Rapid adaptation of n-gram language models using inter-word correlation for speech recognition

In this paper, we study the fast adaptation problem of n-gram language model under the MAP estimation framework. We have proposed a heuristic method to explore inter-word correlation to accelerate MAP adaptation of n-gram model. According to their correlations, the occurrence of one word can be used to predict all other words in adaptation text. In this way, a large n-gram model can be efficiently adapted with a small amount of adaptation data. The proposed fast adaptation approach is evaluated in a Japanese newspaper corpus. We have observed a significant perplexity reduction even when we have only several hundred adaptation sentences.