ISCA Archive - Incorporating probabilities into the dualgram language model (original) (raw)

Incorporating probabilities into the dualgram language model

Colin Matheson, Fergus R. McInnes

It is arguable that for certain applications of speech recognition technology it is useful to employ a language model whose characteristics lie between bigrams and trigrams. While bigrams are efficient and initially effective, the perplexity grows quickly as the model is extended to a point where recognition accuracy is affected. The move to trigrams solves this problem at the expense of increasing the amount of corpus required to levels which may be uneconomic. The DUALGRAM model was proposed in Matheson et al. 1990 [1] as a possible solution to some of these problems and initial results were presented. The present paper reviews the main points of the DUALGRAM approach and reports the addition of probabilities to the model. It has been demonstrated many times that probabilistic bigrams produce better recognition results than simple follow-set models, and similar improvements are to be expected with DUAL-GRAM. Finally, a new method of combining the probabilities derived from the two halves of the model is proposed.