicate). Same-segment, similarity and contrast relations are symmetrical while the
rest are asymmetrical (directed), that is – one of the segments is more important (the
nucleus) than the other (satellite).
According to the coding scheme the three general steps of the annotation process are: (1) the
output of the sentence-splitter is segmented further into clauses and then, if needed,
LTfLL -2008-212578
39
D4.1 Positioning Design
annotators insert intrasentential boundaries for smaller discourse segments; (2) the discourse
segments are grouped thematically and (3) the coherence relations between the segments are
indicated. After the annotation process is finished, the taxonomy may be further adjusted to
improve the descriptive adequacy for the texts in the Computer Science domain.
2. Manual annotation of portfolios as a gold standard for the creation of automatic procedure
for discourse segmentation.
3. The analysis of the obtained discourse structures will provide information that could be
used for (1) the development of constraints over the semantic annotation grammar, (2)
supporting anaphora resolution, and (3) support of sentiment analysis. In addition, we
consider the possibility of creating a rule-based grammar for recognizing the coherence
relations that are unambiguously linguistically marked.
In order to improve the concept annotation, we will test different knowledge-based
techniques that are common for word sense disambiguation. Our main goal is the enrichment
of the concept annotation grammar in order to map the relations between text chunks,
recognized as carriers of the concepts, with relations present in the domain ontology: is-a,
part-of, etc. In the future different algorithms for automatic establishment of lexical chains
(with nouns) may be tested (for an overview see Mihalcea, 2007). Lexical chains and
rhetorical relations, the two types of discourse information, contributing to the text
coherence, will be used for improving the concept annotation. For example, a discourse
segment, nucleus in an elaboration relation, will most probably contain a term, connected via
hypernymy relations with lexical units that belong to the satellite segment.
The discourse annotation will be adapted to the format of the profile. The idea is that
elements of the profile will require some specific kind of language. In such cases the
discourse structure might depend on the peculiarities of the corresponding sub-language. The
output of this functionality will be a segmentation of the text of the profile in discourse
elements and annotation of the relations between them.
5.3.2 Sentiment analysis
The input for this functionality will be the results from the previous above described
functionalities. In order to construct a concept evidence of the learner’s competence, we first
need to extract the concepts which are mentioned within the profile. Then, on the base of the
ontological reasoning, the implied concepts will be added. For example, if the profile’s holder
in IT domain says that he/she has some expertise in XSLT, this automatically means: on more
general level, that he/she has also knowledge of XML and some programming language, and,
on more specific level, that he/she can use XML-based language for the transformation of
XML documents into other XML or “human-readable” documents. We also need to know in
what context each of the concepts in the profile was mentioned by the learner. For example,
behind the discourse relation, called contrast the learner stated two opposite facts: it is useful
to know how to transform documents, but a next step is required – to learn also XSL-FO
language in order to handle formatting objects. From this short context a conclusion can be
drawn that the learner’s position with respect to the knowledge of XSL set of languages is
partly completed. Thus, comparing conceptual information and discourse relations, each
mentioning of a concept will be evaluated by one of the values: ‘well known’, ‘known’, and
‘unknown’[4]. We will use the methods developed in the areas of sentiment and opinion
analysis. As it was already mentioned, a pre-defined requirement list of necessary concepts
LTfLL -2008-212578
40
D4.1 Positioning Design
with definitions will be used in order to estimate the degree of competence, delivered by the
learner in the profile. There will be three types of evaluation: coverage, degree of detail and
relevance. The coverage will be estimated over the number of the mentioned relevant
concepts that match the pre-defined list. The degree of detail will be evaluated over the depth
of the conceptual space. And the relevance will be estimated via the ontological relations
from a given concept to the other co-occurring concepts within the discourse segment.
5.3.3 Construction of a concept evidence of the learner’s competence and knowledge
As it was mentioned above, a concept evidence of the learner’s competence is a set of
concept descriptions extracted from the profile. For the moment we divide this set in the
following subsets: (1) known concepts; (2) partially known concepts; (3) unknown concepts;
and (4) concepts with contradictory usages. The first subset will contain all the concepts
which are evaluated as known in the profile. The second subset will contain concepts that are
mentioned in the profile, but for which there is not enough evidence about the level of
knowledge of the learner with respect to them. The third subset will contain concepts that
definitely are mentioned as unknown by the learner. In the last subset we will include the
concepts for which there are positive and negative evidences about the knowledge of the
learner. In addition to the extracted concepts we will extract links to the occurrences of the
concepts in the text. Within the community of practice, the curriculum part has to be defined
against which the positioning to be done. For example, the curriculum might take XSL as a
whole set of languages, in which each language (XSLT, XPath and XSL-FO) has to be
learned. On the other hand, only XPath might be taken as a learning goal.
The output of this functionality will be used further to compare the concept evidence of the
learner’s competence with the community of practice. The comparisons will use a vector
representation of concept evidence of the learner’s competence and concept evidence of the
target competence. The vector for target competence will be fixed within the learner network.
The vector for learner’s competence will be created by the assessor on the basis of the above
sets of concepts.
The evaluation of the method will be done on two levels. First, for each of the processing
steps, we will create manually gold standard corpus on which to test the corresponding
technology using the usual precision and recall metrics. Second, we will test the method with
respect to the performance of the LSA-based method. The aim is not for the methods to
compete, but to find the best ways to combine them in order to satisfy task goals.
We have described the knowledge rich method preferably as a complement to LSA, rather
than an alternative. We envisage also integration of the two methods. First, in the
construction of a vector space instead of terms from the text the concepts from the conceptual
annotation could be used. In this way, one can abstract over the textual representation of the
concepts. For example, very often in text a super-concept term can be used to denote a subconcept – “system” instead of “computer system”. Also with the sentiment analysis we could
select which concepts to be included in the vector space. It is also possible to combine the
two methods via integrating their results.
LTfLL -2008-212578
41
D4.1 Positioning Design
6. Diagnosing conceptual development (task 4.2)
6.1 Outputs of WP4.2
WP4.2's contribution to the integrated environment will be an elaborated scenario series
developed in association with WP3, and services corresponding to the scenario. These
services are described in more detail below and briefly comprise services for diagnosing
conceptual development and functionality for aggregating the raw data output into formats
meaningful to the end user.
