Parse tree n-grams for spoken language modelling

A method is described for modelling natural language for speech recognition. Its aim is to incorporate the advantages of two previous types of approach; the statistical approach and the formal linguistic approach. The n-gram model (J.K. Baker, 1975), is a statistical model based on corpus data, and is used to estimate the probability of an unseen sequence of words based on this data. A phrase structure grammar is an example of a formal language model which is a representation of a priori linguistic knowledge. Such models can provide a structure, called a parse tree, for a sequence of words. The approach described is to apply n-grams to parse trees. In this way the dependence on a predefined set of grammar rules is avoided, but the advantage of a structural analysis is retained. The main practical difficulty with this approach is that conventionally an n-gram model is used to estimate the probability of a sequence of words, whereas a parse tree is a two dimensional structure. This problem applies both to the training phase, where transition probabilities are estimated, and to the recognition phase, where the probability of a sequence of words is calculated from these transition probabilities.<
>