Main Applicant: Dillenbourg, Pierre

Part 2. Scientific Information

Main Applicant: Dillenbourg, Pierre

Project Title:The effects of modelling on collaborative learning

1.Summary

This proposal is part of a European Collaborative project entitled Scaffolding, structuring and regulating collaborative learning for knowledge construction and sharing . It has been submitted to the European Science Foundation. It includes research labs in The Netherlands, Germany, Norway and Finland. It defines an ambitious and long-term (4 years) partnership.

When students solve a problem collaboratively, they maintain some representation of their partners' goals, knowledge and understanding. This mutual model may be not highly detailed, nor explicit, but is still necessary to build a shared understanding. An alternative view is that learners construct a model of the group interactions as a whole. In both cases, this modeling increases the co-learners' cognitive activity. Our main hypothesis is that this modeling activity leads co-learners to think more deeply about the task and hence to improve their knowledge.

This hypothesis raises methodological difficulties. As soon as one asks a learners what her partner knows, we bias the spontaneous modeling process. Measuring the modeling activity after the task raises other memory biases. Therefore, this project aims to develop new methodological approaches for measuring and manipulating the mutual and group models. These approaches borrow technologies from the field of computer-supported collaborative learning (CSCL).

Three types of CSCL artifacts will be used to investigate modeling activities.

• Scripts are sequences of activities that structure teamwork by defining phases, roles and interactions mode. They may play the role of prosthesis for group modeling.
• Awareness tools are software components that inform about partner's activities and hence facilitate mutual modeling.
• Group mirrors are graphical representations of group actions and interactions and hence support group modeling.

This project aims to answer the following research questions:

1.Which representations do collaborative learners build of their partner (mutual modeling) and of the group (group modeling)? We focus on the cognitive aspects rather than on social or emotional dimensions of these models.

2.Does the process of mutual/group modeling enhance the knowledge acquired by learners?

3.To which extent do the co-learners remember interactions episodes and their relationship to mutual/group modeling?

4.Do CSCL scripts induce more similar mutual and group models?

5.Do CSCL group mirrors and awareness tools lead to build more accurate mutual and group models?

2.Research Plan

2.1.State of research

The goal of our research is to develop the understanding of the mechanisms of collaborative learning in order to improve the existing didactic approaches to make collaborative learning more effective. Therefore, this review is structured in two parts: the first part focuses on the socio-cognitive processes that create knowledge during collaborative interactions; the second part focuses on methods for augmenting the probability that these processes do actually occur during collaborative interaction.

This research belongs to a field referred to as 'computer-supported collaborative learning' (CSCL). This scientific community, which we contributed to launch in the nineties, is now well established, having its own conference series (10), its books series, its society ( and soon its journal. We stress that CSCL is not simply about distance education. Of course, computers and networks enable collaborative learning among students who are in different places and/ or at different times. However, since its very beginning, the research agenda of this community has been to use information technologies to deepen research on learning mechanisms in general and to develop tools that enhance collaborative learning, including when learners are collocated.

A main challenge of CSCL is to open the world of learning technologies, traditional by based on individualistic views of human learning, to the social nature of cognition.

2.1.1.Understanding collaborative learning.

Why would student learn better in groups? How could knowledge emerge from the interactions among students who don't have this knowledge? Initial empirical studies have shown that collaborative learning is often more effective than learning alone, but not systematically (Slavin, 1983). Therefore, a second generation of studies attempted to determine under which conditions collaborative learning would be effective: task features, group size, group heterogeneity, medium features,… (Dillenbourg et al., 1996). However, too many parameters need to be controlled for predicting effects and moreover these parameters interact with each other in a complex way. Therefore, the third generation of studies does not try to control the effectiveness of collaborative learning as if it was a black box but zooms in the collaborative process in order to grasp which interactions produce learning and when these interactions do occur. In other words, the effects of collaborative learning depend on the quality of interactions during collaboration. Several types of interactions have been studied such as the quality of explanations (Webb, 1991), mutual regulation (Blaye, 1988), argumentation (Baker, 1992), conflict resolution (Doise, Mugny & Perret-Clermont, 1975). These various types of interactions have in common that they lead students to verbalize knowledge that would otherwise remain tacit. The methods described in the next section to enhance collaborative learning aim to foster these productive interactions.

Roschelle & Teasley (1995) described collaborative interactions as the process of building and maintaining a shared understanding of the task that learners have to achieve. Since then, this field of research has been much influenced by psycholinguistic research on grounding (Clark & Brennan, 1991). As stressed by Clark, grounding is functional: peers don't reach a level of perfectly shared understanding but the degree of shared understanding that is necessary to perform the task together. This research field has also evolved from socio-cognitive to socio-cultural views on learning. An interesting example is Baker et al.'s (1999) articulation between the notions of grounding and the activity theory. A key point for our research project is that, as pointed out by Schwartz (1999), it is not the fact of having shared a common understanding that generates knowledge, but the effort that learners have to engage in order to construct this shared understanding. We hereafter refer to this as the cognitive effort to build a shared understanding.