6.1.1 The WP4.2 solution scenario, providing formative feedback
We have elaborated a solution scenario (see deliverable D3.2) that depicts the functional
design of a service aiming to provide (semi-)automated formative feedback with the help of
Language Technologies. Using the service, learners could compare evidence of their
knowledge (e.g. text inputs such as essays, blogs, "think alouds" etc.) with reference models
in order to identify possible differences and obtain recommendations of suggested actions to
address the differences. Learners can submit new evidence of their knowledge and receive
formative feedback as often as they want. Moreover, learners can monitor their own learning
process as the service provides also comparisons of the learner’s knowledge evidences
previously submitted.
For tutors, the service will provide a means to monitor the current progress of learners on a
topic, to allow them to take proactive actions to improve learners’ conceptualization of the
topic. This might lower tutor workload. The design considers that the service can be used in
both formal and informal learning settings. Depending how the use of the service is
implemented in the learning context, learners can assume both tutor and learner roles (for
more detail, see WP4.2 Solution Scenario included in D.3.2).
The scenario will be implemented as a set of web services providing (1) learner evidence
collection facilities (data gathering), (2) data extraction and condensation / aggregation
functionality and (3) facilities to compare concepts with the reference models and present the
results to the user.
6.2 Research problems
In undertaking this work, the following research problems arise:
•
•
•
•
•
do potential end users find the scenario realistic?
can we adapt existing concept mapping tools to meet the requirements of the scenario?
with what reference models should the learner's conceptual development be compared?
how should the raw data from the service be aggregated to present meaningful
information to the learner and tutor, to inform their future actions?
what tuning of the language technology services is required to optimise the delivery of
meaningful information?
In the first phase of our work, we undertook a conceptual validation of the showcase scenario
(reported in deliverable D7.2), which showed that learners and tutors do find the scenario
realistic, subject to a number of enhancements and clarifications. It was identified that "the
LTfLL -2008-212578
42
D4.1 Positioning Design
ability for students and tutors to compare concept maps is an under-developed area in the
showcase scenario that was perceived as having the potential to add substantial value to the
service. This feature of the service needs to be developed further to provide a clear picture of
the comparisons that students and tutors will be able to make between concept maps".
This deliverable reports our work with respect to the second and third problems. To address
the second problem, we have compared a number of concept mapping tools. The work is
reported in Section 6.4
With regard to the third problem, we investigate three types of reference model, against
which to compare learner data. The work is reported in Section 6.5.
In later phases, we will address the underlying language technology questions, in cycles of
software development and validation with end users. In view of the importance of the user
interface being meaningful, more frequent validations with end users (learners, tutors) are
indicated.
6.3 Introduction to studies
In our discussion of the theoretical basis of learning, we have indicated the importance of the
inter relationship between the individual learner and communities of practice and the
interaction between building "personal knowing" and "collaborative knowing". We have also
indicated that in order for learners to develop expertise in their specific domain, it is essential
that they recognise the limitations of their understanding and conceptual development and
develop appropriate learning plans, as demonstrated in reflective learning cycles. The
theoretical basis for understanding development of expertise in professional domains
identifies knowledge creation and restructuring as essential components of this process. The
Stahl model and the notions of Communities of Practice indicate that learning from peers has
a key role in enabling individual learners to reach a shared understanding of specific aspects
of their domain. The use of Problem Based Learning, which we indicate is used in
medical education to model aspects of Communities of Practice, is an example of an
educational approach in which peers reach a consensus view of a specific issue, topic or
concept.
In order to provide individual learners with the guidance and "instruction" to enable their
development of expertise, we require reference models. Within this educational context, we
have defined three types, against which learners can compare their understanding of a
specific topic. These are:
•
•
•
Archetypical reference model: based on expert and state-of the art information (e.g.
scientific literature).
Pre-defined reference model (or ‘Theoretical reference model’): considers specific
information based on the curriculum (e.g. course material, tutor notes, relevant reading
materials, etc.).
Emerging group model: considers the concepts and the relations between those concepts
that a group of people (e.g. peers, participants, co-workers, etc.) used most often.
We have concentrated on the pre-defined and emerging approaches to identify or approximate
the conceptual development of learners to underpin the role of Language Technology tools.
Next, we explain how existing applications and tools, namely Leximancer (Smith &
LTfLL -2008-212578
43
D4.1 Positioning Design
Humphreys, 2006) and Pathfinder (Schvaneveldt, 1990), have been used in two different
preliminary explorations as proof of concept of the suitability of these approaches.
In order to assess the individual’s knowledge of a particular domain Goldsmith et al. (1991)
proposed a structural approach to determine how the individual organizes the concepts of
such a domain. This approach involves three steps: knowledge elicitation, knowledge
representation, and evaluation of the representation.
Knowledge elicitation techniques measure the learner’s understanding of the relationships
among a set of concepts (Jonassen et al., 1993a). Methods that support this activity include
card sorting, concept maps, think aloud, or essay questions.
Knowledge representation reflects the underlying organization of the elicited knowledge
(Goldsmith et al., 1991). Advanced statistical methods (e.g. cluster analysis, tree
constructions, dimensional representations, pathfinder nets) are used to identify the structural
framework underlying the set of domain concepts.
Evaluation of the representation relative to some standard (e.g. expert’s organization of the
concepts in the domain) uses one of the following approaches (Goldsmith et al., 1991):
qualitative assessment of derived representations; quantifying the similarities between a
student representation and a derived structure of the content of the domain; or comparing the
cognitive structures of experts and novices.
6.4 Comparison of existing concept mapping tools - which tools to use?
A decision was made to start the exploration with the cognitive map method, which is one of
the most common methods for representing cognitive structures, as a mean to elicit and
represent learner knowledge. The decision was taken on the basis of the appropriateness of
concept maps for representing learners’ representations of subject matter structure and on
research evidence that demonstrates the concept map method is well suited for eliciting
knowledge (Nesbit and Adesope, 2006), and is a better method for evaluating meaningful
learning of learners of different ages than classical assessment methods such as tests and
essays (Jonassen et al., 1997; Novak, 1998). It is important to point out, however, that the
creation of concept maps is a complex and time consuming task that requires training and
practice to understand how the relevant concepts should be identified and how to make the
relation between them.
