Abstractive Summarization of Spoken and Written Conversations Based on Phrasal Queries

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1220–1230, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Abstractive Summarization of Spoken and Written Conversations Based on Phrasal Queries Yashar Mehdad Giuseppe Carenini Raymond T. Ng Department of Computer Science, University of British Columbia Vancouver, BC, V6T 1Z4, Canada {mehdad, carenini, rng}@cs.ubc.ca Abstract We propose a novel abstractive query- based summarization system for conversa- tions, where queries are defined as phrases reflecting a user information needs. We rank and extract the utterances in a con- versation based on the overall content and the phrasal query information. We clus- ter the selected sentences based on their lexical similarity and aggregate the sen- tences in each cluster by means of a word graph model. We propose a ranking strat- egy to select the best path in the con- structed graph as a query-based abstract sentence for each cluster. A resulting sum- mary consists of abstractive sentences rep- resenting the phrasal query information and the overall content of the conversa- tion. Automatic and manual evaluation results over meeting, chat and email con- versations show that our approach signifi- cantly outperforms baselines and previous extractive models. 1 Introduction Our lives are increasingly reliant on multimodal conversations with others. We email for business and personal purposes, attend meetings in per- son, chat online, and participate in blog or forum discussions. While this growing amount of per- sonal and public conversations represent a valu- able source of information, going through such overwhelming amount of data, to satisfy a partic- ular information need, often leads to an informa- tion overload problem (Jones et al., 2004). Au- tomatic summarization has been proposed in the past as a way to address this problem (e.g., (Sakai and Sparck-Jones, 2001)). However, often a good summary cannot be generic and should be a brief and well-organized paragraph that answer a user’s information need. The Document Understanding Conference (DUC)1 has launched query-focused multidocu- ment summarization as its main task since 2004, by focusing on complex queries with very specific answers. For example, “How were the bombings of the US embassies in Kenya and Tanzania conducted? How and where were the attacks planned?”. Such complex queries are appropriate for a user who has specific information needs and can formulate the questions precisely. However, especially when dealing with conversational data that tend to be less structured and less topically focused, a user is often initially only exploring the source documents, with less specific information needs. Moreover, following the common practice in search engines, users are trained to form simpler and shorter queries (Meng and Yu, 2010). For example, when a user is interested in certain characteristics of an entity in online reviews (e.g., “location” or “screen”) or a specific entity in a blog discussion (e.g., “new model of iphone”), she would not initially compose a complex query. To address these issues, in this work, we tackle the task of conversation summarization based on phrasal queries. We define a …

Content truncated to 500 words