Using Confusion Networks for Speech Summarization

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 46–54, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Using Confusion Networks for Speech Summarization Shasha Xie and Yang Liu Department of Computer Science The University of Texas at Dallas {shasha,yangl}@hlt.utdallas.edu Abstract For extractive meeting summarization, previ- ous studies have shown performance degrada- tion when using speech recognition transcripts because of the relatively high speech recogni- tion errors on meeting recordings. In this pa- per we investigated using confusion networks to improve the summarization performance on the ASR condition under an unsupervised framework by considering more word candi- dates and their confidence scores. Our ex- perimental results showed improved summa- rization performance using our proposed ap- proach, with more contribution from leverag- ing the confidence scores. We also observed that using these rich speech recognition re- sults can extract similar or even better sum- mary segments than using human transcripts. 1 Introduction Speech summarization has received increasing in- terest recently. It is a very useful technique that can help users to browse a large amount of speech recordings. The problem we study in this paper is extractive meeting summarization, which selects the most representative segments from the meeting tran- scripts to form a summary. Compared to text sum- marization, speech summarization is more challeng- ing because of not only its more spontaneous style, but also word errors in automatic speech recogni- tion (ASR) output. Intuitively the incorrect words have a negative impact on downstream summariza- tion performance. Previous research has evaluated summarization using either the human transcripts or ASR output with word errors. Most of the prior work showed that performance using ASR output is consistently lower (to different extent) comparing to that using human transcripts no matter whether su- pervised or unsupervised approaches were used. To address the problem caused by imperfect recognition transcripts, in this paper we investigate using rich speech recognition results for summariza- tion. N-best hypotheses, word lattices, and confu- sion networks have been widely used as an inter- face between ASR and subsequent spoken language processing tasks, such as machine translation, spo- ken document retrieval (Chelba et al., 2007; Chia et al., 2008), and shown outperforming using 1- best hypotheses. However, studies using these rich speech recognition results for speech summariza- tion are very limited. In this paper, we demonstrate the feasibility of using confusion networks under an unsupervised MMR (maximum marginal relevance) framework to improve summarization performance. Our experimental results show better performance over using 1-best hypotheses with more improve- ment observed from using confidence measure of the words. Moreover, we find that the selected summary segments are similar to or even better than those gen- erated using human transcripts. 2 Related Work Many techniques have been proposed for the meet- ing summarization task, including both unsuper- vised and supervised approaches. Since we use un- supervised methods in this study, we will not de- scribe previous work using supervised approaches because of the space limit. Unsupervised …

Content truncated to 500 words