Grounding in Multi-modal Task-Oriented Collaboration

5 Conclusions

The protocols indicate the complexity of what we initially summarized as "building a shared conception of the problem." First observation: the subjects do not simply build a shared space of knowledge, but they construct different sub-spaces. These spaces are connected to each other by functional relationships: some information in sub-space X has to be grounded in order to ground information in space Y. But, since the nature of information is different in each space, the grounding mechanisms are also different. It is precisely because the grounding processes are different that we treat these aspects of common ground as different sub-spaces. The nature of information includes factors such as the probability that this information is not shared, the cost of sharing it, the cost if it is not shared, the degree of mutuality of this information requested by the task and the persistence of this knowledge.

Second observation, subjects perform grounding in the different modes of interaction available (MOO action, MOO discussion, whiteboard actions): the information presented in mode X is often grounded by an act in mode Y. Sometimes, the three modes interplay for grounding information. Grounding is not bound inside a mode of interaction. It crosses different modalities in a way which is more complex than we initially expected. The choice of a modality for grounding is also related to the features of the information to be grounded. However, there is some be some flexibility in the relation between information to be grounded and modalities of grounding. If a grounding function cannot be performed through one mode, and if it really needs to be performed, it seems to migrate on another mode. For instance the function "ground basic observed facts" seems to migrate from the whiteboard to the MOO verbs (compare notebooks) from one pair to another. Such a migration makes sense within the theoretical framework of distributed cognition [Dillenbourg95a].

A deeper understanding of the relation between modalities and the nature of information to be grounded would require a fine analysis of temporal issues, namely the relationship between the persistence of information (how long it remains valid) and the persistence of the mode (how long the information remains displayed). People with MOO experience have probably noticed how much simple changes in timing (a message being typed faster than another one) can deeply affect the rules of conversation. A MOO is often treated as synchronous communication tool, but it is not completely synchronous. We infer from those observations that artificial agents should be provided with modality-independent grounding mechanisms that they could dynamically instantiate through different media according to the interaction with the human partner.

We infer from those observations that artificial agents should be provided with modality-independent grounding mechanisms that they could dynamically instantiate through different media according to the interaction with the human partner. These agents should consider the cost of grounding, the belief about how well grounded the material already is, and how necessary it is that the material be (better) grounded. While some of the costs will be medium dependent, the information can be seen as a whole, and the choice of which action to take when can be left to local considerations, such as persistence of the information, and how whether the system has the attention of the user, or if there's room on a graphical display.

Grounding in Multi-modal Task-Oriented Collaboration - 3 SEP 1996

Generated with Harlequin WebMaker