PLSA Based Topic Mixture Language Modeling Approach

In this paper, we propose a method to extend the use of latent topics into higher order n-gram models. In training, the parameters of higher order n-gram models are estimated using discounted average counts derived from the application of probabilistic latent semantic analysis(PLSA) models on n-gram counts in training corpus. In decoding, a simple yet efficient topic prediction method is introduced to predict its topic given a new document. The proposed topic mixture language model (TMLM) displays two advantages over previous methods: 1) having the ability of building topic mixture n-gram LM (n>1) and, 2) without requiring a big general baseline LM. The experimental results show that TMLMs, even using smaller number of topics, outperform LMs implemented using both standard n-gram approach and unsupervised adaptation approaches in terms of perplexity reductions.