Cognitive style is defined by Riding and Rayner (1998) as ...

17
Instructional Science 30(6), 2002. Modelling Cognitive Style in a Peer Help Network Susan Bull 1 and Gord McCalla 2 1 Educational Technology Research Group, Electronic, Electrical and Computer Engineering, University of Birmingham, Edgbaston, Birmingham, B15 2TT, United Kingdom. [email protected] 2 ARIES Laboratory, Department of Computer Science, University of Saskatchewan, Saskatoon, Saskatchewan, S7N 5A9, Canada. [email protected] Abstract: I-Help is a computer system that assists learners as they try to solve problems while learning a subject. I-Help achieves this by supporting a network of peers that help each other out. One component of I-Help selects appropriate peers to assist a student, and then sets up a one-on-one peer help session between the helper and the helpee. The matching of helper to helpee takes into account factors such as a potential helper's knowledge of the topic of the helpee's question; their availability and eagerness to help; and their general helpfulness. Recent work has developed a cognitive style component to supplement these attributes, which enables consideration also of the suitability of a helper’s cognitive style for answering the helpee’s question. This paper describes how modelling individuals' cognitive style can usefully supplement other user model data in a peer help network, and describes how this information is obtained in I-Help. Keywords: cognitive style, learner characteristics, peer help, collaborative learning, networked learning, user modelling, computer mediated learning. Introduction One-on-one peer help networks can support intensive interaction amongst students, with both the helper and helpee benefiting through reflection and tutoring. Computer-based systems can offer useful methods of partnering individuals for such interaction through user modelling – matching helpees with helpers who are suitable according to knowledge of a topic and other factors entails the maintenance of relevant information on all participants (Bull, 1997; Hoppe, 1995; Mühlenbrock et al., 1998; Ogata et al., 1999; Vivacqua, 1999). I-Help is a system that supports such peer help networks, matching helpers and helpees for just-in-time one- on-one help sessions according to a number of attributes represented in the user models of learners. These include: an individual's knowledge level of a range of topics; their helpfulness; their eagerness to participate in I-Help sessions; their availability; the relative importance of various attributes in a learning partner (e.g. relative importance to the helpee of the helper's knowledge level versus their helpfulness versus their immediate availablity, etc.). With such a range of attributes modelled, the utility of I-Help increases as the number of participants increases. Indeed, although there is a place for I-Help in small group learning (Bull and Greer, 2000), I-Help is designed primarily for use where student cohorts are large. A further factor relevant to the educational context is individual differences in the learning process. It is useful to consider such differences when matching partners for peer interaction. This paper describes the incorporation of cognitive style as an additional attribute in I-Help user models, highlighting how information about cognitive style augments the other attributes discussed above, in matching helpers and helpees. The next two sections provide an overview of the I-Help peer network and an overview of cognitive style. The remainder of the paper then describes in greater depth, how cognitive style is used in I-Help. In particular, the paper gives a detailed analysis of I-Help user models. It is shown how cognitive style can be incorporated into I-Help user models and used to enhance the selection of an appropriate peer helper for particular kinds of questions faced by a helpee. There follows a discussion about how a person’s cognitive style can be determined, including an investigation into a questionnaire for extracting provisional cognitive style information from I-Help users, and a discussion of how such a questionnaire could be used in conjunction with internal inferences by I- Help as users use the system, in order to better understand, over time, a user’s cognitive style. The paper concludes with an examination of the issues raised by incorporating cognitive style into a peer help network such as I-Help. I-Help The I-Help peer help network is an environment where students can ask and answer questions about their assignments and courses, based on the metaphor of a help-desk (Greer et al., 1998). I-Help has two integrated parts: a Public Discussion forum, where students post questions and answers to be seen by all their classmates, and a Private Discussion component, the focus of this paper, where students can give and receive help in a one-

Transcript of Cognitive style is defined by Riding and Rayner (1998) as ...

Page 1: Cognitive style is defined by Riding and Rayner (1998) as ...

Instructional Science 30(6), 2002.

Modelling Cognitive Style in a Peer Help Network

Susan Bull1 and Gord McCalla2

1 Educational Technology Research Group, Electronic, Electrical and Computer Engineering, University of Birmingham, Edgbaston, Birmingham, B15 2TT, United Kingdom. [email protected]

2 ARIES Laboratory, Department of Computer Science, University of Saskatchewan, Saskatoon, Saskatchewan, S7N 5A9, Canada. [email protected]

Abstract: I-Help is a computer system that assists learners as they try to solve problems while learning a subject. I-Help achieves this by supporting a network of peers that help each other out. One component of I-Help selects appropriate peers to assist a student, and then sets up a one-on-one peer help session between the helper and the helpee. The matching of helper to helpee takes into account factors such as a potential helper's knowledge of the topic of the helpee's question; their availability and eagerness to help; and their general helpfulness. Recent work has developed a cognitive style component to supplement these attributes, which enables consideration also of the suitability of a helper’s cognitive style for answering the helpee’s question. This paper describes how modelling individuals' cognitive style can usefully supplement other user model data in a peer help network, and describes how this information is obtained in I-Help. Keywords: cognitive style, learner characteristics, peer help, collaborative learning, networked learning, user modelling, computer mediated learning.

Introduction One-on-one peer help networks can support intensive interaction amongst students, with both the helper and helpee benefiting through reflection and tutoring. Computer-based systems can offer useful methods of partnering individuals for such interaction through user modelling – matching helpees with helpers who are suitable according to knowledge of a topic and other factors entails the maintenance of relevant information on all participants (Bull, 1997; Hoppe, 1995; Mühlenbrock et al., 1998; Ogata et al., 1999; Vivacqua, 1999). I-Help is a system that supports such peer help networks, matching helpers and helpees for just-in-time one-on-one help sessions according to a number of attributes represented in the user models of learners. These include: an individual's knowledge level of a range of topics; their helpfulness; their eagerness to participate in I-Help sessions; their availability; the relative importance of various attributes in a learning partner (e.g. relative importance to the helpee of the helper's knowledge level versus their helpfulness versus their immediate availablity, etc.). With such a range of attributes modelled, the utility of I-Help increases as the number of participants increases. Indeed, although there is a place for I-Help in small group learning (Bull and Greer, 2000), I-Help is designed primarily for use where student cohorts are large. A further factor relevant to the educational context is individual differences in the learning process. It is useful to consider such differences when matching partners for peer interaction. This paper describes the incorporation of cognitive style as an additional attribute in I-Help user models, highlighting how information about cognitive style augments the other attributes discussed above, in matching helpers and helpees. The next two sections provide an overview of the I-Help peer network and an overview of cognitive style. The remainder of the paper then describes in greater depth, how cognitive style is used in I-Help. In particular, the paper gives a detailed analysis of I-Help user models. It is shown how cognitive style can be incorporated into I-Help user models and used to enhance the selection of an appropriate peer helper for particular kinds of questions faced by a helpee. There follows a discussion about how a person’s cognitive style can be determined, including an investigation into a questionnaire for extracting provisional cognitive style information from I-Help users, and a discussion of how such a questionnaire could be used in conjunction with internal inferences by I-Help as users use the system, in order to better understand, over time, a user’s cognitive style. The paper concludes with an examination of the issues raised by incorporating cognitive style into a peer help network such as I-Help. I-Help The I-Help peer help network is an environment where students can ask and answer questions about their assignments and courses, based on the metaphor of a help-desk (Greer et al., 1998). I-Help has two integrated parts: a Public Discussion forum, where students post questions and answers to be seen by all their classmates, and a Private Discussion component, the focus of this paper, where students can give and receive help in a one-

Page 2: Cognitive style is defined by Riding and Rayner (1998) as ...

on-one interaction. The Public Discussions are intended for questions requiring a quick, straightforward response, or questions which might elicit multiple viewpoints in the replies (Bull et al., 2001a). The Private Discussions are more suitable for in-depth issues, requiring more intensive one-on-one interaction such as peer tutoring. Nevertheless, it has been found that some students prefer to use one I-Help component over the other for all question types (Greer et al., 2001). A Private Discussions help session is displayed as in Figure 1.