There are already a number of tools for automatic construction and support of the
construction of concept maps: Knowledge Network and Orientation (KNOT, PFNET)
(Clariana et al., 2006); Surface, Matching and Deep Structure (SMD) (Ifenhaler and Seel,
2005); Model Inspection Trace of Concepts and Relations (MITOCAR) (Pirnay-Dummer,
2006 ); Dynamic Evaluation of Enhanced Problem Solving (DEEP) (Spector and Koszalka,
2004); jMap (Jeong, 2008), and ProDaX (Oberholzer et al., 2008). Table 2 depicts these tools
in terms of the data collection they use, the analysis they perform, the data conversion they
use and the comparison they perform.
LTfLL -2008-212578
44
D4.1 Positioning Design
Tool
KNOT
SMD
Data Collection
Analysis
Concept
Quantitative Analysis
pairs/Propositions
Data
Conversion
Pathfinder
Networks
Concept map or Quantitative—analysis Structural
natural language is calculated using
decomposition
tools.
into 3
categories
(manual and
semiautomatic)
Comparison(s)
Direct comparison of
networks with some
statistical results.
Unlimited comparison
MITOCAR Natural language Quantitative—analysis Structural
Paired comparisons
included multiple
composition for semantic and
calculations using tools into 1 category structural
(automatic)
model distance
measure
DEEP
Annotated causal Quantitative/qualitative Structure
maps
—analysis is done
decomposition
into 3
categories
mostly by hand
(automatic)
jMap
Concept maps,
causal maps, or
belief networks
Quantitative analysis – Structural
Superimposes maps of
analysis is calculated decomposition individual (n=1) and
using tools
into link
group of learners (n =
strengths
2+) over a specified
between causal target map
factors and
evidentiary
strength
ProDaX
Association Data,
Cross-Tables,
Two-Way TwoMode Data,
Coordinates,
Scales
Non-Metric
Multidimensional
Scaling/ClusterAnalysis
Unlimited
comparisons, showing
details relative to
concepts
Concept Maps Comparison of maps
based on Procrustean
Transformation/Lossoriented Meta Map
(LOMM)
Table 2: Overview of concept mapping tools (adapted from Shute et al., submitted)
These tools have some common characteristics: (a) they are concerned with conceptual
development of learners; (b) they can (semi-)automatically construct concept maps from a
LTfLL -2008-212578
45
D4.1 Positioning Design
text; (c) they use a sort of distance matrices; (e) they propose a quantitative analysis of the
maps; and (d) most of them pretend to support high levels of learning, namely critical
thinking and problem solving.
Amongst their differences we have found that, even though they all use some sort of
Language Technology analysis, not all of them refer to it explicitly. The SMD and jMap can
use as an input not only text but also concept maps. These tools also differ in the scoring
schemas they use to perform the quantitative analysis: DEEP uses a number of nodes and
links; SMD uses propositions or a number of the links of the shortest path between the most
distant nodes.
Most of the referred concept mapping tools provide opportunities to identify the conceptual
gap between a learner’s concept map and a criterion map (in fact, an expert map), or to
compare a learner’s concept maps in different periods of time. However, only SMD, jMap
and, in some extent DEEP, provide purposely a visualisation of this progression towards the
criterion. Most of these mapping approaches construct and analyse individual maps. jMap
visualises and assesses changes observed in either individual or collective maps.
Nevertheless, jMap is the only tool restricted to producing a particular type of maps, causal
maps.
KNOT, SMD and MITOCAR do report on reliability and validity criteria correlating, as a
typical case, the automatic scores generated by these concept mapping approaches and human
concept mapping scores and human essay scores. Finally, it is worth mentioning that SMD
and MITOCAR report experimental data on the effectiveness of a particular technique as an
increase in similarity between a learner’s map and an expert’s map.
6.4.1 Methods
A first exploration of existing tools that create concept maps from an input test was
performed. The aim was to investigate in which and to which extent existing tools support the
process from knowledge elicitation to evaluation. In particular, the aim was to gain insight in
how flexible and easy to use the tools are (other aspects such as reliability and validity having
been derived from literature).
The following tools have been explored:
•
•
•
CMAP (Institute for Human and Machine Cognition, 2004)
KNOT and Ala-reader (Clariana et al., 2006)
INFOMAP (Peters, 2005)
KNOT is a software tool that generates text proposition files that can be imported from
CMAP to generate concept maps automatically. The conclusion from the initial exploration
was that the process could be used to analyze conceptual development but there are
restrictions on the data that can be used, for instance, the general limit of concept pairs that
can be used. Next, an initial exploration of INFOMAP was performed, to generate an
associative semantic network, based on learner texts. INFOMAP employs a similar approach
to Latent Semantic Analysis, with a focus on word-to-word relations and a context limitation
around the words used for indexing.
To this end, a data set from a Psychology course at the OUNL was used. This data offers
course content, which was considered as the expert level of argumentation. Also documents
LTfLL -2008-212578
46
D4.1 Positioning Design
written by the learners were used. For every part of the course content some related keywords
were generated, then the inter-correlation of the keywords were calculated for one exemplary
chapter. After that, by using a clustering method (nearest neighbour approach), a distance
matrix and clusters in the keywords were generated. Figure 8 shows an exemplary cluster
overview of a chapter. These keyword clusters can be used to identify topical foci of the
documents. An alternative approach to use the keywords and associated other concepts in
documents is Multidimensional Scaling (MDS). With this approach distances between
concepts can be visualized (see Figure 9).
Figure 8: Example of clustering
LTfLL -2008-212578
47
D4.1 Positioning Design
Figure 9: Example of a Euclidian Distance Model derived from Multidimensional Scaling
6.4.2 Results
Clearly Infomap was able to generate a visual representation of the relationship between
concepts from the Psychology course texts at OUNL. The use of Multidimensional scaling,
however, restricted the use of words associated these concepts and it was therefore decided to
investigate tools which non metric Multi dimensional scaling, which allowed the
incorporation of a wider variety of language.
Leximancer and Pathfinder were selected for a further proof of concept. Leximancer
generates concept maps from a document collection using content analysis (based on cooccurrence) and relational analysis (proximity and concept mapping). These maps, or visual
representations, show the concepts identified in the text and the relations between them.
Pathfinder can be used to derive and visualize structured (semantic) networks. It is based on
proximity measures (similarity, correlations, distances, probability) between pairs of concepts
(Clariana et al., 2006).
