Detecting errors in Chinese spoken dialog system using ngram and dependency parsing

In this paper, a hybrid method of detecting ASR error in spoken turns is developed. The erroneous text is locally analyzed first by neighbouring co-occurrence relations using ngram model. Then the text is globally analyzed by long distance dependency relations using a dependency parser. Our experiments show that we can use information from a dependency parsing phase together with n-gram language model not only to detect erroneous ASR hypotheses that can cause understanding errors, but also reliably locate errors and sometimes correct them as the hypotheses are being processed.