The Use of Clustering Techniques for Language Modeling – Application to Asian Languages

Cluster-based n-gram modeling is a variant of normal word-based n-gram modeling. It attempts to make use of the similarities between words. In this paper, we present an empirical study of clustering techniques for Asian language modeling. Clustering is used to improve the performance (i.e. perplexity) of language models as well as to compress language models. Experimental tests are presented for cluster-based trigram models on a Japanese newspaper corpus, and on a Chinese heterogeneous corpus. While the majority of previous research on word clustering has focused on how to get the best clusters, we have concentrated our research on the best way to use the clusters. Experimental results show that some novel techniques we present work much better than previous methods, and achieve up to more than 40% size reduction at the same perplexity.