A Chunk based partial parsing strategy for reranking and normalizing Nbest lists of a speech recognizer

14 A Chunk based partial parsing strategy for reranking and normalizing Nbest lists of a speech recognizer ZAKARIA KURDI Groupe d’Etudes pour l’Oral et le Dialogue, CLIPS, IMAG B.P. 53, , Grenoble Cedex 09, France, E-mail: 14.1 Introduction Interactive spoken dialog provides new challenges for spoken language sys- tems. These systems are faced with two serious problems: 1. One of the most critical problems is the prevalence of speech disfluencies (extragrammaticalities). There are phenomena of speech disfluencies like hesitation, false starts, self repairs, etc. that the speaker utters and which need to be corrected in order to understand the utterance’s meaning. 2. An even more serious problem is the imperfect word accuracy of speech recognizers, particularly when faced with a large vocabulary. In order to overcome these problems many researchers have proposed the integration of high level information sources into language models. Structure based language models employ grammar formalisms richer than weighted finite-state grammars such as Stochastic Context free grammar (SCFG). In this context two kinds of strategies was investigated: In the first one, the high level information was directly integrated into the language model of the recognizer ([13]), ([5]). In the second one, a specific module was added to the system for reranking the Nbest hypothesis of the speech recognition ([23]). Based on ideas from ([2]), the system of Zechner and Waibel is a chunk based partial parsing which is used to generate shallow syntactic structures from speech recognizer output. These representations then serve as the basis for scores used in the task of reranking Nbest lists. The parsing is done after a normalization process that cleans up the 157 ......................................continued...............................................