First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Sentence Speech Recognition Using Semantic Dependency Analysis

Shoichi Matsunaga (1), Shigeki Sagayama (2)

(1) NTT Human Interface Laboratories, Tokyo, Japan
(2) ATR Interpreting Telephony Laboratories, Kyoto, Japan

This paper describes a sentence speech recognition system based on phoneme-based hidden Markov models (HMMs) and two grammatical constraints: a syntactic grammar of phrase structure and a semantic dependency grammar of sentence structure. A joint score, combining acoustic likelihood and linguistic certainty factors derived from phoneme based HMMs and two grammatical constraints, is maximized to obtain the optimal sentence recognition. A semantic analysis algorithm globally optimizes the joint score. This algorithm is based on two key techniques: most likely multi-phrase candidate-detection using the Viterbi algorithm, and breadth-first search for dependency parsing. Where the perplexity of the phrase syntax is 40, this system increases phrase recognition performance in the sentences by approximately 14%, showing the effectiveness of semantic dependency analysis.

Full Paper

Bibliographic reference.  Matsunaga, Shoichi / Sagayama, Shigeki (1990): "Sentence speech recognition using semantic dependency analysis", In ICSLP-1990, 929-932.