The construction of a shared understanding requires that each partner builds some representation of the other partners' beliefs, knowledge or goals. We refer to the construction of this process as mutual modeling, one facet of intersubjectivity (Wertsch, 1985; Bromme, 2000) or audience design (Lockridge & Brennan, 2002). By describing it as a 'model', we do not imply this is a detailed representation of the partner, nor an explicit one. Simply stated, if A wants to (dis-)agree with B, A needs some representation of B's intentions; if A wants to repair B's misunderstanding, A needs some representation of what B understood. Mutual modeling is as functional as the grounding process: the degree of precision in mutual modeling depends on the task, it has to be extremely high if two pilots collaborate on landing a plane, as in Hutchins' (1995) observations, but can be much lower if they discuss about their last party (down to what politeness allows). Moreover, this mutual model is not constructed in a vacuum but is based on multiple inference mechanisms. Common grounds are initialized by the assumptions people make about their partner from cues such as his community membership (age, culture, profession, ...) and from co-presence (e.g. commons ground include any event to which A and B attended together) (Clark & Marshall, 1981). Several scholars studied how this initial model imparts on communication, namely because it can easily be manipulated. For instance, Slugoski et al. (1993) pretends to the subjects that their (fake) partner has received or not the same information. They observed that the subjects adapt to their partner by focusing the explanation on the items that (s)he is supposed to ignore. Brennan (1991) showed that the subjects used different initial strategies in forming queries depending on who they were told their partner was. Other simpler inference mechanisms such as default reasoning rules (e.g. B agrees with me unless he disagrees) are developed according to the conversational context. The mutual modeling could not occur independently from culturally acquired interaction schemas that constrain the space of interpretation. In our project, we use 'scripts' as explicit schema that constrains this space.

Even when mutual modeling is not detailed and explicit, reasoning on what one's partner believes involves some cognitive activity. For Clark & Wilkes-Gibbs (1986), what is important is not individual effort by the receiver of a communicative act, but the overall least collaborative effort. The cost of producing a perfect utterance may be higher than the cost of repairing the problems which arise. For instance, subjects are less careful about adapting utterances to their partner when they know they can provide feedback on her understanding (Schober, 1993). We introduced instead the notion optimal collaborative effort(Dillenbourg et al, 1996) to stress the fact that misunderstanding should not be viewed as something to be avoided (even if this was possible), but as an opportunity to explain, to justify, and so forth. We here enter into the global argument regarding cognitive load in learning activities, namely in discovery learning environments: there is no learning without cognitive load, but overload may hinter learning (Paas, Renkl & Sweller, 2003). In the context of collaborative learning, we understand the cognitive load induced by mutual modeling as part of Schwartz' notion of effort towards a shared understanding. This leads us to the global hypothesis that underlies the present research project: mutual modeling is one mechanism of learning in groups.

This hypothesis is difficult to investigate because the degree of mutual modeling is both difficult to manipulate, as an independent variable, and difficult to measure, as a dependent variable. Measuring is difficult because, as soon as one asks learners what their partner knows, we trigger a modeling process beyond what it was 'naturally'. Controlling is difficult because even in the studies on self-explanation effect, the subjects know that the experimenter will listen their explanations. We describe these difficulties in section 2. Therefore, an important goal of this project is to develop new methodological tools for manipulating and capturing mutual modeling.

Another hypothesis is that participants do not build a representation of their partners' mental states but instead a representation of the interaction process at the group level: instead of modelling who knows what, who does what, who said what, the team members could maintain a representation of what the team knows, has done or has said. We refer to this as the group model. This alternative is directly inspired by the distributed cognition theories (Pea, 1993; Salomon, 1993; Hutchins 1995). The two hypotheses are of course complementary since these two models feed each other. It is therefore our intention to investigate these two hypotheses in parallel.

2.1.2.Enhancing collaborative learning

The evolution of research on collaborative learning has obvious pedagogical implications. Instead of trying to control the conditions that determine the effectiveness of collaborative learning, CSCL designers attempt to (directly) influence the interactions: augmenting the frequency of conflicts, fostering elaborated explanations, supporting mutual understanding, ... Collaboration can be influenced anticipatively, by structuring the collaborative process in order to favour the emergence of productive interactions, or retroactively, by regulating interactions, as tutors do. These two complementary approaches are described below.

Regulating collaborative learning is a subtle art. The tutor has to provide prompts or cues without interfering with the social dynamics of the group. Light human tutoring is a necessary, but expensive resource for computer-supported collaborative learning (hereafter CSCL). Some of the current research aims to design computerized tutors (Inaba & Okamoto, 1996; Barros & Verdejo, 2000; Constantino-González & Suthers, 2002) or to develop tools that facilitate the group regulation by human tutors (Després, 2001). Group tutoring is not directly related to this research project. However, one extension is that tools designed for helping tutors to regulate groups could be given to the group for helping it to regulate. These tools provide the group with a representation of its own processes or interactions (Jermann, 2002; Dillenbourg et al., 2002, Zumbach et al, 2002). We refer to them as group mirrors. Figure 1 compares the representations of group interactions in two electronic forums (Donath, 1999). Group mirrors provide info that would not be available as such to learners. Hence, they constitute a way to play with the degree of mutual modeling and group modeling. This approach to group representations is further described in the next section.

