Grounding in computer-supported collaborative problem solving
|
|  Yüklə 1.41 Mb.
|
Methodological issues
The evolution of research on collaborative learning and/or collaborative problem solving has evolved in three stages (Dillenbourg, Baker, Blaye & O'Malley, 1995). 1 The first generation of studies compared the performance of pairs with the performance of individuals, leading not surprisingly to contradictory results. 2 A second generation of experiments investigated in which conditions collaborative learning was more or less efficient. These study revealed numerous factors such as the composition of the group (number of subjects, objective and subjective heterogeneity, age, gender, ...), the nature of the task and the features of the communication medium. These independent variables interact with each other, creating potential nth order interaction effects. The complexity of such causal relationship is hardly tractable by classical experimental methods. 3 In the third generation of studies, the concept of 'collaboration' has exploded and leaves place for more precise description of interactions between the group members. For instance, Webb (1991) showed that the effects where not the same for all interacting pairs, but it was better for the group members proving elaborated explanations. Schwartz (1995) observed that pairs elaborate more abstract representation than individuals, probably because this representation serves to coordinate different viewpoints, to bridge different perspectives. In summary, when running experiments on collaborative problem solving, the word 'collaborative' refers to the experimental setting or to the instructions given to the subjects. If we want to account for the cognitive effects of interactions in collaborative interactions, one has to look exactly what these subjects did. Therefore, we have chosen an exploratory approach, based on a careful analysis of interactions.
One can observe a similar evolution of research on groupware. Many studies compared the use of groupware versus the non-use. For instance, McLeod (1992) reviewed 42 empirical studies on synchronous group support systems. Her meta-analysis shoes that such systems lead to an increased focus on the task, to more quality of participation for the group members, to take better decisions but after having spent more time and with less consensus. The difference between two groupware systems may be larger than difference between use versus non-use of groupware. For instance, O’Conaill et al (1993) compared group interactions in face-to-face meetings with two types of video conferencing systems, a low quality system (ISDN link, half-duplex line, with transmission lags) versus a better system (full-duplex line, immediate transmission and broadcast-quality image). They compared these 3 settings according to different criteria such as the length of turns, the frequency of backchannel responses (short messages such as ’hmm’, ’uuuh’), and so forth. In the IDSN system, listeners seems to be aware of the disruptive effects of communication lag and half-duplex audio and leave the speaker finishing his point, leading to fewer but longer turns, i.e. to a more more, less spontaneous conversations. Conversely, these systems do not differ regarding to explicit hand-over messages (signaling to the other speaker they have finished their turn, e.g. by adding ’isn't?’ or mentioning the name of the next speaker): in both video conferencing systems, sound and vision are non-directional, while directionality is important in face to face since turning head or eye gaze often announces a speaker change. Comparing 3 video conferencing systems with face-to-face and audio-only settings, Sellen (1995) actually observed that turn taking behavior was unaffected by the lack of visual clues such as selective gaze, as long as the audio quality is good enough to enable the speaker to substitute audio cues to visual cues. These studies illustrate the evolution of research on CSCW: the main relationship between a system and some task performance measures is decomposed into detailed analyses of the relationship between various medium features and various aspects of interaction. Salomon (1990) summarize this issue by discriminating the "effects of technology" and the "effects with technology": one cannot address the former, i.e. the cognitive trace of using some tool, without an appraisal of the latter, i.e. how users function with these media.
The current research belongs to two traditions of research, collaborative problem solving and CSCW, which fortunately converge on the need to detailed analyses of (mainly verbal) interactions. Therefore, we borrowed tools and concepts from pragmatics, especially from the work of Clark and his colleagues. However, the respective contribution from psychology and linguistics differ by two main points: the implicit criteria used for evaluation interactions and the scale of analysis.
If one considers the efficiency of communication, it is more advantageous to minimize the necessary effort for grounding interactions. This does necessarily mean that the speaker has to foresee and avoid all possible problems. What is important is not individual effort by the receiver of a communicative act, but the overall Least Collaborative Effort (Clark, 1986). The cost of producing a perfect utterance may be higher (if it is even possible) than the cost of collaboratively repairing those problems which do arise. However, we are less concerned by the economy of interaction than by the cognitive effects which may come out interactions. When two partners misunderstand, they have to build explanations, justify themselves, often make explicit some knowledge which would otherwise remain tacit and therefore reflect on their own knowledge, and so forth. This extra effort for grounding, even if it slows down interaction, may lead to better understanding of the task or ti better performance in the longer term. Hence, we rather talk in terms of Optimal Collaborative Effort (Dillenbourg, Traum & Schneider, 1996). As suggested by the word ’optimal’, those grounding efforts have to remain subordinated to the accomplishment of the task, i.e. to the effective need for grounding knowledge.
This study aims to describe how two agents elaborate (or fail to elaborate) a joint understanding of the problem they have to solve. The 'shared understanding' is used by psychologists and by linguists, but with a different scale. In psychology, the notion of 'shared understanding' is an intuitively appealing way of discriminating collaboration from cooperation, but it is far from being operational. At the opposite, in dialogue studies, 'shared/mutual understanding' refers to grounding mechanisms (acknowledgment, repair, request for acknowledgment, ...) by which one agent verifies that his utterance has been understood as he meant by his partner and repairs it if misunderstanding occurs. There is a circular relationship between 'shared understanding' of one utterance (micro level) and 'shared understanding' of the task and the underlying concepts involved (macro level). On one hand, an utterance makes only sense with respect to some context of reference. On the other hand, a shared understanding of the task is built through a complex sequence of utterances which have to be individually (more or less) understood. But how does this shared understanding of the task emerges from a complex structure of grounding episodes? There exists a large gap between describing an episode 3-5 utterances and understanding how a shared understanding emerges progressively through 754 of such episodes. Our solution will be to define a meso level, aggregating grounding episodes data into larger categories. We namely describe grounding according to the knowledge being negotiated and the phase in the problem solving process. Clark and Schaefer (1989) have emphasized that the degree of grounding varies according the task. By 'grounding criterion', they refer to the extent to which some piece of information has to be fully shared or not. For instance, you need to agree with your backer about the prize of bread, not about european politics. The grounding criterion does not only vary between tasks but also during the task. Our categories attempt to account for the variations of the grounding criterion during the task itself.
|