Language modeling for spoken dialogue system based on sentence transformation and filtering using predicate-argument structures

We present a novel scheme of language modeling for a spoken dialogue system by effectively exploiting the back-end documents the system uses for information navigation. The proposed method first converts sentences in the document, which are written and plain style, into spoken question-style queries, which are expected in spoken dialogue. In this process, we conduct dependency analysis to extract verbs and relevant phrases to generate natural sentences by applying transformation rules. Then, we select sentences which have useful information relevant to the target domain and thus are more likely to be queried. For this purpose, we define predicate-argument (P-A) templates based on a statistical measure in the target document. An experimental evaluation shows that the proposed method outperforms the conventional method in ASR performance, and the sentence selection based on the P-A templates is effective.