Word Activation Forces-Based Language Modeling and Smoothing

N-gram language models are useful for modeling the local dependencies of word occurrences but not for capturing global word dependencies. When the window size n is limited, the n-gram is weak in terms of capturing long distance dependencies. Long-distance Dependency information has long been proven useful in language model. However, the improved performance of long-distance LMs over conventional n-gram models generally comes at the cost of increased decoding complexity and model size. Word Activation Forces has been proven a simple and human-comparable accurate measure to identify word closest associates. In this paper, Word Activation Forces-Based language model is proposed to capture the long distance dependency between words, but which is as fast for decoding as a conventional word n-gram. As shown by experiments on broadcast news, the proposed language modeling and smoothing can significantly reduce the perplexity of language models and word error rate with moderate computational cost.