LTfLL -2008-212578
48
D4.1 Positioning Design
6.5 Studies on potential reference models using Leximancer and Pathfinder
6.5.1 Pre–defined Reference Model
6.5.1.1 Methods
A data collection protocol was defined to elicit and represent a learner’s knowledge. This
protocol combines a think aloud procedure with a cognitive map method to provide a suitable
and appropriate measure of the learner’s representation of the subject matter structure. As a
proof of concept these tools have been explored to in two different ways to generate (a) the
pre-defined reference model and (b) an emerging group model (see Berlanga et al., 2009 for
details).
For the generation of a pre-defined reference model, a combination of Leximancer and
Pathfinder was used. A small randomly selected group of Year 2 undergraduate medical
students (N = 12) recorded their summaries of a specific PBL case and were asked to speak
for 5 minutes on the Bioscience mechanisms which formed the basis of the condition and
how they might treat the conditions, using this knowledge. The recording were made under
standard conditions in the University of Manchester Medical School, transcribed and used as
the text for Leximancer (Figure 10). Text from tutor notes and supporting materials.were used
to generate the predefined reference model. Pathfinder was used to identify similarities and
differences between results from learners and those form the pre-defined model.
Figure 10: Part of transcribed student think aloud
6.5.1.2 Results
The concept maps from the students and the pre-defined reference model differ in the level of
detail (see Figure 11) . Whereas the student concept map included detailed concepts, the predefined reference model encapsulated the concepts and gave the panoramic view of the
knowledge. Furthermore, the student map can be characterized as the description of a disease
process, while the pre-defined reference model is at the (auto)immune system level. Finally,
the latter includes both a diagnostic part, and more signs and symptoms.
6.5.1.3 Conclusions
These results suggest that even if the learning material explains the reasons and conditions of
a problem (“the why”), novice students represent their understanding by indicating only
procedural knowledge, mentioning how to solve a problem (“the how”). This might imply
that the tutor notes and learning materials might not be ideal to generate a pre-defined
reference model. The materials are written from a perspective that requires more expertise
than the novice student can achieve at that point of time. Consequently, this might not be a
LTfLL -2008-212578
49
D4.1 Positioning Design
good basis for deriving the pre-defined reference model, nor for providing formative
feedback.
Figure 11: Concept map for a learner (left) and the pre-defined reference model (right) (Leximancer)
6.5.2 Emerging Group Model
6.5.2.1. Methods
Only Leximancer was used in these experiments, in which OUNL employees in an informal
learning situation were the subjects. They used a similar “think aloud”protocol to that used in
the previous experiments, except that on that occasion the subject matter was “Learning
Networks”
6.5.2.2 Results
The results indicated the ten most used concepts and their relevance automatically, as well as
the relations of each concept with other concepts. Figure 12 depicts the so-called emerging
reference model for the concept Learning Networks as it arises from all concepts and the
relations between concepts. It also visualizes the position of the individual learners in relation
to the model, by indicating which concepts the speaker mentioned.
Further, a concept map was generated for individual employees for whom the ten most used
concepts were identified. These were compared to identify similarities and differences
between the emerging reference model and employees’ concept maps. It seems feasible to
generate individual formative feedback reports that present differences and similarities.
Future work involves validation of the reliability of the emerging reference model and the
formative feedback report.
LTfLL -2008-212578
50
D4.1 Positioning Design
Figure 12: Example of an emerging reference model (Leximancer)
6.6 Conclusions and next steps
In summary, the pre-defined reference model approach seems to provide little information to
generate a formative feedback report, since it contained information that might be at a “too
high level” for a learner at a specific point of time. In fact this is in line with what we have
argued before regarding theories of expertise. It could be the case that, at a specific point of
time, learners do not have the expertise level described in the pre-defined reference model,
which will consider the ultimate learning goal but not the different levels of expertise a
learner will go through. The emerging reference model approach seems to solve this issue.
The set of concepts that is used by most people at a specific point in time might provide
better evidence of the level of abstraction and relations between concepts. This approach
might provide better guidance as in resembles the Zone of Proximal Development (Vygotsky,
1978) of the learner and it could be also seen as a way to build socially a shared
understanding of a concept, a unit of understanding that is shared by a particular group and
context. The approach, however, will require a better appreciation of the learner’s knowledge
representation – by contextualizing both the learner’s knowledge and the situation in which
the knowledge will be applied – and requires mechanisms to keep the model updated.
This work informed the development of the WP4.2 scenario. The scenario now makes
provision for two reference models (pre-defined model and emerging group model).
Undoubtedly, more research is needed to establish how learners would benefit the most from
comparing their conceptual development with these models: whether it is good strategy for
learners to see comparisons with both models or, whether, depending on their level of
expertise, comparisons with different models will be made available. The type of reference
model used may depend on the level of learner development. The emerging reference model,
which is based on concepts and their inter relationships, generated by peers, would most
likely be of use for an individual learner at a novice level, as at this stage it would
correspond to his/her Zone of Proximal Development (Vygotsky, 1978). As expertise
LTfLL -2008-212578
51
D4.1 Positioning Design
develops, the emerging reference model may still be appropriate, depending on the
development stage of the practice group as a whole, but pre-defined reference model may be
more suited to a more advanced learner.
6.7 Design of the WP4.2 service
While Leximancer and Pathfinder were used in the showcase, they were not considered
suitable as a basis for the LTfLL conceptual development service (see deliverable D7.2).
Leximancer does provide all the required functionality; however it is a proprietary application
and cannot be customized to the requirements of the scenario. Pathfinder also does not
provide the required functionality and difficulties were experienced in using it alongside
other applications. A decision was made to develop a custom tool based on LSA.
Regarding future work on developing the WP4.2 service, there are three process steps that are
needed: (i) data gathering and evidence collection; (ii) extraction of structure and
condensation; and (iii) comparison of the conceptual structures (Figure 13).
Figure 13: Process steps for monitoring conceptual development
First, data gathering and capturing techniques need to provide functionality for easy
collection of evidence contained in learning texts using the various tools provided by and
beyond the project. Such texts include texts created by the learner, such as study notes,
summaries, reviews, discussion articles etc. Further possible evidence sources could include
short texts provided in chats, fora, comments, etc. A simple evidence production tool such as
a learning diary (blog) as outlined in the section "Texts available for analysis" is also an
important source of material from individual learners
The second challenge is the development of web services for the extraction of structure into
a condensed, meaningful representation reflecting the conceptual information in the learner
evidence. Web services will extract the input texts into the desired representational structure,
thereby unveilling its conceptual structure. The representation format chosen so far are
graphs of connected terms produced by LSA (similar to the concept maps produced by
Leximancer, which was shown in the initial studies).
