Integrating Speech Recognition and Natural Language LTAG Techniques with Weighted Synchonized Automata
Résumé
This paper will focus on the conceptual and technical design of a system architecture devoted to the integration of the various natural spoken language processing levels. Our goal is to avoid the well known constraints and limitations imposed by the usual combination of stochastic recognition models and statistical n-gram language models. To do so, we investigate the intensive use of weighted finite-state automata at each processing level to gather the static information. These automata are obtained directly or by linearizing tree structures and are then minimalized to get a optimal sharing of common substructures. Our hypothesis is that this property of sharing directly leads to a high processing factorization and consequently to an improved computational efficiency. Experiments are currently carried out to integrate an analytical segmentation system, a stochastic phonetic recognition module and a parser based on synchronous LTAG (Lexicalized Tree Adjoining Grammars) [Shieber 90] [Lopez 98] for syntactic and semantic descriptions to analyse a spoken task-oriented man-machine dialog corpus. The segmentation stage is based on hierarchical multi-level lattices and produces a weighted automaton of predicted broad phonetic classes [Husson 96]. The phoneme recognition stage is led by usual phonetic hidden Markov models. Minimilized phonetic automata are computed from the canonical phonetic transcription of all words of the corpus [Laporte 93]. The syntactic and semantic lexicalized tree grammar are semi-automatically built from a corpus collected by an wizard of Oz experiment. The trees are linearized and the resulting automata are minimalized with similar techniques as described in [Evans 98]. We provide some results which exhibit the high syntactical substructures sharing rate which is obtained. These different automata are synchronized with links between states. The decoding and parsing modules produce temporary concurrent analysis structures attached to at least one state of the corresponding automaton. Distributed control modules are needed to apply and exploit local synchronization constraints on the activated synchronized states. These synchonization techniques allow us to implement all kind of hypothesis manipulation for both deductive and abductive reasoning.