Figure 1: An example help session in I-Help Private Discussions A virtual economy underlies the I-Help Private Discussions, and students spend (or receive) I-Help credit units (ICUs) as they seek (or give) help (Vassileva et al., 1999). The virtual economy is intended to motivate students to offer help and to constrain students from making frivolous requests for help. The Private Discussions component includes a personal agent for each student. This personal agent works to find the best learning partners for its user, with the assistance of a matchmaker agent that suggests potential learning partners, and through subsequent negotiations with the personal agents of these partners (Vassileva et al., 1999). These negotiations are based on the contents of the user models of the participants. I-Help is designed to benefit not only those students with help requests, who receive answers to their questions, but also their peers providing the help. Students offering help will often learn further through their formulation of a response, since the need to articulate their answer clearly and the act of writing for an authentic audience should facilitate reflection. Cognitive style A person's cognitive style is defined by Riding and Rayner (1998) as comprising fixed characteristics relating to methods of information processing and organisation. This notion of stability is supported by findings suggesting a cerebral basis for differences in cognitive style, as revealed by EEG Alpha readings (Riding et al., 1997). In accordance with these physiological findings, longitudinal studies have indicated cognitive style to be quite stable over three and a half years (Clapp, 1993). Other aspects of learning such as learning style or learning strategies may be less stable: for example, it has been argued that teaching, assessment, course organisation and the particular task can affect students' approaches to learning (Laurillard, 1979; Newble and Hejka, 1991; Ramsden, 1979; Ramsden and Entwistle, 1981), as can prior knowledge (Moran, 1991) and content area (Westman, 1993). It has also been suggested that learning style can be a matter of habit (Valley, 1997). For learner modelling purposes in I-Help, where students could be asking questions on a range of topics, in a range of departments and institutions, and perhaps using I-Help in different types of course within a single institution, a more stable attribute (such as cognitive style) is most useful, if it can be recorded. Reviewing the literature on cognitive style, Riding and Cheema (1991) found more than thirty labels which they grouped into two independent dimensions:

1. wholist-analytic e.g. Witkin et al.'s (1977) field-dependence/independence; Pask's (1976) holist/serialist; 2. verbal-imagery e.g. Riding et al.'s (1989) verbal/imagery style.

Page 3: Cognitive style is defined by Riding and Rayner (1998) as ...

The wholist-analytic (WA) dimension refers to the extent to which an individual processes information in wholes or separate parts; the verbal-imagery (VI) dimension relates to the degree to which an individual represents information during thinking in verbal or image form. Individuals who can adapt to both aspects of the WA dimension are intermediates; those accommodating each aspect of the VI dimension are bimodals. User modelling in I-Help As described in the introduction, the I-Help matchmaker agent takes a variety of information into account when suggesting potential learning partners. Cognitive style is therefore only one of a range of factors that could be considered when the matchmaker agent ranks potential partners for one-on-one help sessions. Table 1 shows part of the user model of an individual, to illustrate the range of attributes modelled. The standard default values are also shown for comparison. (The maximum value for numeric representations of user attributes is 10.)

Table 1: A computed I-Help user model

User Attributes

I-Help / user helpfulness

I-Help / user eagerness readiness cognitive

style maximum

discussions privacy

user 1 7.5 / 8 6.89 / 10 online analytic imager 15 show all

default 5 / 3 5 / 3 offline - 7 protected Knowledge Level

I-Help / user topic 1

I-Help / user topic 2

I-Help / user topic 3

I-Help / user topic 4

I-Help / user topic 5

I-Help / user topic 6

user 1 8.75 / 9 6.5 / 7 5 / 5 6.25 / 6 5.5 / 6 8 / 7 default 2 / 2 2 / 2 2 / 2 2 / 2 2 / 2 2 / 2

Currency min helper price

max helpee price

Preference and Banned Lists

preferred helpers

banned people

banned topics

user 1 2 10 user 1 user 3, user 16 user 65 topic 3 default 3 6 default - - - Weighting of Attributes

helper's helpfulness

helper's eagerness

helper's readiness

helper's cognitive style

helper's knowledge currency

user 1 6 8 4 8 6 2 default 6 6 6 6 6 6

Although the user model in Table 1 has been assembled for illustrative purposes, many of the attributes are distributed and fragmented throughout the personal agents of the I-Help participants, and attributes are not computed until they are required. This lack of a 'constructed' user model occurs in part because user model fragments reside in different locations; in part because they are drawn from different sources; in part because data may be contradictory; in part because user model data will be differentially applicable for different modelling purposes; and in part because updates to user model fragments may be received at any moment (see Bull et al., 2001b; McCalla et al., 2000). For example, the helpfulness attribute is drawn from three sources: self assessment; helpee evaluations of helpers after completion of a help session; and user response rate in both components of I-Help (Public as well as Private Discussions). Self assessments of helpfulness reside in the user's personal agent; peer evaluations are located in the agents of all peers who have evaluated an individual; and user response rate is represented in the I-Help database, which can be accessed only by the matchmaker. Given the fragmented nature of representation of the various attributes, it is likely that some of the data will be contradictory. However, since new evidence may arrive at any time (for example, after some user action or a peer evaluation), there is no need to reconcile such conflicts until the information is required. Furthermore, different modelling purposes will demand the use of different data. Even within the peer matching context under focus here, reconstruction of an attribute will only be important in matchmaking if a helpee has so weighted it. Thus, a complete computed user model as shown in Table 1 will never actually exist – only those aspects relevant in a particular context will be constructed (and will be combined according to a weighted linear equation that is appropriate to that context). There are five kinds of user model fragment:

1. User-given information on attributes such as the user's helpfulness, eagerness to help, cognitive style, knowledge level of the various topics, maximum number of concurrent discussions and privacy level – i.e. what information about the user should be available to others (Figure 2).

2. I-Help's own assessment of the user's helpfulness (based on response rate and helpee evaluations of previous help sessions), eagerness (based on participation to date), readiness (whether the user is online, or likely to be online soon) and knowledge level (based on peer evaluations).

Page 4: Cognitive style is defined by Riding and Rayner (1998) as ...

3. User-provided lists of people the individual would like to be given preference as helper; people with whom the user does not wish to interact; topics about which the user does not wish to provide help (Figure 3).

4. The user's stated weightings for the importance of the various attributes in a helper: whether it is more important that their helper knows the subject in detail versus whether they should be able to respond soon versus their helpfulness, etc. (Figure 3).

5. The maximum currency in I-Help Credit Units that users will pay for help, and the minimum they will accept for providing help (Figure 3).

Figure 2: User-provision of user model data

Figure 3: Assigning weightings to helper attributes Some user model information is provided directly by users, as illustrated in Figures 2 and 3, and some is based on user activity – for example, readiness is calculated according to whether the user is currently or frequently online. Some representations, such as eagerness, helpfulness and knowledge level, utilise both user and system data. In addition to user activity in I-Help, some of the system data (helpfulness and knowledge level) is drawn partly from short peer evaluation questionnaires which are presented after a help session is completed. Some preferences are provided only by the user: e.g. privacy level, maximum number of concurrent discussions and preferences in a learning partner. Users can add people to a 'banned list' or 'preferred helper list', to give higher weighting to those with whom they would particularly like to interact, or to ensure that they are not matched with someone with whom they wish no contact. Topics can also be banned.

Page 5: Cognitive style is defined by Riding and Rayner (1998) as ...

The rate of earning and spending I-Help Credit Units is also under the user's control – they can set the maximum price they are prepared to bid for help and the minimum they will accept for offering help. The currency is intended in part as motivation for students to provide help to others, and this does appear to motivate some students, while others interact ignoring the I-Help economy (see Greer et al., 2001). As with other attributes, the importance of earning and saving I-Help Credit Units can be weighted by users. Cognitive style is thus only one of a number of attributes modelled, and its importance will depend on the comparative weighting (assigned by a helpee) of the cognitive style attribute for helpers. With a low rating, cognitive style will have little influence on the matching of partners. However, if weighted higher, the impact of cognitive style on the matchmaking will be greater. Cognitive style in the I-Help user model Sadler-Smith and Riding (1999) suggest three responses to acknowledge individual differences in learners: (1) matching instructional methods, media and assessment to the preferences of learners; (2) mismatching preferences to encourage the development of broader approaches to learning; (3) taking only the purpose of instruction into account, when selecting methods and media of instruction. User modelling of learner attributes has been described in relation to adaptation in a computer-based learning environment (e.g. Milne et al., 1997), as have approaches that assign learners to different versions of a program, or to different teaching strategies within a program, based on cognitive or learning style information (e.g. Clarke, 1993; Groat and Musson, 1995). For I-Help the situation is a little different: I-Help has no content knowledge. Thus, the system does not adapt its presentation or teaching, but instead it makes the choice of a helper that is adapted to the immediate needs of the helpee. Therefore, in our context, the questions regarding cognitive style are: (1) should a helpee be helped by someone with a similar cognitive style?; (2) should the helper and helpee have dissimilar cognitive styles?; or (3) should the purpose of the interaction determine the cognitive style of a helper? Since helpers are not experts, we have chosen to focus on the purpose of the interaction (as in question 3 above): that is, it would be most appropriate to choose helpers with a cognitive style appropriate for the kind of explanation required by the helpee’s question. Insofar as cognitive style is important in matching a helper to a helpee’s request (which is determined by weights assigned by the helpee), a helper will receive questions that they are better equipped to answer in terms of their cognitive style. Moreover, a helpee should more readily appreciate the helper’s explanations, as they are likely to be authored by a person with an understanding of the requirements of a good answer in the context of the particular question. Thus, as illustrated in Figure 4, unlike other attributes in the I-Help user model (knowledge level, readiness, eagerness, helpfulness), the desired cognitive style of a helper depends on the helpee's question type, rather than being generally applicable in every situation. For some questions a particular cognitive style in a helper is more suitable, regardless of the cognitive style of the helpee, whereas for other kinds of question a helper-helpee cognitive style match is sought.

