Tree-based language model dedicated to natural spoken dialogs systems

Within the framework of Natural Spoken Dialog systems , we propose in this paper a method which automatically builds, from a training corpus, a set of Language Models (LMs) organized as a binary tree and called a Tree-based LM (TLM). Each LM corresponds to a specific dialogue situation, where the general LM is attached to the root node and the leaves represent the more specialized ones. Such LMs can be used to automatically adapt the decoding process to the dialog situation. We propose a two-pass decoding strategy, which implements this idea: a LM is dynamically selected from the TLM according to a list of Nbest hypotheses produced by a first decoding process, then this LM is used to perform a rescoring process on the word graph previously calculated.