Figure 1. A group mirror (Donath, 1999)

Structuring collaborative learning is achieved by semi-structured communication interfaces and/or by the application of scripts for collaborative learning.

A semi-structured communication interface is usually a text-based communication tool that supports the grounding mechanisms for instance by making explicit which utterance is being acknowledged, which object is being referred to or which type of speech act is being uttered. Examples of these are given by Baker & Lund (1997), Suthers et al., (1995); Jermann & Schneider, 1997 (Ref) and Veerman & Treasure Jones (1999).

Figure 2: Example of semi-structured interface: buttons in the bottom part offer pre-defined communication acts and sentence openers (Soller, 2002)

A collaboration script (Aronson et al, 1978; O'Donnell & Dansereau, 1992) is a set of instructions prescribing to how the group members should interact, how they should collaborate and/or how they should solve the problem. Fisher (REF) establishes a difference between socials scripts, that structure interactions between group members, and epistemic scripts, that structure interactions with the task. The collaboration process is not left open but structured as a sequence of phases, each phase corresponding to a specific task where group members have a specific role to play. A well-known script is the 'reciprocal teaching' approach set up by Palincsar & Brown (1984): one peer reads a text paragraph and the other questions him/her about his/her understanding, for the next paragraph roles are shifted. Many variations of this script exist such as peer tutoring (O'Donnell & Dansereau, 1992; Fantuzzo et al, 1989) or peer teaching (Reiserer, Ertl & Mandl, 2002).

The semi-structured interfaces structure interactions at the micro-level (at the utterance level) and scripts at the macro level (at the activity level) (Pfister, to appear). Both approaches convey a prescriptive model of group interactions: they do not capture how peers represent collaborative learning but provide them with such a representation. In some cases, the author intends that the group member internalize the script as a model of collaboration, while in other situations the students are simply expected to play it. While the effects of semi-structured interfaces have not been clearly established (Veerman & Treasure Jones, 1999), there exist reports of the effectiveness of CSCL scripts and scripting collaboration became a hot research topic (Kollar et al, 2003; Weingerger et al, to appear). Scripts offer the opportunity to turn basic research into classroom guidelines. At the same time, they also raise the risk of drifting away from the very idea of collaborative learning, by forcing interactions that do not match the social dynamics of groups, what we called 'over-scripting' (Dillenbourg, 2002). The assumption underlying this project is that the work on CSCL scripts would greatly benefit from a deeper understanding of how people represent collaborative learning and, vice-versa could provide researchers with ways to manipulate these group models. Therefore this project articulates basic research on mutual modeling and group modeling with more applied research on the effects of CSCL scripts.

This project also relies on research in computer-supported cooperative work (CSCW), a field that has always inspired CSCL. The design of shared editors was based on the WYISISIS principle: what you see is what I see; all group users have the same view (Stefik et al, 1987). This principle must be relaxed when the group includes more members or when the task is more complex. In these cases, users need different views. In order to sustain group coordination, designer created 'awareness tools' (Greenberg & Gutwin, 1998), illustrated in figure 3. These software components inform one partner about what the others are doing, where they are, etc.

As a green user, I see that my blue partner reads the same paragraph as me while my red partner is reading much below. /
When jointly editing this graph, we both see that we are looking at different subsets of the graph but with some overlap.

Figure 3: Awareness tools developed by Greenberg and Gutwin.(REF)

2.2.Participants Research

Our team has carried out research in CSCL since its beginning. Our current lab in EPFL has been created in 2002 but several team members (Jerman, Nova, Sangin, Girardin and Dillenbourg) have previously conducted research on this issue at TECFA, the educational technology unit within the Faculté de Psychologie et des Sciences de l'Education, Université de Genève. We report our recent activities with the same structure as the previous review. This report includes graduates theses, postgraduate theses, doctoral theses and projects funded by the NSF, by OFES and by private companies.

2.2.1.Understanding collaborative learning

Our research projects aims at getting a deeper understanding of collaboration process. For instance, we have seen pairs build more abstract representations than individuals, as Schwartz (1995) found before, but that these more abstract representations did not lead to higher transfer performance (Mondoux, Auderset & Dillenbourg, 2004). We are currently running experiments on the effectiveness of static versus dynamic pictures in individual and pair conditions (Bétrancourt, Dillenbourg & Montarnal, 2003). Our main question is whether the interaction effect between the presentation format (static/dynamic) and the social condition (solo/duo) is a matter of cognitive load or grounding mechanisms. We also explored if the learning is different gains were different when the learners have to tailor their explanations to the explainees, we –due to methodological difficulties, we did not find clear results (Ploetzner et al, 199)