With the contribution of visitors from IRST, a large number of methods for estimating probabilities of stochastic Language Models (LM) has been compared and a system has been implemented for computing LM probabilities from a large corpus. This system works in conjunction with another system that generates an original and compact lexical network representation with incorporated bigram probabilities. These two systems have been used to generate our version of the Air Travel Information System based on corpora made available from the Linguistic Data Consortium (LDC). New interesting theoretical results have been obtained on parsing under the control of stochastic context-free grammars. They have been published in October 94 on the IEEE Transactions on Pattern Analysis and Machine Intelligence. A research is in progress for combining these results with semantic interpretation under the control of rules automatically learned with Classification and Regression Trees. The learning method as well as experimental results on its use have been published on the April 95 issue of the IEEE Transactions on Pattern Analysis and Machine Intelligence. Ways of deriving dialog strategies in person-robot dialogues are described with the Mobile Robot project.
R. DeMori, G. Antoniol, A. Corazza (IRST, Trenton, Italy), M. Federico, C. Pateras