Combination of words and word categories in varigram histories

This paper presents a new kind of language models: category/word varigrams, This special model type permits a tight integration of word-based and category-based modeling of word sequences. Any succession of words and word categories may be employed to describe a given word history. This provides a much greater flexibility than previous combinations of word-based and category-based language models. Experiments on the WSJ0 corpus and the 1994 ARPA evaluation data indicate that the category/word varigram yields a perplexity reduction of up to 10 percent as compared to a word varigram of the same size, and improves the word error rate (WER) by 7 percent. Compared to a linear interpolation of a word-based and a category-based n-gram, the WER improvement is about 4 percent.