常識判断を用いた音声理解システムの構築

摘要

The speech recognition technology is necessary to achieve a smooth conversation between a person and a computer. However, it is difficult to use flexible conversation by voice in a present speech recognition system, with restrictions on the speaker, the environment, and the vocabulary. Then, in this paper, the Speech Understanding System is constructed to obtain high accuracy by correction of language process based on large-scale knowledge base for existing speech recognition result. First, it corrects the speech that misses the sound model and the difference in the user environment by the Revision Method by the Degree of Coincidence using the Knowledge Base to Correct Recognition Result that corrected using information on the knowledge base for the recognition result with some sound models. Second, it thinks that some words in sentence often have a commonsense relation between them and it proposes the method for correcting using the Commonsense Judgment System to consider a commonsense relation between the words.%人とコンピュータの円滑な会話を実現するには音声認識技術が必要である．しかし現在の音声認識装置は話者や語彙などに制約があり，音声による柔軟な会話に用いることは難しい．本稿では，既存の音声認識結果に対して大規模知識べースに基づく言語処理による補正を行うことで高い精度を得られる音声理解システムを構築する．まず認識結果補正知識べースを用いた一致度補正方式によって，複数の音響モデルの認識結果に対して知識べースの情報を用いて補正を行い，音響モデルから外れる音声や利用環境による揺らぎを補正する．さらに，文章を構成する単語はそれらの間で常識的な関係を持っていることが多いと考え，常識判断システムを用いて語と語の常識的な関連性を考慮した補正手法を提案する．