A Study on Statistical Methods for Automatic Speech Summarization

This dissertation proposes a new automatic speech summarization method through word extraction. In this method, a set of words maximizing a summarization score indicating an appropriateness of summarization is extracted from automatically transcribed speech. This extraction is performed according to a target compression ratio using a dynamic programming technique sentence by sentence. The extracted set of words is then connected to construct a summarization sentence. The summarization score consists of a word significance measure, a confidence measure, linguistic likelihood, and a word concatenation probability. The word concatenation score is determined by a dependency structure in the original speech given by Stochastic Dependency Context Free Grammar (SDCFG). This summarization process aims to maintain the original meaning as much as possible within a limited number of words. In order to make abstracts, we then extend this method to summarization of a set of multiple utterances (sentences). This is done by adding a rule which restricts application of the score beyond the sentence boundaries. This summarization technique can preserve more words inside information rich utterances and shortening or even completely deleting less informative ones. This paper proposes a summarization method for multiple utterances using a two-level Dynamic Programming (DP) technique in order to reduce the amount of calculation. Japanese and English broadcast news speech which has been transcribed using a largevocabulary continuous-speech recognition (LVCSR) system is summarized using our proposed method. In addition, lecture speech is also transcribed and summarized to make abstracts. and compared with manual summarization by human subjects. The manual summarization results are combined to build a word network. This word network is used to calculate the word accuracy of each automatic summarization result using the most