Linguistic Model Adaptation for Speech Summarisation

In this paper we extend the work done on the two-stage summarisation method described in [1] by focusing on adapting the linguistic component to make it more suited for the summarisation task. In particular we examine methods for adapting the linguistic models (LiM) automatically to improve performance, using either unigram, bi-gram or trigram information from different sources of data. Experiments were performed both on spontaneous speech, using 9 talks taken from the Translanguage English Database (TED) corpus [2], and speech read from text, using 5 talks from CNN broadcast news from 1998. For each of those talks, human (TRS) and speech recogniser (ASR) transcriptions along with human summaries were used. The talks were used for both development and evaluation with a rotating form of cross-validation [3]. The objective measure of summary quality used in this paper is summarisation accuracy (SumACCY) [4]. The full process is described in [5].