Topic Detection of an Speech Dialogue P

This paper proposes a method for topic detection of an utterance for speech dialogue processing including speech translation. There are three important points to consider in realizing topic detection of an utterance. The first is to obtain information from a short utterance. The second is to deal with a broad topic. The third is to have robustness against speech recognition errors. Our topic detection method is suitable as a solution to these problems. To deal with the first two problems, our method detects the topic of an utterance according to the relation factor between the topics and the words included in utterances. The third problem is solved by merging words with different surface forms. To verify the performance of our proposed method, we carried out experiments by using a broad-coverage travel conversation corpus. The experimental results show that the proposed method achieved 79.3% topic detection accuracy for correct transcriptions, and 75.6% accuracy for recognizer’s outputs whose WER was 12.8%.