Automatic n-gram language model creation from web resources

This paper describes an automatic building of N-gram language models from Web texts for large vocabulary continuous speech recognition. Although a huge amount of well-formed texts are needed to train a model, collecting and organizing such text corpus for every task by hand needs a great labor. We need the language model to update frequently to cover the current topics. To deal with this problem, we propose an automatic language model creation method by collecting Web texts via keywordbased Web search engines. We can build a task-dependent language model by selecting suitable keywords for the task. A text filtering algorithm based on character perplexity is developed to extract proper Japanese texts from Web texts. A language model for a medical consulting task created by the proposed method shows the higher word recognition rate by 11.4% than that of a conventional newspaper language model.