Incorporating proximity information for relevance language modeling in speech recognition

Language modeling (LM), aiming to provide a statistical mechanism to associate quantitative scores to sequences of words, has long been an interesting yet challenging problem in the field of speech and language processing. Although the ngram model remains the predominant one, a number of disparate LM methods have been developed to complement the n-gram model. Among them, relevance modeling (RM) that explores the relevance information inherent between the search history and an upcoming word has shown preliminary promise for dynamic language model adaptation. This paper continues this general line of research in two significant aspects. First, the so-called “bag-of-words” assumption of RM is relaxed by incorporating word proximity evidence into the RM formulation. Second, latent topic information is additionally explored in the hope to further enhance the proximity-based RM framework. A series of experiments conducted on a large vocabulary continuous speech recognition (LVCSR) task seem to demonstrate that the various language models deduced from our framework are very comparable to existing language models.