Adaptive topic - dependent language modelling using word - based varigrams

This paper presents two extensions of the standard interpolated word trigram and cache model, namely the extension of the trigram model by useful word m--grams with m ? 3 resulting into a varigram model, and the addition of topic--specific trigram models. We give the criteria for selecting useful m--grams and for partitioning the training corpus into topic-- specific subcorpora. We apply both extensions, separately and in combination, to corpora of 4 and 39 million words taken from the Wall Street Journal Corpus and show that high reductions in perplexity of up to 19 % on the largest corpus are achieved. We also performed some recognition experiments.