A novel decoding framework for extractive speech summarization with Rhetorical Structure modeling

摘要

We propose a novel decoding framework (Hybrid RSD-SE framework: Rhetorical Structure Decoding-Summary Extraction framework) for extracting summaries with rhetorical structure information from speech. Rhetorical structure hidden in speech data which helps us understand speech easier is always under-utilized. The hybrid RSDSE framework automatically decodes this underlying information in order to provide better-organized summaries by combining the process of rhetorical structure extraction with the process of summarization. We show that the hybrid RSD-SE framework gives a 82.24% ROUGE-L F-measure, a 3.33% absolute increase in lecture speech summarization performance compared to the baseline systems [1], [2].