Language Modeling With Dynamic Bayesian Networks Using Conversation Types and Part of Speech Information

In this paper we investigate whether more accurate modeling of differences in language in different types of conversations, e.g. formal presentations vs. spontaneous conversations can improve the quality of a language model. We also investigate whether the modeling of sentence lengths can improve a language model. A language model is an important component of statistical natural language processing systems, such as automatic speech recognizers and spelling checkers, that judges the plausibility of sentence hypotheses. Standard language modeling approaches rely on statistics over word sequences. Our experiments show that modeling the conversation type and part-of-speech tags sequences improves the language models, while modeling sentence length does not.