Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Statistical Language Modeling with a Class Based N-Multigram Model

Sabine Deligne

IBM T. J. Watson Research Center, Yorktown Heights, NY, USA

In this paper, we report on speech recognition experiments 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 Likelihood criterion, using a generalized Expectation-Maximization algorithm. In our speech recognition experiments on a database of air travel reservations, the 2-multigram model allows a reduction of 19% of the word error rate with respect to the usual trigram model, with 25% fewer param eters than in the trigram model. We also report on a scheme where some a priori information is introduced in the model via semantic tagging.


Full Paper

Bibliographic reference.  Deligne, Sabine (2000): "Statistical language modeling with a class based n-multigram model", In ICSLP-2000, vol.3, 119-122.