question

readiness

helpfulnesseagerness

knowledge

cognitivestyle

Figure 4: Cognitive style and other attributes modelled in I-Help The most obvious relevant aspect of cognitive style for I-Help is Riding and Cheema's (1991) wholist-analytic dimension, since this deals with the way in which an individual processes information – by drawing relationships between parts of a whole, or as separate independent parts and details. Therefore the wholist-analytic dimension was included in I-Help. Riding and Cheema's verbal-imagery dimension perhaps appears at first consideration to be less relevant to I-Help, since I-Help interactions are textual. However, the verbal-imagery dimension is useful here, and was also included since, according to the model proposed by Riding and Sadler-Smith (1992), imagery may be able to some extent compensate for the absence of wholism for an analytic-imager (i.e. the individual might be able to gain an overall perspective through the application of imagery). Similarly, the wholist-verbal learner may be able to draw on his or her verbal thinking to help overcome any difficulties arising from their lesser level of analytic processing. Thus, there may be a possibility of considering styles according to wholist versus analytic

Page 6: Cognitive style is defined by Riding and Rayner (1998) as ...

processing taken in combination with verbal and imagery thinking, resulting in a ordering from wholist-imager to analytic-verbaliser (Riding and Caine, 1993). Another finding, reported by Riding and Calvey (1981), is that imagers (11 year olds) find it easier to recall descriptive text than acoustically complex text, and vice versa for verbalisers. Hence, if this also applies to adults, even in a non-graphical environment information about the verbal-imagery dimension of cognitive style could be useful for peer matching. What this means is that, despite currently being a text-only system, I-Help could make use of the VI dimension in finding appropriate peer helpers, depending on the type of question being asked. Consider finding a peer helper for a question such as 'does anyone have a good reference for [topic]?': it is likely that an imager will recommend resources helpful to another imager, and verbalisers for verbalisers. This is supported by the finding of Riding and Watts (1997) that, given a choice of learning materials, verbalisers are more likely to select textual materials over graphical ones, and vice versa for imagers. Riding and Ashmore (1980) discovered that individuals (11 year olds) recalled more of the information presented to them in their preferred mode. Riding and Douglas (1993) noted that imagers (15-16 year olds) learned better from text plus picture mode, while verbalisers learned similarly from text plus text and text plus picture. Barker et al. (2000) found strong verbalisers and imagers benefited most from multimedia presentations that matched their cognitive style. However, not every study shows such direct effects of the VI dimension. In fact, some experimental work has revealed no significant differences between the performance of matched and mismatched groups (Pillay, 1998). It has even been found that expert and novice undergraduate verbalisers learned programming concepts better from a text plus graphical metaphor, as opposed to text plus text, and novice imagers learned better from text plus text (McKay, 1999a; 1999b). Taking the overall cognitive style construct into account (where readings from both dimensions are combined), McKay (2000) recommends providing a text plus graphical metaphor for wholist-verbalisers, analytic-verbalisers and wholist-imagers, and text only for analytic-imagers. The interactions between cognitive style and visual-verbal perception and expression are therefore obviously quite complex, but what is clear is that imagery seems important for some students. Finally, future versions of I-Help will allow some degree of graphical interaction to allow greater flexibility of expression. Riding and Douglas (1993) found that imagers in their text plus picture group produced more pictures in their answers than did verbalisers from this group. Cox (1999) observed individual differences in spontaneously produced graphical and non-graphical external representations during analytical reasoning. Therefore, the planned extension to I-Help to allow graphical interaction, while enabling helpees to more easily receive help in their preferred mode on the verbal-imagery dimension (where appropriate), will also be less restrictive on helpers. Thus, future versions of I-Help will have an even stronger requirement for user modelling on the VI dimension of cognitive style, to complement modelling on the WA dimension. As indicated above, in I-Help it is not the case that students should necessarily be matched with someone having the same cognitive style for all interactions, nor is it the case that they should always be mismatched. This decision depends on the type of question. For each question type, one of the cognitive style dimensions may be more important than the other. When students enter their question into I-Help, they also state the question type as illustrated in Figure 5.

Figure 5: Indicating the type of question

Page 7: Cognitive style is defined by Riding and Rayner (1998) as ...

The preferred cognitive style for a helper for each of these question types is described below: • What are the details of…?

For this question type an analytic helper is sought, regardless of whether this matches with the cognitive style of the helpee, because analytics tend to grasp the details of a topic more readily than wholists. In the absence of a suitable analytic helper, a wholist-verbaliser can be selected in the hope that the verbal aspect will compensate for the lack of analytic tendency (Riding and Sadler-Smith, 1992).

• How does this fit with other things? For this type of question a wholist helper is required, irrespective of the cognitive style of the helpee, because wholists will be better equipped to provide an adequate answer where an overview of a topic is needed. If there is no suitable wholist helper, an analytic-imager is sought.

• Can you suggest any good resources for…? The aim is to match individuals on the verbal-imagery dimension since, if Riding and Ashmore's (1980) findings apply, a verbaliser is more likely to find useful, and recommend, materials with a more textual bias, whereas an imager might recommend more graphical resources. If McKay's (1999a; 1999b) contrary results are applicable in our context, assuming students are at the same knowledge level, the same argument favouring matching applies nevertheless: those with one cognitive style will recommend materials suitable for that cognitive style. McKay’s (2000) results suggest that for this question type, it may also be useful to consider the overall cognitive style construct for the next version of I-Help. Bimodals are ideally matched with other bimodals for resource recommendations, since they are more likely to get a greater range of recommendations from bimodals who, like themselves, can easily learn from both modalities.

• How do you…? / Can you clarify what I need to do? This is essentially the same as 'what are the details of…?', applicable in procedural domains. It is given separately to help learners classify their questions more easily.

• Miscellaneous general question General questions cover all questions not included in the above. The default is to match all learners on the wholist-analytic dimension. The importance of cognitive style information for general questions will be more varied than in the previous cases. In some cases cognitive style will have little influence. However, general questions could include more complex exchanges, where matching might become useful.

• Simple question requiring a straightforward answer This kind of question is best asked in the Public Discussions component of I-Help. However, it has been found that some students have a strong preference for using one component of I-Help over the other (Greer et al., 2001), and it is therefore included here to cater to students who prefer the one-on-one alternative. For questions requiring a straightforward answer, cognitive style is ignored, since it is unlikely to be relevant.

With the exception described above for the recommendation of resources, intermediate or bimodal helpees can be matched with any cognitive style on that same dimension, as they are flexible. However, intermediate or bimodal helpers on the dominant dimension for the particular question type are selected only for intermediate or bimodal helpees, as it is not possible to predict how they will respond. Students who have not provided cognitive style information may sometimes be given preference in the selection process, over those who have supplied data. For example, for a 'details' question, if no suitable analytic helper is available, someone with no cognitive style representations in their user model will be selected over a wholist, since this helper may be analytic. In the future, as we gain more experience in the use of cognitive style as a factor in matching helpers to helpees, we may acquire a more subtle understanding of the interactions among the dimensions of cognitive style in the I-Help context. At that point, it may be useful to expand the above matching rules to consider the overall cognitive style construct – the combined readings from both dimensions, for example identifying an individual’s cognitive style as wholist-imager (see Riding, 1991; 1998). Supporting this view is a large-scale study of first year University students carried out by McKay (2000), where the cognitive style construct was explored among students learning programming concepts. McKay’s results suggest that this construct view is important in analysing how people learn. In particular, the study recommends presentation using a text plus graphical metaphor for wholist-verbalisers, analytic-verbalisers and wholist-imagers, and text only for analytic-imagers. This may be particularly relevant to requests for resources in I-Help. However, these results may not directly apply to other question types in I-Help, as we are concerned with how students offer and process explanations, rather than the more general problem of learning. We hypothesise also that for some question types, one of the cognitive style dimensions may be more important than the other. Inclusion of the more flexible intermediate and bimodal parts of the continua are also important in our context. Therefore, for I-Help it would be important to recognise the cognitive style construct as a point in the two dimensional space defined by the VI and WA dimensions, rather than taking an absolute choice of one value from each dimension.

