Incorporating Speaker and Discourse Features into Speech Summarization

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 367–374, New York, June 2006. c⃝2006 Association for Computational Linguistics Incorporating Speaker and Discourse Features into Speech Summarization Gabriel Murray, Steve Renals, Jean Carletta, Johanna Moore University of Edinburgh, School of Informatics Edinburgh EH8 9LW, Scotland gabriel.murray@ed.ac.uk, s.renals@ed.ac.uk, jeanc@inf.ed.ac.uk, j.moore@ed.ac.uk Abstract We have explored the usefulness of incorporat- ing speech and discourse features in an automatic speech summarization system applied to meeting recordings from the ICSI Meetings corpus. By an- alyzing speaker activity, turn-taking and discourse cues, we hypothesize that such a system can out- perform solely text-based methods inherited from the field of text summarization. The summariza- tion methods are described, two evaluation meth- ods are applied and compared, and the results clearly show that utilizing such features is advanta- geous and efficient. Even simple methods relying on discourse cues and speaker activity can outper- form text summarization approaches. 1. Introduction The task of summarizing spontaneous spoken di- alogue from meetings presents many challenges: information is sparse; speech is disfluent and frag- mented; automatic speech recognition is imper- fect. However, there are numerous speech-specific characteristics to be explored and taken advantage of. Previous research on summarizing speech has concentrated on utilizing prosodic features [1, 2]. We have examined the usefulness of additional speech-specific characteristics such as discourse cues, speaker activity, and listener feedback. This speech features approach is contrasted with a sec- ond summarization approach using only textual features—a centroid method [3] using a latent se- mantic representation of utterances. These indi- vidual approaches are compared to a combined ap- proach as well as random baseline summaries. This paper also introduces a new evalua- tion scheme for automatic summaries of meeting recordings, using a weighted precision score based on multiple human annotations of each meeting transcript. This evaluation scheme is described in detail below and is motivated by previous find- ings [4] suggesting that n-gram based metrics like ROUGE [5] do not correlate well in this domain. 2. Previous Work In the field of speech summarization in general, re- search investigating speech-specific characteristics has focused largely on prosodic features such as F0 mean and standard deviation, pause information, syllable duration and energy. Koumpis and Re- nals [1] investigated prosodic features for summa- rizing voicemail messages in order to send voice- mail summaries to mobile devices. Hori et al. [6] have developed an integrated speech summariza- tion approach, based on finite state transducers, in which the recognition and summarization compo- nents are composed into a single finite state trans- ducer, reporting results on a lecture summariza- tion task. In the Broadcast News domain, Maskey and Hirschberg [7] found that the best summariza- tion results utilized prosodic, lexical, and structural features, while Ohtake et al. [8] explored using only prosodic features for summarization. Maskey and Hirschberg similarly found that prosodic fea- tures alone resulted in good quality summaries of 367 Broadcast News. In the meetings domain (using the ICSI cor- pus), Murray et al. [2] …

Content truncated to 500 words