On enhancing katz-smoothing based back-off language model

Though the statistical language modeling plays an important role in speech recognition, there are still many problems that are difficult to be solved such as the sparseness of training data. Generally, two kinds of smoothing approaches, namely the back-off model and the interpolated model, have been proposed to solve the problem of the impreciseness of language models caused by the sparseness of training data. By expanding the idea of back-off to the re-estimation of not only the unseen word pairs but also all word pairs, a back-off model based modified method is proposed, referred to as the Enhanced Katz smoothing with deleted interpolation (EKSWDI). A uniform expression and two simplified versions for this modified model are also given. Experiments on a Chinese pinyin-to-character conversion system and the perplexity measure show that the proposed model has a better performance than the Katz smoothing method does.