DAI interaction protocols as control strategies in a natural language processing system

The goal of the paper is to define and show the relevance for a constrained communication between the various autonomous agents genuinely embedded in a distributed natural language processing (NLP) system. NLP raises the problem of ambiguities and therefore the multiple solutions derived. Architectures based on sequential levels, in which each module corresponds to a linguistic level (preprocessing, morphology, syntax, pragmatics, semantics) have shown their limits. A sequential architecture does not provide an adequate framework for exchanging necessary information between different modules in order to reduce ambiguities. We present a distributed approach to NLP able to solve ambiguities by cooperation of different linguistic agents. It turns out that proposing a collaborative learning interaction protocol as a means to program distributed sentence analysis leads to a restriction of ambiguities. The resulting system can operate with partial analyses at different classical levels of analysis, change strategies according to the applications or the corpus in question.