Sentence-extractive automatic speech summarization and evaluation techniques

Abstract

This paper presents sentence extraction-based automatic speech summarization techniques for making abstracts from spontaneous presentations. We propose a summarization technique using dimension reduction based on singular value decomposition which effectively focuses on the most salient topics of each presentation. With this technique, sentence location information, which is used for text summarization, is combined to extract important sentences from the introduction and conclusion segments of each presentation. We also investigate the combination of confidence measure and linguistic likelihood to effectively extract sentences with less recognition error. Experimental results show that the dimension-reduction-based method incorporating sentence location information, the confidence measure, and linguistic likelihood achieves the best automatic speech summarization performance in the condition of 10% summarization ratio. This paper also presents objective methods for evaluating automatic speech summarization methods. The correlation analysis between subjective and objective evaluation scores confirms that summarization accuracy, sentence F-measure, and 2 and 3-gram recall are the most effective among the objective evaluation metrics investigated in this paper.