Page 8: Cognitive style is defined by Riding and Rayner (1998) as ...

What cognitive style offers As stated above, it is not suggested that cognitive style should be the only source of information used to match learners for help sessions – it should be used alongside other attributes (e.g. eagerness; knowledge level; relative importance of various helper characteristics such as knowledge level versus helpfulness versus cognitive style, etc.). While this results in reduced impact of cognitive style representations, these remain sufficiently influential in cases where they might be important. The following illustrates the benefit that inclusion of cognitive style representations in user models brings to I-Help, through examples of how it influences the ranking of potential helpers. In ranking suitable potential helpers with reference to the range of attributes, the matchmaker agent weighs suggestions according to the helpee's preferences in a helper. The matchmaker checks that none of the top-ranked people have been banned by the helpee, and that with this new request they would still be within their stated limit of concurrent help sessions. The helpee's agent then negotiates with the personal agents at the top of the list over the price (see Vassileva et al., 1999), modifying the list if necessary. The final top five helpers are then sent the help request, and the first to accept it takes on the role of helper. An example of matchmaker ranking is given in Table 2, showing the top eight ranked potential helpers (of 76) for a question taken from a previous I-Help deployment in a java programming course. Cognitive style was not modelled during that deployment1, but it is illuminating to see what the impact might have been, had information on the individuals' cognitive styles been included in the matchmaking.

Table 2: Matchmaker ranking of helpers without cognitive style

Potential Helpers helper 1 helper 2 helper 3 helper 4 helper 5 helper 6 helper 7 helper 8

ranking 69.97 49.52 47.1 44.32 39.1 37.55 35.97 28.38 The cut-off point for being notified of the help request was helper 5 (since a fixed number of the top ranked helpers receive the request). Helper 5's score is quite close to the scores of helpers 6 and 7, but helpers 6 and 7 never saw the request, and therefore did not get a chance to answer it. As described previously, cognitive style matching and mismatching depends on the specific question type. This request was of the type 'what are the details of…?', preferring an analytic helper. No data on cognitive style was recorded at this time, but it may have been that helper 5 was a wholist, and helpers 6 and 7 were analytics. If this were true, helper 6 may have been a more appropriate recipient of the help request than helper 5. Indeed, cognitive style may have had implications further up the list. While it is unlikely that helper 1 would have been displaced, given the obvious suitability in other areas (helper 1 had particularly high scores for system and user knowledge level, helpfulness and eagerness), if any of helpers 2, 3 or 4 had been wholists, their scores may have been low enough for helpers 6 and even 7 to replace them on the list. The score for helper 8 is lower – probably too low to displace the actual recipients of the request, even if helper 8 were analytic. While initial study has suggested matchmaker rankings to be generally quite effective even with this reduced set of attributes (Bull et al., 2001b), given that cognitive style is likely to be important for some students, there are some students for whom the cognitive style attribute should lead to better pairings. Moreover, in large courses, the more information available to the matchmaker about each individual, the more fine-grained the selection of good potential helpers for an individual is likely to be. For cognitive style, this should benefit to some extent all helpees except those who set the cognitive style component in their helper preferences to 'unimportant' (the lowest point on a five-point scale). Nevertheless, even for these helpees, the fact that the cognitive style of a potential helper need not be considered during the matchmaking process is useful information. With many potential helpers from which to select, the more information available, the better. Table 3 shows the matchmaker rankings for a request for resources in a languages course with 43 participants, from a recent I-Help deployment, this time with cognitive style included as a component of the user models. The helpee was a verbaliser (on the scale: strong verbaliser, verbaliser, bimodal, imager, strong imager); hence for a question about resources, a verbaliser or strong verbaliser would be a strong candidate as helper. The helpee had assigned equal weighting to all helper attributes.

Table 3: Matchmaker ranking of helpers with cognitive style

Potential Helpers helper 1 helper 2 helper 3 helper 4 helper 5 helper 6 helper 7 helper 8

ranking 62.76 57.84 57.72 55.92 51.72 50.085 48.84 48.72 cognitive style

strong verbaliser

strong verbaliser verbaliser strong

verbaliser bimodal verbaliser verbaliser bimodal

Page 9: Cognitive style is defined by Riding and Rayner (1998) as ...

It can be seen that no imagers appeared on this list. However, two bimodals are evident, one making it to the final 5. Since the helpee's user model representation for the importance of the helper's cognitive style was equivalent to the importance of the other helper attributes, this allowed other factors to influence the choice. The request was labelled 'urgent', and helper 5 was online at the time the request was submitted. Helper 5 also had high user-assessed knowledge level (7) and high user and system eagerness (both 10). Helper 8's user assessment of knowledge of the topic, and user and system evaluation of eagerness to participate were high (4, 10 and 10 respectively), outweighing the representation for cognitive style. (Although 4 appears low, at this stage of the course the average self assessed knowledge level for the topic was 2.53, with a median of 2 and a range of 0-7.) Since the importance of the helper's knowledge level and eagerness were set at the same weighting as cognitive style in the helpee's user model, this enabled this greater influence for these higher-scoring attributes. Helper 9 was a verbaliser, but did not make it higher on the list mainly because of a lower user eagerness score (6) than the other potential helpers on the list.

Table 4: Partial user models of potential helpers at different rankings

Attributes I-Help helpfulness

user helpfulness

I-Help eagerness

user eagerness

I-Help knowledge

user knowledge readiness cognitive

style

helper 1 6.25 8 10 10 - 6 offline (back soon)

intermediate strong verb.

helper 5 6.25 6 10 10 - 7 online analytic bimodal

helper 10 5 6 5 10 - 2 offline wholist bimodal

Table 4 shows part of the user models of potential helpers from different points on the ranked list from the example in Table 3, to illustrate how the different attributes can influence helper rankings. No peer evaluations concerning these helpers had been performed at the time of the help request; therefore there was no alteration of system knowledge based on peer evaluations. Retaining currency was weighted by the helpee as 'unimportant', and so the earning currency attribute in the helper user models was largely ignored in the matchmaking process (and has been omitted from Table 4). The helpee had rated all other helper attributes as equally important. The excerpt from the ranked list in Table 4 therefore looks quite straightforward. However, if any of these attributes had been differently weighted, a different list from that in Table 3 would likely have resulted. If the question had been of another type, again, a different ranking would have been calculated. For example, a question asking for clarification of details would have favoured an analytic helper. In that case helper 5 would probably have been higher. From the example in Table 3, it can be seen that the cognitive style of a helper can make a difference even when it is weighted similarly to the other attributes in the user model of the helpee. In such cases, when the other attributes have a particularly high score, those attributes will be more dominant. Otherwise cognitive style will have more weight in determining potential helpers. Different results ensue if the user assigns greater or lower importance to the helper cognitive style attribute, relative to the other helper attributes. This is illustrated in tables 5 and 6, where the cognitive style weighting from the above example has been manipulated to be low and high, respectively.

Table 5: Matchmaker ranking of helpers with cognitive style assigned a low weighting

Potential Helpers helper 1 helper 5 helper 3 helper 12 helper 8 helper 2 helper 10 helper 13

ranking 47.76 46.22 45.72 44.3 42.84 42.72 41.42 40.92 cognitive style

strong verbaliser bimodal verbaliser imager bimodal strong

verbaliser imager strong imager

If the helpee from Tables 3 and 4 had set the user model representation to 'unimportant' (lowest on the five-point scale) for the weighting of cognitive style of a helper, the ranking in Table 5 would have been calculated (assuming no other changes in the helpee model). A greater range of helper cognitive styles is now evident. The top five still contains three of the original recipients of the help request, as they scored high across the various attributes in general, but especially for self assessed knowledge level, where helpers 3 and 5 tied for the highest score of 7, and helper 1 scored 6. Helper 1 also scored higher than any other learner for user helpfulness (8) and scored relatively high for system helpfulness (6.25). Nearly all other users scored 5 – the default, as the question was early in the I-Help deployment. These three helpers also changed their user eagerness to the maximum (10), indicating their enthusiasm to help.

Page 10: Cognitive style is defined by Riding and Rayner (1998) as ...

The former helper 2 has moved to sixth position. Helper 4 from Table 3 has moved to ninth position, having a lower level of knowledge of the topic (2) than most of the other students higher on the list. As stated above, the helper now in position 5 had high scores for user and system eagerness and user knowledge level. The helper now in position 4 also had high scores for user and system eagerness (both 10) and knowledge level (5), and was online at the time of the request which was marked urgent, thus raising the importance of the readiness attribute.

