Statistical language modeling with a class based N-multigram model

In this paper, we report on speech recognition experi ments with an n-multigram language model, a stochastic model which assumes dependencies of length n between variable-length phrases. The n-multigram probabilities can be estimated in a class-based framework, where both the phrase distribution and the phrase classes are learned from the data according to a Maximum Likeihood cri terion, using a generalized Expectation-Maxiization al gorithm. In our speech recognition experiments on a database of air travel reservations, the 2-multigram mode allows a redction of 10% of the word error rate with respect to the usual trigram model, with 25% fewer param eters than in the trigram mode. We also report on a scheme where some a priori information is introduced in the odel ia seantic tagging.