The processing chain is provided as a set of modular services that can be flexibly configured
in order to sanitise, tokenise, relate, and aggregate the data from raw input to the conceptual
representation at the output. The analytical part of the service condenses the raw information
such that the learner gets an overview first and details on demand.
The way in which the data is aggregated is important in determining a meaningful output and
the design of the data aggregation will require careful thought. The service will output, for
example, selected terms and relations extracted from the text, similarities and differences
between the individual terms and the pre-defined reference model, etc.
LTfLL -2008-212578
52
D4.1 Positioning Design
In the third step, the service will provide a comparison of the conceptual structures
between the evidence of one learner at a certain point in time and that of another learner, the
same person at a different point in time, a pre-defined reference model or a emergent model.
This third step faces the challenge of translating the structural representation of the second
step into a surface representation of the differences that can be shown and can be understood
by the user. As well as considering visualisation methods, navigation and interaction issues
will be addressed. The service should therefore help learners to understand the comparison,
for example by using visual clues and contextualized help.
For the conceptual development service to be adopted, it is essential that its user interface and
reports meet the requirements of the users and provide neither too much nor too little
information. Iterative validation of the user interface with potential real users will take place
to inform improvements to the outputs. The accuracy of the service is important, i.e. how
good the extraction service is compared to human extraction (with dual or more codings to
balance inter-rater bias). Technology acceptance testing goes beyond accuracy and evaluates
whether the service is appreciated by learners, tutors, and other stakeholders (perceived
usefulness and ease of use).
Additional technical information on the showcase implementation can be found in deliverable
D2.2. Currently, development progresses towards the version 1 release candidate. Its
technology will be documented in D2.3, the stakeholder validation in D7.3, and verification
aspects in D4.2.
LTfLL -2008-212578
53
D4.1 Positioning Design
7. Conclusions
7.1 WP4.1
In summary, WP4.1 supports the positioning of an individual learner by means of qualitative
and qualitative feedback that is based on knowledge about learner language usage inferred
from the learners’ texts and learning materials.
WP4.1 approaches (i.e. knowledge poor LSA based and knowledge rich) rely on identifying
evidence of language usage in texts specifically surface level phrases and language usage
patterns. While the knowledge poor subtask infers linguistic knowledge from texts only, the
knowledge rich subtask makes use of concepts that are determined externally from text by
use of ontologies.
This report has described how positioning relates to communities of practice and speech
genres and how languages technologies can be used for positioning. The WP4.1 service
measures integration in such communities of practise by calculating distances based on
textual features and terminology. Moreover, the Stahl cycle has been used to show how
integration into expert communities of practice corresponds with level of expertise. For
WP4.1, formative feedback comes in the form of commentary on usage of language (e.g.
phrases and terminology) that can be used to facilitate effective and responsive
communication with experts in the community of practice.
In the next cycle of the project, WP4.1 will focus of on the development of version 1 of the
services for positioning based on the design presented here. For supporting the design and the
validation of the relevant language technologies for positioning, existing data sets (e.g.
graded and annotated online discussion in the medical domain) and new data sets (e.g. graded
and annotated computer science text materials in German) are being built. As the data sets are
being consolidated, multiple different configurations of the language technologies are being
tested. Experiments will be carried out to determine the best use of linguistic patterns
(significance of phrases, subject-verb-object grammatical patterns, etc.). Moreover, we expect
to experiment with new suffix array algorithms for extracting discontinuous phrases (tandem
repeats) as a new source of evidence. Then we will seek to find the best balance between the
knowledge poor and knowledge approaches, in order to build a positioning service that can
provide the most useful quantitative and qualitative feedback. This system will be validated
from the perspective of the user.
7.2 WP4.2
We conclude from our initial studies that we can visualise a learner’s ability to relate concepts
to one another within a specific domain and compare this to reference models. The type of
reference models used, based on either materials from the appropriate curriculum or
generated by communities of peer learners, situates these results within Stahl’s learning
model. Our ability to produce such a comparison in a way that is meaningful to end users will
form the basis for the next stage in production of the services based on WP 4.2, namely
provision of feedback to learners, by tutors and other educational stakeholders.
LTfLL -2008-212578
54
D4.1 Positioning Design
We also conclude that the tools used to produce these results, were not best suited to meeting
the stakeholder requirements captured in the scenario. The next step will be the development
of a customised approach based on LSA, for which the focus will be both the extraction and
aggregation of data. The final format of the service will depend on iterative validation by all
user groups.
LTfLL -2008-212578
55
D4.1 Positioning Design
References
Abouelhoda,M. I., Kurtz, S. & Ohlebusch E. (2004) Replacing suffix trees with enhanced suffix arrays. J. of
Discrete Algorithms, 2(1):53–86.
Alexander, P.A. (2003). The development of expertise: The journey from acclimation to proficiency.
Educational Researcher, 32, 10-14.
Bakhtin, M. M. (1981) The Dialogic Imagination: Four Essays, University of Texas Press
Bakhtin, M. M. (1986) Speech Genres and Other Late Essays. Trans. by Vern W. McGee. Austin, Tx: University
of Texas Press.
Bekkerman, R. and Allan, J. (2004) Using bigrams in text categorization. Technical Report IR-408, Center of
Intelligent Information Retrieval, UMass Amherst.
Berlanga, A. J., Brouns, F., Van Rosmalen, P., Rajagopal, K., Kalz, M., & Stoyanov, S. (2009). Making Use of
Language Technologies to Provide Formative Feedback. In S. D. Craig & D. Dicheva (Eds.), AIED 2009
Workshops Proceedings Volume 10. Natural Language Processing in Support of Learning: Metrics, Feedback
and Connectivity. Proceedings of the 14th International Conference in Artificial Intelligence in Education,
Workshops Proceedings (pp. 1-8). July, 6-7, 2009, Brighton, United Kingdom.
Berlanga, A. J., Kalz, M., Stoyanov, S., Van Rosmalen, P., Smithies, A., & Braidman, I. (2009b). Using
Language Technologies to Diagnose Learner’s Conceptual Development. In I. Aedo., N. Chen, Kinshuk, D.
