Summarizing Spoken and Written Conversations

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 773–782, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Summarizing Spoken and Written Conversations Gabriel Murray and Giuseppe Carenini Department of Computer Science University of British Columbia Vancouver, BC V6T 1Z4 Canada Abstract In this paper we describe research on sum- marizing conversations in the meetings and emails domains. We introduce a conver- sation summarization system that works in multiple domains utilizing general conversa- tional features, and compare our results with domain-dependent systems for meeting and email data. We find that by treating meet- ings and emails as conversations with general conversational features in common, we can achieve competitive results with state-of-the- art systems that rely on more domain-specific features. 1 Introduction Our lives are increasingly comprised of multimodal conversations with others. We email for business and personal purposes, attend meetings in person and remotely, chat online, and participate in blog or forum discussions. It is clear that automatic summa- rization can be of benefit in dealing with this over- whelming amount of interactional information. Au- tomatic meeting abstracts would allow us to prepare for an upcoming meeting or review the decisions of a previous group. Email summaries would aid corpo- rate memory and provide efficient indices into large mail folders. When summarizing in each of these domains, there will be potentially useful domain-specific fea- tures – e.g. prosodic features for meeting speech, subject headers for emails – but there are also un- derlying similarites between these domains. They are all multiparty conversations, and we hypothe- size that effective summarization techniques can be designed that would lead to robust summarization performance on a wide array of such conversation types. Such a general conversation summarization system would make it possible to summarize a wide variety of conversational data without needing to develop unique summarizers in each domain and across modalities. While progress has been made in summarizing conversations in individual domains, as described below, little or no work has been done on summarizing unrestricted, multimodal conversa- tions. In this research we take an extractive approach to summarization, presenting a novel set of conver- sational features for locating the most salient sen- tences in meeting speech and emails. We demon- strate that using these conversational features in a machine-learning sentence classification framework yields performance that is competitive or superior to more restricted domain-specific systems, while having the advantage of being portable across con- versational modalities. The robust performance of the conversation-based system is attested via several summarization evaluation techniques, and we give an in-depth analysis of the effectiveness of the indi- vidual features and feature subclasses used. 2 Related Work on Meetings and Emails In this section we give a brief overview of previous research on meeting summarization and email sum- marization, respectively. 773 2.1 Meeting Summarization Among early work on meeting summarization, Waibel et al. (1998) implemented a modified version of the Maximal Marginal Relevance algorithm (Car- bonell and Goldstein, 1998) applied to speech tran- scripts, presenting the …

Content truncated to 500 words