Random forests and the data sparseness problem in language modeling

摘要: Abstract Language modeling is the problem of predicting words based on histories containing already hypothesized. Two key aspects language are effective history equivalence classification and robust probability estimation. The solution these hindered by data sparseness problem. Application random forests (RFs) to deals with two simultaneously. We develop a new smoothing technique randomly grown decision trees (DTs) apply resulting RF models automatic speech recognition. This method complementary many existing ones dealing study our approach in context n-gram type which n − 1 present history. Unlike regular models, have potential generalize well unseen data, even when longer than four words. show that superior best known technique, interpolated Kneser–Ney smoothing, reducing both perplexity (PPL) word error rate (WER) large vocabulary state-of-the-art recognition systems. In particular, we will statistically significant improvements contemporary conversational telephony system applying only one its models.