Grounding in Multi-modal Task-Oriented Collaboration

4 Preliminary results

At the time of this writing, we have had 7 pairs perform this task.[3] Pairs 1 and 2 communicated with voice as well as through MOO and whiteboard, and so we concentrate here on the other 5. These five pairs found the correct solution respectively in 2 h 45 (pair 3), 2h 27 (pair 4); 1h 53 (pair 5); 1 h 41 (pair 6) and 2 h 24 (pair 7).[4] Protocols also reveal a wide variety in how the pairs chose to interact. Pair 3 did not use at all the whiteboard (despite our repeated invitation to do it), while pairs 6 and 7 completely filled the available space. Some pairs (1,4,6,7) use the whiteboard as a space to paste 'post-it' notes, not fully exploiting the graphical features, especially spatial positions. Other pairs draw more elaborated schema: pair 2 draws a timetable; pairs 4 starts such a timetable but stops; pair 5 uses labelled arrows between objects to represent more complex concepts. Most pairs establish a code for marking the degree of suspicion for each character: they use colours, they cross out notes, or they put a label on those who are not suspect any more.

All pairs pass through three stages, reflecting the nature of the task: collecting information, organizing information (generally on the whiteboard) and making inferences. However, the pairs proceed differently through these stages:

The "methodical detectives" (e.g., Pairs 6 and 7) collect information in a systematic way, room by room, suspect by suspect. This task is performed cooperatively: they split the domain and explore individually. These two pairs barely talk at all during the first hour. They report all collected information on the whiteboard. Then, there is usually a 'pivot sentence' such as "We should put some order in our information isn't?" (protocol 7.2 - we translate) or "Let make a list of the persons who could steal the gun" (protocol 6.2). The second part of the work is the collaborative, it alternates phases of organization (basically, working on the whiteboard representation) and inferences. This style of deduction can be likened to forward chaining - reasoning from observed facts to eventual inferences regarding the goal.
The "intuitive detectives" (e.g. Pairs 4 and 5) interact as soon as they find some information that they judge potentially interesting. They make inferences and opportunistically decide on further investigation. The three processes of collecting data, organizing them and making inferences are performed in an interleaved fashion. This style of deduction can be likened to backward chaining - reasoning about unsubstantiated suspicions to decide which facts would be useful to gather.

We report two main observations: (1) the pairs seem to build different shared spaces, through different grounding mechanisms, and (2), grounding is often performed across different modalities.

[3] The protocols of MOO actions and communication and snapshots from the whiteboard movies are available by WWW (http://tecfa.unige.ch/tecfa/tecfa-research/cscps/bootnap.html).

[4] Pair 3 did not perform a training task and one subject was not previously familiar with the MOO.

4.1 - Grounding mechanisms are use to built shared sub-spaces
4.2 - Grounding across different modes of interaction

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

Generated with Harlequin WebMaker