Table 6: Matchmaker ranking of helpers with cognitive style assigned a high weighting

Potential Helpers helper 1 helper 2 helper 4 helper 3 helper 6 helper 9 helper 5 helper 15

ranking 72.76 67.72 65.92 65.72 58.085 57.34 55.72 54.92 cognitive style

strong verbaliser

strong verbaliser

strong verbaliser verbaliser verbaliser verbaliser bimodal verbaliser

If the helpee had awarded cognitive style a high weighting ('very important' – top of the five-point scale), the ranking shown in Table 6 would have resulted. Helper 5, although not appearing on the final list, was still quite high because of being online at the time of the urgent request, and because of having many high scoring attributes, which together still outweighed the cognitive style weighting. All other potential helpers at the top of the list were verbalisers or strong verbalisers. The helper now in position 8 had been fifteenth in the original ranking, where cognitive style had equal importance to the other attributes. Increasing the importance of a helper's cognitive style against the other attributes in the helpee's user model will usually result in the recipients of the request all having the most suitable cognitive style to answer the type of question asked, as in this example. This section has shown how, when considered important by a helpee, helpers can be selected according to their cognitive style – with those put forward having a cognitive style suited to answering the helpee's specific question. For helpees who afford little importance to cognitive style, however, cognitive style will have a negligible effect on the selection of helpers. This may occur, for example, if a student has an urgent request that may require advanced knowledge. In this case they may view readiness and knowledge level to be the most important attributes in a helper. In the more general case, a student who has rarely experienced difficulties that could be ascribed to a mismatch between cognitive style and instructional methods, may also view the cognitive style attribute as less important. Identifying cognitive style in I-Help Cognitive style data is of benefit in many circumstances for the matching of learning partners. If such data is to be used, however, cognitive style information about each student must be obtained and maintained. How can this be achieved? The choices include using: 1. a front-end questionnaire (e.g. Clarke, 1993; Kwok and Jones, 1995); 2. a questionnaire external to the application, whose results can be entered into I-Help (e.g. Soloman and

Felder, 1999); 3. a program aimed at measuring cognitive style, the results of which can be transferred to I-Help (e.g. Riding,

1998); 4. stereotypes based on the 'typical' style identified for students in different disciplines; 5. inference by I-Help, based on the success or otherwise of peer interactions to date. Each of these alternatives has positive and negative characteristics, which are discussed below. Front-end questionnaires Given our focus on Riding and Cheema's (1991) wholist-analytic and verbal-imagery cognitive style dimensions, any questionnaire developed for I-Help must encompass these attributes. Front-end questionnaires are potentially useful, since these are directly integrated with the application. However, a difficulty resides in persuading users to take the time to complete questionnaires before commencing their use of a learning support system. This is especially true in the context of I-Help, since it is designed as a means for learners to find answers to their problems quickly. Thus, lengthy questionnaires may not actually be completed by many individuals. Requiring the submission of cognitive style data might result in many users avoiding the system for this factor alone – having to spend time identifying cognitive style before students can ask their first pressing question at the very time they are experiencing some problem, is likely to have a negative effect. Clarke (1993) undertook an investigation with Ford's (1985) Study Preference Questionnaire, with the aim of identifying a single question to be used as a front-end indicator of an individual's serialist-holist style (Pask,

Page 11: Cognitive style is defined by Riding and Rayner (1998) as ...

1976), to enable configuration of presentation for an individual. This approach clearly avoids the problem of requiring students to complete a lengthy questionnaire. However, the extent to which such attributes can be measured with a single question is in doubt. Indeed, Clarke identified a different question to be the most reliable, than that suggested by Ford. Subsequently, Kwok and Jones (1995) used a 13-question adaptation of Ford's original 18 items, as identified by Clarke to produce greater reliability. Nevertheless, even with this version no significant differences were observed between presentation format and matched and mismatched learning style groups. Thus, while a short integrated questionnaire could be practical for I-Help, it may be difficult to find an acceptable compromise between questionnaire length and questionnaire dependability. For a short questionnaire, this should be an initial measure applicable only until sufficient information is obtained through alternative means to verify or disconfirm the initial data. External questionnaires Similar problems exist for external questionnaires as for front-end ones that are integrated with the system: i.e. accuracy of results versus completion time. However, additional activity will be demanded from students as they will have to transfer their results into I-Help. Furthermore, while an integrated questionnaire will be understood as related to I-Help, an external one may not be obviously relevant from the perspective of a given group of I-Help users. Soloman and Felder's (1999) 44-item Index of Learning Styles Questionnaire (ILS) provides such an example. The ILS measures four dimensions: active-reflective, sensing-intuitive, verbal-visual, global-sequential. It can be completed and evaluated on the web, or printed and evaluated according to scoring instructions, thus allowing the flexibility necessary for a large-scale system such as I-Help (which in its most recently completed version had deployments in Canada, the U.K. and France). The ILS is also freely accessible, a second requirement for our distributed context. The third requirement, of considering wholist-analytic and verbal-imagery dimensions, is also met (albeit with different terminology). Nevertheless, many of the ILS questions, while possibly identifying style adequately – the authors acknowledge scarcity of data on validity (Felder, n.d.) – may not be perceived as relevant to I-Help users. For example, the following might be accepted as relevant to the academic context, but perhaps only to literature courses:

When I'm analyzing a story or novel (a) I think of the incidents and try to put them together to figure out the themes (b) I just know what the themes are when I finish reading and then I have to go back and find the incidents

that demonstrate them

Other questions may be viewed as even less relevant:

When I meet people at a party, I am more likely to remember (a) what they looked like (b) what they said about themselves

This lack of obvious relevance of such questions to their subject of study, at least among most of the students in I-Help’s current deployments, may negatively affect student motivation to complete such a questionnaire. In addition to the problem of perceived relevance in our context, applicable to both integrated and non-integrated questionnaires is a difficulty with questions such as the above that force a polar choice. One student in our study (see below) reported having notable trouble deciding which answer was most appropriate for each of the ILS visual-verbal questions. In every case he eventually chose the visual option, but his difficulty indicates that in reality he is probably bimodal. Thus, questionnaires operating a Likert scale may be more appropriate. Cognitive style identification programs In an attempt to overcome the subjectivity of self-report questionnaires, Riding (1991) developed the Cognitive Styles Analysis program (CSA). The CSA provides an individual’s position on the cognitive style construct. With our concentration on Riding and Cheema's (1991) WA and VI style dimensions, the components of the cognitive style construct, the CSA was an obvious candidate for consideration for the identification of cognitive style in the I-Help context. The CSA has three components: true/false judgement of appearance and conceptual category statements; same/different judgement of complex geometrical figures; judgement of whether a simple geometrical figure is contained within a complex one. Response time ratios are used to indicate an individual's cognitive style. It has been argued that the CSA is a better approach to measuring cognitive style since it directly evaluates task performance, as opposed to questionnaire approaches that evaluate strategies manifested by the underlying style (Riding and Rayner, 1998; Riding and Sadler-Smith, 1992).

Page 12: Cognitive style is defined by Riding and Rayner (1998) as ...

Nevertheless, there are also problems with the CSA. It takes about 10 minutes to complete which, while acceptable in many contexts, is possibly too long for many I-Help users. In I-Help students usually seek answers in real time, and responding to a series of questions at the time they need help would be a major deterrent to using I-Help. Making a prior investment of time to complete the CSA before help is actually needed, is also unlikely to be seen by students as appealing when no immediate gain can be expected. Completion time is not the only problem with the CSA in the I-Help context. The CSA has similar problems to the external questionnaires in that the results obtained would need to be transferred to I-Help. Furthermore, availability of the CSA may also be a problem when there are thousands of users at different locations. Each organisation would need to purchase the CSA program. Perceived relevance is also a concern here – the role of the geometrical components of the CSA will not be obvious to students, without reading the accompanying documentation. Even if they did this, the relevance to I-Help would need to be explained in order for students to appreciate its utility. Moreover, even the textual section does not appear, on the surface, related to the academic context – with questions such as the following (requiring a response of 'wrong' or 'right', indicated by a single key depression):

Bacon and lawyer are the same type Fire engine and strawberry are the same colour

