Informational Redundancy and Resource Bounds in Dialogue

The view of language as action is wellentrenched in computational theories of discourse, however previous work has not paid particular attention to the fact that agents' resource bounds must a ect the way that dialogue proceeds just as it a ects agents' behavior in general. This paper presents results on the role of discourse strategies in improving performance on collaborative tasks when agents are resource-limited. The resource limitation addressed here is limits in inferential capacity, and the interaction of discourse strategies with this limitation is explored through a dialogue simulation environment, Design-World, in which inferential capacity can be parameterized. The results reported here demonstrate that discourse strategies that incorporate redundant information can improve the performance of inference-limited agents to the point of logical omniscience. The gures in this postworkshop version of the AAAI Reasoning about Mental States workshop paper are modi ed so that the simulation comparison results are plotted as di erences, rather than requiring the reader to visually make the comparison, in response to a suggestion from Mark Liberman. Other than this minor change, this paper is identical to the one presented at the workshop. Thanks to Aravind Joshi for funding this work under ARO grant DAAL03-89C0031PRI and DARPA grant N00014-90-J-1863 at the University of Pennsylvania, and to Hewlett Packard, U.K., and to an NSF award for 1991 Science and Engineering Institute in Japan