Opinosis: A Graph Based Approach to Abstractive Summarization of Highly Redundant Opinions

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 340–348, Beijing, August 2010 Opinosis: A Graph-Based Approach to Abstractive Summarization of Highly Redundant Opinions Kavita Ganesan and ChengXiang Zhai and Jiawei Han Department of Computer Science University of Illinois at Urbana-Champaign {kganes2,czhai,hanj}@cs.uiuc.edu Abstract We present a novel graph-based summa- rization framework (Opinosis) that generates concise abstractive summaries of highly re- dundant opinions. Evaluation results on sum- marizing user reviews show that Opinosis summaries have better agreement with hu- man summaries compared to the baseline ex- tractive method. The summaries are readable, reasonably well-formed and are informative enough to convey the major opinions. 1 Introduction Summarization is critically needed to help users better digest the large amounts of opinions ex- pressed on the web. Most existing work in Opin- ion Summarization focus on predicting sentiment orientation on an entity (Pang et al., 2002) (Pang and Lee, 2004) or attempt to generate aspect-based ratings for that entity (Snyder and Barzilay, 2007) (Lu et al., 2009)(Lerman et al., 2009)(Titov and Mcdonald, 2008). Such summaries are very infor- mative, but it is still hard for a user to understand why an aspect received a particular rating, forcing a user to read many, often highly redundant sen- tences about each aspect. To help users further di- gest the opinions in each aspect, it is thus desirable to generate a concise textual summary of such re- dundant opinions. Indeed, in many scenarios, we will face the problem of summarizing a large number of highly redundant opinions; other examples include sum- marizing the ‘tweets’ on Twitter or comments made about a blog or news article. Due to the sub- tle variations of redundant opinions, typical extrac- tive methods are often inadequate for summarizing such opinions. Consider the following sentences: 1. The iPhone’s battery lasts long, only had to charge it once every few days. 2. iPhone’s battery is bulky but it is cheap.. 3. iPhone’s battery is bulky but it lasts long! With extractive summarization, no matter which single sentence of the three is chosen as a sum- mary, the generated summary would be biased. In such a case, an abstractive summary such as ‘iPhone’s battery is cheap, lasts long but is bulky’ is a more complete summary, conveying all the necessary information. Extractive methods also tend to be verbose and this is especially problem- atic when the summaries need to be viewed on smaller screens like on a PDA. Thus, an informa- tive and concise abstractive summary would be a better solution. Unfortunately, abstractive summarization is known to be difficult. Existing work in abstractive summarization has been quite limited and can be categorized into two categories: (1) approaches us- ing prior knowledge (Radev and McKeown, 1998) (Finley and Harabagiu, 2002) (DeJong, 1982) and (2) approaches using Natural Language Genera- tion (NLG) systems (Saggion and Lapalme, 2002) (Jing and McKeown, 2000). The first line of work requires considerable amount of manual effort to define schemas such as frames and templates that can be filled …

Content truncated to 500 words