next up previous
Next: Introduction


Speech Recognition using Valency Patterns

Jin'ichi Murakami, Ryo Saeki, and Satoru Ikehara

Dept. of Information and Knowledge Eng., Tottori University,
4-101 Koyama-Minami, Tottori 680-8550, Japan

{murakami| saeki | ikehara}


We examined the effectiveness of the valency patterns when they were included in a Japanese sentence speech recognition algorithm. These valency patterns contained detailed information on case elements and structural usages of 6,000 Japanese predicates. The patterns were extracted from Nihongo Goi Taikei published by Iwanami [4].

Nihongo Goi Taikei is a large dictionary containing a total of 300 thousand entries (Japanese words) and about 14,000 valency patterns and 2,710 types (groups) of noun.

We used a speech recognition system that employed a bigram model to generate the $N$-best (8 best) output sentences. Next, we selected candidate sentences by using the valency patterns. Finally, we evaluated the changes in the sentence recognition rate when the valency patterns were used.

The results of experiment were that the rate improved for 23 of 70 sentences and the rate of the first candidate was 67%. This means that the valency patterns are useful for Japanese speech recognition.

Keywords: speech recognition, sentence recognition rate, the valency patterns, $N$-gram


<< PDF File >>
next up previous
Next: Introduction
Jin'ichi Murakami 2005-08-25