Sampson & L. Zaitseva (Eds.), Proceedings of the 9th IEEE International Conference on Advanced Learning
Technologies (ICALT2009) (pp. 669-673). July, 14-18, 2009, Riga, Latvia: IEEE.
Bhattacharya, M. Heinrich, E., and Rayudu, R. (2006). Work in progress: E-portfolios in computer science
and engineering education. Frontiers in Education, 36th Annual Conference, 11-12.
Bleakley A, Bligh J (2008) Students learning from patients: Let's get real in medical education. Advances in
Health Sciences Education 13:89-107
Boshuizen, H. P. A., & Schmidt, H. G. (2008). The development of clinical reasoning expertise; Implications
for teaching. In: J. Higgs, M. Jones, S. Loftus, & N. Christensen (Eds.), Clinical reasoning in the health
professions, (3rd comp. rev. ed.). Oxford: Butterworth-Heinemann/Elsevier.
Bowker, G., C. and Star, S. (1999) L. Sorting Things Out: Classification and Its Consequences (Inside
Technology). The MIT Press.
Braidman I, Regan M, Cowie B (2008) Online reflective learning supported by student facilitators. Medical
Education 42:528-529
Brown, J. S. & Adler, R. P. (2008) Minds on fire; Open education, the long tail, and Learning 2.0. Educause
Review, 16-20, 22, 24, 26, 28, 30, 32.
Brown, J.S. & Duguid, P. (1991). Organizational Learning and Communities-of-Practice: Toward a Unified
View of Working, Learning, and Innovation, Organization Science, 2 (1), 40-57.
Burek, G., Pietsch, C., and De Roeck, A. (2007). SVO triple based Latent Semantic Analysis for recognising
textual entailment. Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing
(WTEP). 45th Annual Meeting of the Association of Computational Linguistics (ACL 2007), Prague. 113-118.
2007.
Centre for Recording Achievement ePortfolios briefing paper. Retrieved November, 2008 from
http://www.recordingachievement.org/downloads/eportfolios-briefing-paperSGRW11.doc.
Charlin, B., Boshuizen, H. P. A., Custers, E., & Feltovich, P. J. (2007). Scripts and clinical reasoning. Medical
Education, 41(12), 1178-1184(7).
Chinn, C. A., & Brewer, W. F. (1993). The role of anomalous data in knowledge acquisition: A theoretical
framework and implications for science instruction. Review of Educational Research, 63, 1-51.
Clariana, R., Koul, R., and Salehi, R. (2006). The criterion-related validity of a computer-based approach for
scoring concept maps. International Journal of Instructional Media, 33(3), 317-325.
LTfLL -2008-212578
56
D4.1 Positioning Design
Cover, T. and Hart P. (1967) Nearest neighbor pattern classification. Information Theory, IEEE Transactions on,
13(1):21– 27.
Dornan T., & David T. (2000) Adult Learning and Continuing Education Diabetic Medicine 17: 78-80
Dysthe, O. (1996) 'The Multivoiced Classroom: Interactions of Writing and Classroom Discourse'. Written
Communication, 13.3: 385-425
Eraut, M. (1994). Developing professional knowledge and competence. London: RoutledgeFalmer.
Ericsson, K. A. (ed., 1996). The road to excellence: The acquisition of expert performance in the arts and
sciences, sports and games. Mahwah, NJ: Erlbaum.
Ericsson, K. A. (1999). Creative expertise as superior reproducible performance: innovative and flexible aspects
of expert performance, Psychological Inquiry 10, 4, 329–361
Ericsson, K. A. (2004). Deliberate practice and the acquisition and maintenance of expert performance in
medicine and related domains. Academic Medicine, 10, S70-S81.
Ertmer, P. A., Newby, T. J. (1993). Behaviorism, cognitivism, constructivism: Comparing critical features from
an instructional design perspective. Performance Improvement Quarterly, 6 (4), 50-70.
Galley M., and McKeown K. (2003). Improving word sense disambiguation in lexical chaining. In Proceedings
of 18th International Joint Conference on Artificial Intelligence (IJCAI), Acapulco, Mexico.
Gangemi, A., Navigli, R., and Velardi, P. (2003). The OntoWordNet Project: extension and axiomatization of
conceptual relations in WordNet. R. Meersman et al. (Eds.), Conference on Ontologies, Databases and
Applications of SEmantics (ODBASE 2003), Catania, (Italy), Springer.
Garrison DR, Anderson T, Archer W (2000) Critical inquiry in a text-based environment: computer conferencing
in higher education. . The Internet and Higher Education 2:87 - 105.
Goldsmith, T. E., Johnson, P. J., and Acton, W. H. (1991). Assessing structural knowledge. Journal of
Educational Psychology, 83, 88–96.
Guarino, N. (2000). Invited Mini-course on Ontological Analysis and Ontology Design. First Workshop on
Ontologies and lexical Knowledge Bases - OntoLex 2000. Sozopol, Bulgaria.
Guba EG and Lincoln YS (2005), “Paradigmatic controversies, contradictions and emerging confluences” in
The Sage Handbook of Qualitative Research, ed. Denzin NK and Lincoln YS, Sage Publications Ltd
Hatano, G. & Inagaki, K. (1986). Two courses of expertise. In: H. W. Stevenson, H. Azuma, & K. Hakuta (Eds). Child developmen
Japan (pp. 262-272). New York: W H Freeman.
Hattie, J., Biggs, J., & Purdie, N. (1996). Effects of learning skills interventions on student learning: A metaanalysis. Review of Educational Research, 66, 99-136.
Hobbs, J. (1985). On the coherence and structure of discourse. CSLI TR 85-37. Stanford.
Ifenhaler, D., and Seel, N. M. (2005). The measurement of change: learning-dependent progression of mental
models. Technology, Instruction, Cognition and Learning, 2(4), 317-336.
Institute for Human and Machine Cognition. (2004). Cmap tools: University of West Florida.
Jeong, A. (2008). jMap v. 104. Available at http://dev22448-01.sp01.fsu.edu/ExcelTools/jmap
Jonassen, D. H., Beissner, K., and Yacci, M. (1993a). Structural knowledge: techniques for representing,
conveying, and acquiring structural knowledge. Hillsdale, NJ: Lawrence Erlbaum Associates.
