Improving Japanese language models using POS information

In this paper, part-of-speech (POS) information is used to improve the performance of a Japanese language model (LM). The POS bigram is used to tackle the sparseness problem of the training data. Additionally, due to the characteristics of the Japanese language, part of the Japanese syntax information can be integrated into the POS bigram, through POS combination rules. Based on the Japanese syntax grammar, the POS combination rules determine if a POS pair is prohibited in Japanese language. The Japanese POS bigram table not only includes the POS pairs that occurred in the training corpus, but also includes all the prohibited POS pairs. The confusion in the search space can be reduced by explicitly modeling the prohibited POS pairs. In this work, a series of experiments have been carried out to investigate the impact of the POS bigram with prohibited POS pairs on the recognition search space. The framework of fast generation of the language model look-ahead (LMLA) probabilities based on POS bigram information is also presented in this paper. The experimental results showed that compared to the traditional word n-gram model, the LM with POS bigram information achieves significant improvement in both word accuracy and the speed of Japanese LVCSR system.