Response generation based on statistical machine translation for speech-oriented guidance system

An example-based response generation is a robust and practical approach for a real-environment information guidance system. However, this framework cannot reflect differences in nuance, because the set of answer sentences are fixed beforehand. To overcome this issue, we have proposed response generation using a statistical machine translation technique. In this paper, we make use of N-best speech recognition candidates instead of manual transcription used in our previous study. As a result, the generation rate of appropriate response sentences was improved by using multiple recognition hypothesis.