Jonassen, D. H., McAleese, T. M. R. & Duffy, T. M. (1993b). A Manifesto for a constructivist approach to
technology in higher education. In Duffy, T. M., Lowyck, J. & Jonassen, D. H. (Eds.) Designing environments
for constructive learning: Implications for instructional design and the use of technology, Heidelburg, FRG:
Springer-Verlag
Jonassen, D., Reeves, T., Hong, N., Harvey, D., and Peter, K. (1997). Concept mapping as cognitive learning
and assessment tools. Journal of interactive learning research, 8(3/4), 289-308.
Kalz, M., Van Bruggen, J., Rusman, E., Giesbers, B., and Koper, R. (2007). Positioning of learners in Learning
Networks with content, metadata and ontologies. Interactive Learning Environments, 15(2), 191–200.
LTfLL -2008-212578
57
D4.1 Positioning Design
Kelly GA (1955) The Psychology of personal constructs, vol. 1 and 2. Norton, New York.
Kennedy, A. & Inkpen, D. (2006). Sentiment classification of movie reviews using contextual valence shifters.
Computational Intelligence. 22 (2), 110-125.
Kim, S. & Hovy, E. (2006). Extracting opinions, opinion holders, and topics expressed in online news media
text. In Proceedings of ACL/COLING Workshop on Sentiment and Subjectivity in Text, Sydney, Australia.
Kim, D. K., Kim, M. and Park, H. (2008) Linearized suffix tree: an efficient index data structure with the
capabilities of suffix trees and suffix arrays. Algorithmica, 52(3):350–377.
Knowles, M. S. (1975) Self-Directed Learning. A guide for learners and teachers, Englewood Cliffs: Prentice
Hall/Cambridge.
Landauer, T. K. & Dumais, S. T. (1997). A solution to Plato's problem: The Latent Semantic Analysis theory of
the acquisition, induction, and representation of knowledge. Psychological Review, 104, 211-240.
Lave, J., and Wenger, E. C. (1991). Situated Learning: legitimate peripheral participation. New York: Cambridge
University Press.
Liu, B. (2008). Web Data Mining: Exploring Hyperlinks, Contents, and Usage Data. Springer.
Lown N, Davies I, Cordingley L, Bundy C, Braidman I (2009) Development of a method to investigate medical
students' perceptions of their personal and professional development. Adv Health Sci Educ Theory Pract
14:475-86.
Merrifield, J., McIntyre, D. and Osaigbovo, R. (2000). Mapping APEL: Accreditation of prior experiential
learning in English higher education. London: Learning from Experience Trust.
Mieg, H. A. (2009). Two factors of expertise? Excellence and professionalism of environmental experts. High
Ability Studies, 20(1), 91-115.
Mihalcea R. (2007). Word Sense Disambiguation, Encyclopedia of Machine Learning, Springer.
Moilanen, K. and Pulman, S. (2007). Sentiment Composition. In Proceedings of Recent Advances in Natural
Language Processing (RANLP 2007). Borovets, Bulgaria, 378-382.
Nesbit, J., and Adesope, O. (2006). Learning with concept and knowledge maps: A meta-analysis. Review of
Educational Research, 76, 413-448.
Newman DR, Johnson C, Webb B, Cochrane C (1997) Evaluating the quality of learning in computer supported
co-operative learning. Journal of the American Society for Information Science 48:484 – 496
Nokes, T. J., & Ohlsson, S. (2003). Declarative transfer from a memory task to a problem solving task.
Cognitive Science Quarterly, 3, 259-296.
Nonaka, I., Nayana, R., & Konno, N. (2000). SECI, Ba and Leadership: a unified model of dynamic knowledge
creation. Long Range Planning, 30, 5-34.
Novak, J. D. (1998). Learning, creating and using knowledge: concept maps as facilitative tools in schools and
corporations. Mahwah, NJ: Erlbaum.
Norman, G. R., Eva, K., Brooks, L., & Hamstra, S. (2006). Expertise in medicine and surgery, In: K. A.
Ericsson, N. Charness, P. J. Feltovich, & R. R. Hoffman (Eds.), The Cambridge handbook of expertise and
expert performance, (pp 344-347). Cambridge: Cambridge University Press.
O’Neill, P. A., Willis, S. C., and Jones, A. (2002). A model of how students link problem-based learning with
clinical experience through ‘‘Elaboration’’. Academic Medicine, 77, 552–561.
Oberholzer, R., Egloff, S., Ryf, S., and Läge, D. (2008) Prodax. Datenauswertung im Bereich der Skalierung.
Benutzerhandbuch. Available at http://www.prodax.ch/ProDaX-Handbuch.pdf
Pang, B and Lee, L. (2004). A Sentimental Education: Sentiment Analysis Using Subjectivity Summarization
Based on Minimum Cuts. In Proceedings of ACL 2004, 271-278.
Peters, S. (2005). Infomap NLP software: an open-source. Package for natural language processing. Available at
http://infomap-nlp.sourceforge.net/
Pirnay-Dummer, P. (2006). Expertise and model building. MITOCAR. Unpublished doctoral dissertation.
University of Freiburg, Freiburg, Germany.
LTfLL -2008-212578
58
D4.1 Positioning Design
Rimon, M. (2005). Sentiment Classification: Linguistic and Non-linguistic Issues. In Proceedings of Israel
Association for Theoretical Linguistics, 21.
Salomon, G., & Perkins, D. N. (1989). Rocky roads to transfer: Rethinking mechanisms of a neglected
phenomenon. Educational Psychologist, 24, 113–142.
Salton, G. and Buckley, C. (1988) Term-weighting approaches in automatic text retrieval. In Information
Processing and Management, pages 513–523.
Sargeant, J Mann, K, Sinclair, D Ferrier S et al (2006) Learning in practice: Experiences and perceptions of
high-scoring physicians Academic Medicine 81: 655 – 660.
Schauer, H. (2000). Referential structure and coherence structure. TALN 2000. Lausanne, Switzerland.
Schmidt, H. G., Norman, G. R., & Boshuizen, H. P. A. (1990). A cognitive perspective on medical expertise:
Theory and implications. Academic Medicine, 65, 611-621.
Schön, D. (1987). Educating the reflective practitioner; toward a a new design for teaching and learning in the
professions. San Francisco: Jossey-Bass.
Shulman, L. S. (1986). Those who understand: Knowledge growth in teaching. Educational Researcher, 15 (2),
4-14.
