Variable-length category-based n-grams for language modelling

This report concerns the theoretical development and subsequent evaluation of n-gram language models based on word categories. In particular, part-of-speech word classifications have been employed as a means of incorporating significant amounts of a-priori grammatical information into the model. The utilisation of categories diminishes the problem of data sparseness which plagues conventional word-based n-gram approaches, and therefore yields a fundamentally more compact model. Furthermore, it allows the use of larger n, and a strategy by means of which successively longer n-grams are selectively added to the model according to a cross-validation likelihood criterion is proposed. This enables the model compactness to be maintained while allowing longer range effects to be modelled where they benefit performance. The language modelling approach was applied to the LOB corpus in order to assess its effectiveness. When compared with models of corresponding complexity constructed according to conventional n-gram methods, it is found that the proposed procedures render language models exhibiting superior performance. Furthermore, comparison with word-based n-gram models shows that comparable performance may be achieved at a large reduction in model size. An ultimate aim of the described work is to construct language models from very large text corpora, the contents of which are generally not annotated with the required part-of-speech classifications. For this reason the use of the category-based language model as a statistical tagger is introduced as a means of automatically determining this information, and is shown by means of tests on the LOB corpus to yield very good tagging accuracies.