N-gram distribution based language model adaptation

This paper presents two techniques for language model (LM) adaptation. The first aims to build a more general LM. We propose a distribution-based pruning of n-gram LMs, where we prune n-grams that are likely to be infrequent in a new document. Experimental results show that the distribution-based pruning method performed up to 9% (word perplexity reduction) better than conventional cutoff methods. Moreover, the pruning method results in a more general ngram backoff model, in spite of the domain, style, or temporal bias in the training data. The second aims to build a more task-specific LM. We propose an n-gram distribution adaptation method for LM training. Given a large set of out-of-task training data, called training set, and a small set of task-specific training data, called seed set, we adapt the LM towards the task by adjusting the n-gram distribution in the training set to that in the seed set. Experimental results show non-trivial improvements over conventional methods.