Shute, V. (2008). Focus on Formative Feedback. Review of Educational Research, 78 (1), 153-189.
Shute, V. J., Jeong, A. C., Spector, J. M., Seel, N. M., and Johnson, T. E. (submitted). Model-Based Methods for
Assessment, Learning, and Instruction: Innovative Educational Technology at Florida State University. In
Yearbook Educational Media and Technology.
Schvaneveldt, R. W. (1990). Pathfinder associative networks: Studies in knowledge organization.
Nordwood,NJ: Ablex.
Simov, K., Peev, Z., Kouylekov, M., Simov, E., Dimitrov M. and Kiryakov, A. (2001). CLaRK - an XML-based
System for Corpora Development. In Proceedings of the Corpus Linguistics 2001 Conference, 558 -560
Simov, K. and Osenova, P. (2007). Applying Ontology-Based Lexicons to the Semantic Annotation of Learning
Objects. In Proc. Of RANLP 2007 workshop: Natural Language Processing and Knowledge Representation for
eLearning Environments. Borovets, 26. September 2007.
Simov, K. and Osenova, P. (2008). Language Resources and Tools for Ontology-Based Semantic Annotation. In
Proc. at the Workshop ‘OntoLex 2008’ at LREC 2008.
Smith, A. E., & Humphreys, M. S. (2006). Evaluation of unsupervised semantic mapping of natural language
with Leximancer concept mapping. Behavior Research Methods, 38(2), 262-279.
Spector, J. M., and Koszalka, T. A. (2004). The DEEP methodology for assessing learning in complex domains
(Final report to the National Science Foundation Evaluative Research and Evaluation Capacity Building).
Syracuse, N.Y: Syracuse University.
Spiro, R. J., Feltovich, P. J., Jacobson, M. J., & Coulson, R. L. (1992). Cognitive flexibility, constructivism, and
hypertext: Random access instruction for advanced knowledge acquisition in ill-structured domains. In T. M.
Duffy & D. H. Jonassen (Eds.), Constructivism and the technology of instruction: A conversation. Hillsdale, NJ:
Lawerence Erlbaum Associates, 57-76.
Stahl, G. (2006) Group Cognition: Computer Support for Building Collaborative Knowledge. Cambridge: MIT
Press.
Sternberg, R. J. & Frensch, P. A. (1992). On being an expert: A cost-benefit analysis. In R. R. Hoffman (Ed.),
The psychology of expertise: Cognitive research and empirical AI (pp 191-203). New York: Spinger.
Thijssen, J. G. L., & van der Heijden, B. I. J. M. (2003). Evaporated talent? Problems with talent development
during the career. International Journal of Human Resources Development and Management, 3(2), 154–170.
van Bruggen, J., Sloep, P., van Rosmalen, P., Brouns, F., Vogten, H., Koper, R., and Tattersall, C. (2004). Latent
Semantic Analysis as a tool for learner positioning in learning networks for lifelong learning. British Journal of
Educational Technology, 35 (6), 729-738.
Van der Heijden, B. I. J. M. (2000). The development and psychological evaluation of a multi-dimensional
measurement instrument of professional expertise, High Ability Studies, 11(1), 9-39.
LTfLL -2008-212578
59
D4.1 Positioning Design
van Merriënboer, J. J. G., Clark, R., & de Croock, M. B. M. (2002). Blueprints for complex learning: The
4C/ID-model. Educational Technology, Research and Development, 50, 39-64.
van Rosmalen, P., Sloep, P., Kester, L., Brouns, F., De Croock, M., Pannekeet, K., et al. (2008). A learner
support model based on peer tutor selection. Journal of Computer Assisted Learning, 24(1), 74-86.
Vosniadou, S., & Ortony, A. (1989). Similarity and analogical reasoning: A synthesis. In S. Vosniadou & A.
Ortony (Eds.), Similarity and analogical reasoning (pp. 1–17). Cambridge: Cambridge University Press.
Vygotsky, L.S. (1978). Mind in Society. Cambridge, MA: Harvard University Press.
Weggeman, M. C. D. P. (2000). Kennismanagement: de praktijk [Knowledge management: the practice].
Schiedam: Scriptum.
Wenger E (2001) Supporting communities of practice: a survey of community-oriented technologies. Selfpublished report available at www.ewenger.com/tech,.
Wenger, E. C. (1998). Communities of Practice, learning, meaning and identity. New York: Cambridge
University Press.
Wertsch, J. V. (1991) Voices of the mind: A sociocultural approach to mediated action . Cambridge, MA :
Harvard University Press
Wiebe, J. (2002). Instructions for Annotating Opinions in Newspaper Articles. Department of Computer Science
Technical Report TR-02-101, University of Pittsburgh, Pittsburgh, PA.
Wolf, F. & Gibson, E. (2006). Coherence in Natural Language. Data structures and applications. A Bradford
Book. The MIT Press.
Woods, N. N. (2007). Science is fundamental: the role of biomedical knowledge in clinical reasoning. Medical
Education, 41, 1173-1177.
Woods, N. N., Brooks, L. R., & Norman, G. R. (2005). The value of basic science in clinical diagnosis: Creating
coherence among signs and symptoms. Medical Education, 39, 107-112
Woods, N. N., Brooks, L. R., & Norman, G. R. (2007). It all makes sense: Biomedical knowledge, causal
connections and memory in the novice diagnostician. Advances in Health Sciences Education, 12, 417–426.
Yamamoto, M. and Church, K.W. (2001) Using suffix arrays to compute term frequency and document
frequency for all substrings in a corpus. Comput. Linguist., 27(1):1–30.
LTfLL -2008-212578
60
D4.1 Positioning Design
[1]
Optionally the tokenizer could be set to eliminate such punctuation marks. The
bullet point makes a good example here, however, due to its burstiness.
[2]
This definition and the following definitions are similar to those found in
Abouelhoda et al 2004.
[3]
We assume here that the text is padded with unique beginning of string and end of
string sentinels so that indexing at 0 or 7 makes sense.
[4]
In the process of experiments with the actual data we will refine this scale of
values.
[5]
Notice that in this first exploration “expert knowledge” has been defined as the
course content. The “expert knowledge” can be also seen as the knowledge a learner, who is
considered by the tutor as an expert learner, has in a particular context.
LTfLL -2008-212578
61