Towards Unrestricted Syntax for Speech Understanding
スポンサーリンク
概要
- 論文の詳細を見る
Traditional models of syntax assume each utterance can be given a complete parse tree. This view is not well suited for understanding spoken language utterances where the speaker speaks in a fragmented, spontaneous style. To make a speech-friendly parser, there have been many efforts to modify traditional parsers; but I have taken a different approach: to radically re-examine the syntactic aspects of speech understanding. The goal is an "unrestrictive parser", meaning a syntactic mechanism suitable for understanding all utterances, thereby removing the need to restrict the user to speak grammatically. The first issue is dealing with the inherent uncertainty of speech input, especially for noisy inputs. Any recognizer produces large numbers of word hypotheses. So that the parser can do its work, these are traditionally reduced somehow into one or a few "sentence hypotheses" (typically this reduction relies on a separate "language model"). Doing so can lose a lot of information; thus syntax is a bottleneck. To eliminate this problem, syntax should work directly from the raw word hypotheses, numerous as they are The second issue is feedback. Because purely bottom-up speech recognition is hard, effective mobilization of "higher-level knowledge" to aid the recognizer is a major goal in speech understanding research. Traditionally semantic constraints have not been applied here (with two exceptions, 1. when syntactic and semantic processing is integrated into one module, as in semantic grammars, and 2. when semantic constraints are used for filtering complete interpretations.) The difficultly has been the lack of a syntactic mechanism which can, given some semantic feedback, compute the implications of that feedback for the word hypotheses. The figure roughly indicates some of the kinds of information flow needed for speech understanding.
- 一般社団法人情報処理学会の論文
- 1994-03-07