Not only are such items likely to be perceived as irrelevant by students using I-Help, a more general concern with the CSA is its choice of some of the comparisons. Several subjects in our study (see below) spontaneously reported assessments on individual items that they felt were 'unfair' (feedback is provided). For example, the instructions state that "'ball and tennis are the same type' is wrong because they are not both sports". Item 40 of the CSA reads: "basketball and swimming are the same type". One subject interpreted 'basketball' as referring to the ball, and not the sport, as in the example in the instructions. However, 'basketball' was here intended to refer to the sport, and this subject therefore received the feedback that their answer was incorrect. Other subjects disagreed with the CSA's interpretation of whether 'cream' and 'paper' are the same colour; similarly, 'omelette' and 'waffle'. Applicable to our context is the 'fire engine'/'strawberry' example above. While the CSA expects a positive response, local fire 'trucks' are lime-coloured. Furthermore, some subjects found the instructions for the section on whether a simple geometrical figure is contained in a more complex one, confusing – they interpreted this to mean that they should answer 'yes' if the simple figure was within the outside boundary of the complex figure, instead of whether it appeared as part of the complex figure. Even if some of the CSA questions and instructions were modified, other points such as completion time, perceived relevance, and in particular, availability of the CSA, would still be a problem for I-Help. Thus, despite comprehensive questioning in the categories listed above, and the use of response time ratios to calculate an individual’s position on the cognitive style construct, the CSA does not appear suitable for use in conjunction with I-Help. Stereotypes Stereotypes (Rich, 1983) have long been applied in user modelling. In I-Help, the use of stereotypes would be appropriate if there were sufficiently strong tendencies for learners of a particular discipline or specialisation to have similar cognitive styles. Indeed, there has been some research into this issue in the context of learning style, based upon Kolb's (1984) model. Drew et al. (1998) found 86.5% of their (52) basic surgical trainees to be convergent or accommodative, according to Kolb's model: i.e. they were oriented towards problem-solving and practical involvement. Differences in learning style preferences have also been identified for business students, according to their major field of study: finance students preferred the abstract (as measured by Kolb's Learning Style Inventory); marketing students preferred the concrete; and accounting students tending to have quite balanced learning styles (Baldwin and Reckers, 1984; Brown and Burke, 1987). Further, Lonka and Lindblom-Ylänne (1996) found that students from two disciplines differently understood the concept of learning: medical students tended to view learning as intake of knowledge, whereas psychologists had a more constructivist view. Default stereotypes in I-Help would need to reflect research findings relating to the user's major academic subject, which would need to be supplied by the student or course tutor. This approach would be most beneficial in deployments where courses comprise students from multiple disciplines. Nevertheless, with the domain-independent nature of I-Help (examples have so far been described for computer science (Greer et al., 2001) and medical education (Greer and Bull, 2000), and I-Help has also been deployed in law and language domains), not all future user types will be studying subjects or specialisations where rigorous cognitive style studies have been performed. Furthermore, since some individuals in a particular course or program will deviate from any identified 'norm', initialised stereotypes of this kind will be set incorrectly for such students. As was argued for the case of short questionnaires, stereotypes would need, then, to be used only as an approximation, to be improved as more data becomes available for individuals.

Page 13: Cognitive style is defined by Riding and Rayner (1998) as ...

I-Help inference Once a student has been involved in several help sessions, patterns relevant to cognitive style may begin to emerge for that individual. For example, subsequent to the completion of a help session, each member of a pair evaluates their partner. The helpee's evaluation form includes questions on the knowledge level, utility of the help, and helpfulness of the helper. If, for example, helpees consistently rate a particular helper higher for their responses to questions preferring a wholist helper than for questions better answered by analytics, or if it is mostly wholists who add the helper to their preferred helper list, this might suggest that the helper is a wholist. Thus, responses to evaluation questions across all evaluations, and in comparison with the cognitive style representation in the helper and helpee user models, could be taken as an indication of whether an existing representation is likely to be accurate. There is a danger, however, that the representations in the models of peer evaluators could also be incorrect. Hence, changes to a user model based on peer evaluations could in theory be based on faulty data. Nevertheless, in order for serious inaccuracies to accrue, many different peer evaluators would have to have similarly inaccurate user model representations in order to effect an inappropriate change in any particular individual's user model. This is increasingly unlikely as the number of interactions grows and the number of peers providing the evaluations becomes large. Since I-Help is intended to be used long term by large numbers of students, it seems likely that over time inaccuracies would be weeded out. However, for the early interactions of any individual there will be no reliable cognitive style representations inferred for their user model. Thus a method of acquiring initial representations is still required. Representing initial cognitive style in I-Help The previous sections illustrated the difficulty of finding an adequate method of identifying individuals' cognitive styles for our context. Self-report questionnaires, cognitive style identification programs or stereotypes may result in inaccurate user model representations, and therefore must be considered subject to later updating. I-Help inferencing based on peer evaluations will only be reliable once sufficient evaluations have been completed. Hence some combination of the above is required. There is, as yet, insufficient evidence concerning the cognitive styles of students following different subject areas, for the application of stereotypes as an initial approximation, to a domain-independent system such as I-Help. The CSA program takes too long to complete for this context, and furthermore, users may not have access to the program. Therefore a short questionnaire approach was taken, to obtain the initial cognitive style instantiation for the user models. A provisional short 4-item questionnaire (I-HQ) was developed as a means of encouraging completion, the results of which remain open to subsequent revision. This is necessary because the validity of the I-HQ is not claimed – a study was conducted to compare the results of the I-HQ with Soloman and Felder's (1999) ILS and Riding's (1998) CSA (amongst other instruments). Subjects were 7 faculty, 11 graduate students and 21 undergraduates: T = 39 (one subject's results were discarded subsequent to her admission of having responded randomly to some questions). Problems have already been noted for the use of these instruments in conjunction with I-Help and, not surprisingly, no statistically significant results were obtained using Pearson's product-moment correlation coefficient. (For I-HQ and the relevant parts of ILS for the WA dimension r = .40, for the VI dimension r = .60; for I-HQ and CSA for the WA dimension r = -.13, for the VI dimension r = .36; for CSA and the relevant parts of ILS for the WA dimension r = -.27, for the VI dimension r = -.03). The I-HQ is composed of questions clearly related to the (current) I-Help academic contexts, to secure perceived relevancy from the viewpoint of students. Two questions were designed to relate to Riding and Cheema's (1991) wholist-analytic dimension, and two to their verbal-imagery dimension. While not a compulsory part of the system, the I-HQ is integrated in a manner that encourages students to contribute information directly to their user model alongside other information they deem important (see Figure 1). The provisional I-HQ is shown in Figure 6.

Page 14: Cognitive style is defined by Riding and Rayner (1998) as ...

Figure 6: The provisional I-Help Cognitive Styles Questionnaire The first and third questions relate to the wholist-analytic dimension; the second and fourth to the verbal-imagery dimension. For each dimension the combination of responses to the relevant questions results in the provisional identification of individuals as: strong wholists, wholists, intermediates, analytics, strong analytics; and strong verbalisers, verbalisers, bimodals, imagers, strong imagers. Once this attribute is initialised in the user model it becomes subject to change if appropriate, as a result of increasing external evidence from peer evaluations. Although not shown to be statistically significant, this short questionnaire may be good enough for I-Help purposes if appropriate inferences about each student (as discussed earlier) can be drawn using the questionnaire categorisations as a starting point. Our future work on cognitive styles will investigate this issue in detail, starting with a more extensive investigation of suitable questions for the I-HQ, and a consideration of the number of questions to include. However, even with an improved I-HQ, the need is still for a relatively short questionnaire to sustain a high completion rate. Thus, whatever the eventual form, the I-HQ will remain a starting point for obtaining cognitive style representations. I-Help will subsequently use a student's answers to the I-HQ (and other information it has about the student) as an index into appropriate stereotypes that will include the initial representation of the student's cognitive style. Thus, rather than using pre-assigned stereotypes for an individual according to their domain of study which, as indicated previously, is infeasible, stereotypes will be assigned according to, amongst other information, I-HQ responses. It is not necessary that initial stereotypes be absolutely correct, since stereotype values can be refined over time through interactions with a variety of users (Rich, 1989). In the long term, with thousands of students using I-Help, the stereotypes will evolve and should become increasingly accurate, as user model fragments are combined and recombined until stable stereotypes emerge. For example, it might be found that a 'helpful verbaliser', while frequently selected as a helper for other verbalisers, and rated highly by them, might receive lower ratings for the helpfulness attribute from imager helpees. The likelihood of evaluations from imagers being fewer in number is high, given that cognitive style matching is often sought on the verbal-imagery dimension. Thus, the helpfulness measure of a potential helper, while appropriate for verbaliser helpees, could become inappropriately high with reference to imager helpees. The evolution of stereotypes over time, from recombining user model fragments, will accommodate such occurrences. With many peer-peer interactions on which to draw, it should become easier to assign new students to appropriate stereotypes based on interactions of previous students, as has been suggested elsewhere in student-system interaction contexts (Kay, 1995; Milne et al., 1997; Winter and McCalla, 1999). Of course, however accurate a stereotype may be for a general class of students, any given student may not be completely stereotypical. Moreover, the knowledge the system has about a given student is continuously being updated, with new information being gained as the student uses I-Help, and old information perhaps being found to be outdated or even wrong. This means that there must be on-going maintenance of the user model, even up to the need to change the assigned stereotype. Traditional belief revision techniques (as in de Kleer, 1986; Doyle, 1979, for example) will not be adequate to carry out this maintenance because they do not work for fragmented models and do not incorporate stereotypes into their reasoning. A good starting point for a belief revision system with the appropriate characteristics would be Huang's (1993) approach, which assumes that beliefs are fragmented into many separate 'frames of mind', only some of which are in focus at any given time. Moreover, Huang also incorporates stereotypes and provides algorithms for changing stereotypes if beliefs that are in focus contradict the current stereotype (Huang et al., 1991). Recent work on Bayes Nets (e.g. Horvitz,

Page 15: Cognitive style is defined by Riding and Rayner (1998) as ...

