Introduction

A language model is primarily based on two types of information. One is grammatical and syntactic information, and the other is stochastic and statistical information. Currently, a word $N$-gram model has usually been used in speech recognition system[1]. The word $N$-gram model is local statistical information on language and does not include syntactic information.

Conventionally, some language models of speech recognition systems use context-free grammar [2] or network grammar [3]. However, these language models have not been proved whether these models are effective or not. One of the reasons is that the number of rules and vocabulary used by them are limited. In this paper, we will examine the effectiveness of a large number of the valency patterns in Japanese sentence speech recognition.

We used the valency patterns of Nihongo Goi Taikei, a dictionary published by Iwanami [4]. These valency patterns contain detailed information on case elements and structural usages of 6,000 Japanese predicates. The dictionary contains a total of 300-thousand Japanese words, about 14,000 valency patterns, and 2,710 semantic categories of noun that are composed in a 12-layer hierarchy.

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

Note on the paper; the strings enclosed by ¡Ö and ¡× are the Japanese Kanji-Kana expression, and the strings enclosed by " and " indicate the Japanese pronunciation for the sentence. The strings enclosed by ' and ' are the English translation. UGS means an ungrammatical sentence or a semantically incorrect sentence.