Domain-Independent Abstract Generation for Focused Meeting Summarization

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1395–1405, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Domain-Independent Abstract Generation for Focused Meeting Summarization Lu Wang Department of Computer Science Cornell University Ithaca, NY 14853 luwang@cs.cornell.edu Claire Cardie Department of Computer Science Cornell University Ithaca, NY 14853 cardie@cs.cornell.edu Abstract We address the challenge of generating natu- ral language abstractive summaries for spoken meetings in a domain-independent fashion. We apply Multiple-Sequence Alignment to in- duce abstract generation templates that can be used for different domains. An Overgenerate- and-Rank strategy is utilized to produce and rank candidate abstracts. Experiments us- ing in-domain and out-of-domain training on disparate corpora show that our system uni- formly outperforms state-of-the-art supervised extract-based approaches. In addition, human judges rate our system summaries significantly higher than compared systems in fluency and overall quality. 1 Introduction Meetings are a common way to collaborate, share information and exchange opinions. Con- sequently, automatically generated meeting sum- maries could be of great value to people and busi- nesses alike by providing quick access to the es- sential content of past meetings. Focused meet- ing summaries have been proposed as particularly useful; in contrast to summaries of a meeting as a whole, they refer to summaries of a specific as- pect of a meeting, such as the DECISIONS reached, PROBLEMS discussed, PROGRESS made or AC- TION ITEMS that emerged (Carenini et al., 2011). Our goal is to provide an automatic summariza- tion system that can generate abstract-style fo- cused meeting summaries to help users digest the vast amount of meeting content in an easy manner. Existing meeting summarization systems re- main largely extractive: their summaries are com- prised exclusively of patchworks of utterances se- lected directly from the meetings to be summa- rized (Riedhammer et al., 2010; Bui et al., 2009; Xie et al., 2008). Although relatively easy to con- struct, extractive approaches fall short of produc- ing concise and readable summaries, largely due C: Looking at what we’ve got, we we want an LCD dis- play with a spinning wheel. B: You have to have some push-buttons, don’t you? C: Just spinning and not scrolling, I would say. B: I think the spinning wheel is definitely very now. A: but since LCDs seems to be uh a definite yes, C: We’re having push-buttons on the outside C: and then on the inside an LCD with spinning wheel, Decision Abstract (Summary): The remote will have push buttons outside, and an LCD and spinning wheel inside. A: and um I’m not sure about the buttons being in the shape of fruit though. D: Maybe make it like fruity colours or something. C: The power button could be like a big apple or some- thing. D: Um like I’m just thinking bright colours. Problem Abstract (Summary): How to incorporate a fruit and vegetable theme into the remote. Figure 1: Clips from the AMI meeting corpus (Mc- cowan et al., 2005). A, B, C and D refer to distinct …

Content truncated to 500 words