1997; VanLehn et al., 1998) might also be usefully incorporated into Huang's approach, since Bayes Nets are an excellent means of tracking dependencies amongst knowledge and are an efficient way of propagating the side effects of change should refinements in the initial user model be required. Using techniques such as these, it should be possible for I-Help to evolve an increasingly accurate representation of each student's cognitive style and other characteristics, and interactions between them, without requiring the user to spend a great amount of time inputting information themselves which, in the I-Help context, is unrealistic. Conclusions and future directions This paper has described how the I-Help user models are used to help match peers for one-on-one help sessions, focussing on the cognitive style component of the models. The importance of cognitive style is differentially applicable. Learners who are not concerned about cognitive style (for example, intermediate-bimodals may have had little experience with problems adapting to an explanation or resource) may set their cognitive style component to 'unimportant', thus reducing its weight compared to the attributes considered more crucial. Similarly, students who have had bad learning experiences that they attribute to teaching style or unsuitable materials, etc., may increase the weight of the cognitive style aspect. I-Help is primarily deployed amongst large cohorts. In such contexts there is much scope for including many factors in the user models of individuals. Indeed, the more attributes recorded, the more useful the matchmaker's ranking is likely to be. An issue such as cognitive style is particularly relevant since, instead of being one of many (often static) preferences of the helpee, its relevance is to the helpee's question type, where for some questions and some helpees, cognitive style might be an indicator of the extent of success of an interaction. This applies particularly in our context where peers are not trained helpers. If I-Help is adopted across a whole course in an academic milieu or even a whole career in a workplace setting, it will become feasible to use a more detailed measuring instrument. I-Help could serve as an experimental testbed for obtaining information about cognitive style, a testbed that is completely authentic and non-artificial. Different instruments could be tested in this context. With I-Help deployments planned that will incorporate dozens of courses, in many different subjects, with thousands of users, some using the system over several years, the amount of information about cognitive style and other attributes lurking in the I-Help database is potentially huge. Not only will it be possible to examine cognitive style itself, but there may be interesting interactions between attributes kept in the fragmented I-Help user models. The stereotypes that will eventually emerge from the distributed user model fragments will be of great interest. Finally, pedagogical implications of cognitive style may emerge, as situations in which a help session was successful (or not) are correlated with the cognitive styles of the helper and helpee. This should add to the general debate on the application of cognitive style information in education. Should the I-Help experiments with cognitive style confirm its value, then cognitive style may become a standard attribute in user modelling to be used well beyond I-Help. Modelling cognitive style may prove useful in intelligent tutoring systems generally, perhaps even more so than in I-Help, since this information would allow the system to choose an approach to teaching the student that is appropriate to the student's cognitive style and the nature of the concept being learned. Notes 1 There have been 4 major deployments of I-Help to date (see Greer et al., 2001), and only in the last two has cognitive

style information been gathered about students. The examples in this section are drawn from two of these deployments, as appropriate to the point under discussion.

Acknowledgements This work was funded by the Canadian TeleLearning Network of Centres of Excellence and the Natural Sciences and Engineering Research Council of Canada. Thanks to Lori Kettel and Mike Winter for programming the cognitive style component of the I-Help user models. I-Help was developed in the ARIES Lab, Department of Computer Science, University of Saskatchewan. A large number of people have been involved in the I-Help project. In particular, Jim Greer, Julita Vassileva and Ralph Deters should be acknowledged for their major leadership role. Our gratitude also goes to the many graduate students and summer students who have contributed ideas to I-Help, the many undergraduate students who have used various prototypes over the years, and in particular to the experimental subjects who helped us explore the I-HQ questionnaire and who used the versions of I-Help that incorporated cognitive style.

Page 16: Cognitive style is defined by Riding and Rayner (1998) as ...

References Baldwin, B.A. and Reckers, P.M.J. (1984). Exploring the Role of Learning Style Research in Accounting Education Policy.

Journal of Accounting Education 2(2): 63-76. Barker, T., Jones, S., Britton, C. and Messer, D. (2000). Individual Cognitive Style and Performance in a Multimedia

Learning Application. Presented at EURO Education 2000 Conference. Aalborg, Denmark. Brown, H.D. and Burke, R.C. (1987). Accounting Education: A Learning Styles Study of Professional-Technical and Future

Adaptation Issues. Journal of Accounting Education 5: 187-206. Bull, S. (1997). A Multiple Student and User Modelling System for Peer Interaction. In R. Schäfer and M. Bauer, eds, 5 GI-

Workshop, Adaptivität und Benutzermodellierung in interaktiven Softwaresystemen, pp. 61-71. Universität des Saarlandes. Online proceedings: http://w5.cs.uni-sb.de/~abis97/proceedings/bull.ps

Bull, S. and Greer, J. (2000). Peer Help for Problem-Based Learning. In S.S. Young, J. Greer, H. Maurer and Y.S. Chee, eds, Proceedings of ICCE/ICAI 2000: Learning Societies in the New Millenium, Creativity, Caring and Commitments (Volume 2), pp. 1007-1015. National Tsing Hua University of Taiwan.

Bull, S., Greer, J., McCalla, G. and Kettel, L. (2001a). Help-Seeking in an Asynchronous Help Forum. In R. Luckin and B. du Boulay, eds, Proceedings of Workshop on Help Provision and Help Seeking in Interactive Learning Environments, pp. 9-21. International Conference on Artificial Intelligence in Education.

Online proceedings: http://www.cogs.susx.ac.uk/users/bend/aied2001/bull.pdf Bull, S., Greer, J., McCalla, G., Kettel, L. and Bowes, J. (2001b). User Modelling in I-Help: What, Why, When and How. In

M. Bauer, P.J. Gmytrasiewicz and J. Vassileva, eds, User Modeling 2001, pp. 117-126. Berlin Heidelberg: Springer-Verlag. Also available from: http://www.eee.bham.ac.uk/bull/papers/UM01.html

Clapp, R.G. (1993). Stability of Cognitive Style in Adults and Some Implications, A Longitudinal Study of the Kirton Adaption-Innovation Inventory. Psychological Reports 73: 1235-1245.

Clarke, J. (1993). Cognitive Style and Computer-Assisted Learning: Problems and a Possible Solution. Association for Learning Technology Journal 1(1): 47-59.

Cox, R. (1999). Representation, Construction, Externalised Cognition and Individual Differences. Learning and Instruction 9: 343-363.

de Kleer, J. (1986). An Assumption-Based Truth Maintenance System. Artificial Intelligence Journal 28(2): 127-162. Doyle, J. (1979). A Truth Maintenance System. Artificial Intelligence Journal 12: 231-272. Drew, P.J., Cule, N., Gough, M. Heer, K., Monson, J.R.T., Lee, P.W.R., Kerin, M.J. and Duthie, G.S. (1998). Optimal

Education Techniques for Basic Surgical Trainees: Lessons from Education Theory. Journal of the Royal College of Surgeons of Edinburgh 44: 55-56.

Felder, R. (not dated). Index of Learning Styles (ILS). http://www2.ncsu.edu/unity/lockers/users/f/felder/public/ILSpage.html. Accessed: 27th March 2001.

Ford, N. (1985). Learning Styles and Strategies of Postgraduate Students. British Journal of Educational Technology 16(1): 65-79.

Greer, J. and Bull, S. (2000). Computer Support for Collaboration in Medical Education. Clinical and Investigative Medicine 23(4): 270-274.

Greer, J., McCalla, G., Cooke, J., Collins, J., Kumar, V., Bishop, A. and Vassileva, J. (1998). The Intelligent Helpdesk: Supporting Peer Help in a University Course. In B.P. Goettl, H.M.Halff, C.L. Redfield, V.J. Shute, eds, Intelligent Tutoring Systems, pp. 494-503. Berlin Heidelberg: Springer-Verlag.

Greer, J., McCalla, G., Vassileva, J., Deters, R., Bull, S. and Kettel, L. (2001). Lessons Learned in Deploying a Multi-Agent Learning Support System: The I-Help Experience. In J.D. Moore, C.L. Redfield and W.L. Johnson, eds, Artificial Intelligence in Education, pp. 410-421. Amsterdam: IOS Press.

Groat, A. and Musson, T. (1995). Learning Styles: Individualizing Computer-Based Learning Environments, Association for Learning Technology Journal 3(2): 53-61.

Hoppe, H.U. (1995). The Use of Multiple Student Modelling to Parameterize Group Learning. In J. Greer, ed, Proceedings of World Conference on Artificial Intelligence in Education, pp. 234-241. Charlottesville, VA: AACE.

Horvitz, E. (1997). Agents with Beliefs: Reflections on Bayesian Methods for User Modeling. In A. Jameson, C. Paris and C. Tasso, eds, Proceedings of the Sixth International Conference on User Modeling, pp. 441-443. Wien, New York: Springer.

Huang, X. (1993). Inconsistent Beliefs, Attention, and Student Modelling. International Journal of Artificial Intelligence in Education 3(4): 417-428.

Huang, X., McCalla, G.I., Greer, J.E. and Neufeld, E. (1991). Revising Deductive Knowledge and Stereotypical Knowledge in a Student Model. User Modeling and User-Adapted Interaction 1 (1): 87-115.

Kay, J. (1995). The UM Toolkit for Cooperative User Modelling. User Modeling and User-Adapted Interaction 4: 149-196. Kolb, D.A. (1984). Experiential Learning. Englewood Cliffs, NJ: Prentice Hall. Kwok, M. and Jones, C. (1995). Catering for Different Learning Styles. Association for Learning Technology Journal 3(1):

5-11. Laurillard, D. (1979). The Processes of Student Learning. Higher Education 8: 395-409. Lonka, K. and Lindblom-Ylänne, S. (1996). Epistemologies, Conceptions of Learning, and Study Practices in Medicine and

Psychology. Higher Education 31: 5-24. McCalla, G., Vassileva, J., Greer, J. and Bull, S. (2000). Active Learner Modelling. In G. Gauthier, C. Frasson and K.

VanLehn, eds, Intelligent Tutoring Systems. pp. 53-62. Berlin Heidelberg: Springer-Verlag. McKay, E. (1999a). An Investigation of Text-Based Instructional Materials Enhanced with Graphics. Educational

Psychology 19(3): 323-335.

Page 17: Cognitive style is defined by Riding and Rayner (1998) as ...

McKay, E. (1999b). Exploring the Effect of Graphical Metaphors on the Performance of Learning Computer Programming Concepts in Adult Learners: A Pilot Study. Educational Psychology 19(4): 471-487.

McKay, E. (2000). Measurement of Cognitive Performance in Computer Programming Concept Acquisition: Interactive Effects of Visual Metaphors and the Cognitive Style Construct, Journal of Applied Measurement 1(3): 257-291.

Milne, S., Cook, J., Shiu, E. and McFadyen, A. (1997). Adapting to Learner Attributes: Experiments Using an Adaptive Tutoring System. Educational Psychology 17(1-2): 141-155.

Moran, A. (1991). What can Learning Styles Research Learn from Cognitive Psychology? Educational Psychology 11(3-4): 239-245.

Mühlenbrock, M., Tewissen, F. and Hoppe, H.U. (1998). A Framework System for Intelligent Support in Open Distributed Learning Environments. International Journal of Artificial Intelligence in Education 9(3-4): 256-274.

Newble, D.I. and Hejka, E.J. (1991). Approaches to Learning of Medical Students and Practising Physicians: Some Empirical Evidence and its Implications for Medical Education. Educational Psychology 11(3-4): 333-342.

Ogata, H., Sueda, T., Furugori, N. and Yano, Y. (1999). Augmenting Collaboration Beyond Classrooms through Online Social Networks. In G. Cumming, T. Okamoto and L. Gomez, eds, Advanced Research in Computers and Communications in Education – Proceedings of ICCE '99, pp. 277-284. Amsterdam: IOS Press.

Pask, G. (1976). Styles and Strategies of Learning. British Journal of Educational Psychology 46: 128-148. Pillay, H. (1998). An Investigation of the Effect of Individual Cognitive Preferences on Learning through Computer-Based

Instruction. Educational Psychology 18(2): 171-182. Ramsden, P. (1979). Student Learning and Perceptions of the Academic Environment. Higher Education 8: 411-427. Ramsden, P. and Entwistle, N.J. (1981). Effects of Academic Departments on Students' Approaches to Studying. British

Journal of Educational Psychology 51: 368-383. Rich, E. (1983). Users are Individuals: Individualizing User Models. International Journal of Man-Machine Studies (18):

199-214. Rich, E. (1989). Stereotypes and User Modeling. In A. Kobsa and W. Wahlster, eds, User Models in Dialog Systems, pp. 35-

51. Berlin: Springer-Verlag. Riding, R. (1991). Cognitive Styles Analysis. Birmingham: Learning and Training Technology. Riding, R. (1998). Cognitive Styles Analysis. Birmingham: Learning and Training Technology. Riding, R.J. and Ashmore, J. (1980). Verbaliser-Imager Learning Style and Children's Recall of Information Presented in

Pictorial Versus Written Form. Educational Studies 6(2): 141-145. Riding, R.J., Buckle, C.F., Thompson, S. and Hagger, E. (1989). The Computer Determination of Learning Styles as an Aid

to Individualized Computer-Based Training. Educational and Training Technology International 26(4): 393-398. Riding, R. and Caine, T. (1993). Cognitive Style and GCSE Performance in Mathematics, English Language and French.

Educational Psychology 13(1): 59-67. Riding, R.J. and Calvey, I. (1981). The Assessment of Verbal-Imagery Learning Styles and their Effect on the Recall of

Concrete and Abstract Prose Passages by 11 Year Old Children. British Journal of Psychology 72: 59-64. Riding, R. and Cheema, I. (1991). Cognitive Style – an Overview and Integration. Educational Psychology 11(3-4): 193-

215. Riding, R. and Douglas, G. (1993). The Effect of Cognitive Style and Mode of Presentation on Learning Performance.

British Journal of Educational Psychology 63: 297-307. Riding, R.J., Glass, A., Butler, S.R. and Pleydell-Pearce, C.W. (1997). Cognitive Style and Individual Differences in EEG

Alpha During Information Processing. Educational Psychology 17(1-2): 219-234. Riding, R. and Rayner, S. (1998). Cognitive Styles and Learning Strategies. London: David Fulton Publishers. Riding, R. and Sadler-Smith, E. (1992). Type of Instructional Material, Cognitive Style and Learning Performance.

Educational Studies 18(3): 323-340. Riding, R.J. and Watts, M. (1997). The Effect of Cognitive Style on the Preferred Format of Instructional Material.

Educational Psychology 17(1-2): 179-183. Sadler-Smith, E. and Riding, R. (1999). Cognitive Style and Instructional Preferences. Instructional Science 27: 355-371. Soloman, B.A. and Felder, R.M. (1999). Index of Learning Styles Questionnaire. http://www2.ncsu.edu/unity/lockers/users/f/felder/public/ILSdir/ilsweb.html Valley, K. (1997). Learning Styles and Courseware Design. Association for Learning Technology Journal 5(2): 42-51. VanLehn, K., Niu, Z., Siler, S. and Gertner, A.S. (1998). Student Modeling from Conventional Test Data: A Bayesian

Approach without Priors. In B.P. Goettl, H.M. Halff, C.L. Redfield and V.J. Shute, eds, Intelligent Tutoring Systems, pp. 434-443. Berlin Heidelberg: Springer-Verlag.

Vassileva, J., Greer, J., McCalla, G., Deters, R., Zapata, D., Mudgal, C. and Grant, S. (1999). A Multi-Agent Design of a Peer-Help Environment. In S. Lajoie and M. Vivet, eds, Artificial Intelligence in Education, pp. 38-45. Amsterdam: IOS Press.

Vivacqua, A.S. (1999). Agents for Expertise Location, Proceedings of the AAAI Spring Symposium on Intelligent Agents in Cyberspace, Stanford, CA.

Westman, A.S. (1993). Learning Styles are Content Specific and Probably Influenced by Content Areas Studied. Psychological Reports 73: 512-514.

Winter, M. and McCalla, G. (1999). The Emergence of Student Models from and Analysis of Ethical Decision Making in a Scenario-Based Learning Environment. In J. Kay, ed, UM99 User Modeling, pp. 265-274. Wien, New York: Springer.

Witkin, H.A., Moore, C.A., Goodenough, D.R. and Cox, P.W. (1977). Field-Dependent and Field-Independent Cognitive Styles and their Implications. Review of Educational Research 47: 1-64.