Online Debate Summarization using Topic DirectedSentiment AnalysisSarvesh Ranade, Jayant Gupta, Vasudeva Varma, Radhika MamidiInternational Institute of Information Technology, Hyderabad{sarvesh.ranade | jayant.gupta}@research.iiit.ac.in{vv | radhika.mamidi}@iiit.ac.inABSTRACTSocial networking sites provide users a virtual community in-teraction platform to share their thoughts, life experiencesand opinions. Online debate forum is one such platformwhere people can take a stance and argue in support oropposition of debate topics. An important feature of suchforums is that, they are dynamic and increase rapidly. Insuch situations, ef f ective opinion summarization approachesare needed so that readers need not go through the entiredebate. This paper aims to summarize online debates by ex-tracting highly topic relevant and sentiment rich sentences.The proposed approach takes into account topic relevant,document relevant and sentiment based features to capturetopic opinionated sentences. ROUGE scores are used toevaluate our system. Our system signif i cantly outperformsseveral baseline systems and show 5.2% (ROUGE-1), 7.3%(ROUGE-2) and 5.5% (ROUGE-L) improvement over thestate-of-the-art opinion summarization system. The resultsverify that topic directed sentiment features are most im-portant to generate ef f ective debate summaries.1. INTRODUCTIONWith the exponential growth in the use of World WideWeb, online users express themselves on continuously emerg-ing social networking sites. These sites provide users a widevariety of choices. Micro-blogging sites like twitter allowsthem to express their opinion (140 characters) on trendingtopics. Users can express their views and share their expe-riences on popular web blogs like wordpress, blogspot, etc.Facebook allows community oriented interaction, restrictedwithin one’s friend circle. E-commerce users can provideproduct specif i c reviews on online shopping sites.Amongst social networking platforms, online debate sites(‘convinceme.net’, ‘4forums.net’, ‘onlinedebate.net’) have be-come popular in recent times. These online debate forumsprovide users an option to express their opinion about theirfavorite debate topics [28]. From the research point of view,they provide a rich collection of dif f ering opinions on vari-Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for prof i t or commercial advantage and that copiesbear this notice and the full citation on the f i rst page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specif i cpermission and/or a fee. WISDOM ’13, August 11 2013, Chicago, USA.Copyright 2013 ACM 978-1-4503-2332-1/13/08 ...$15.00.ous topics. In ideological two-sided debates, users supporttheir stance by cleverly stating arguments supporting theirstance or opposing other stance [27]. Conversation sentencesbetween two users are very common in this multi-party con-versation. People rebut to another user’s post and expresstheir viewpoint on other’s opinion [2]. Because of dynamicnature of debates and large number of posts (194 per debate)they contain, it is essential to generate ef f ective summariesfor them so that readers need not go through the entire de-bate.To help this cause, we need to summarize online debatessuch that, after reading the summaries, user gets a goodidea about the information debate presents and the opinionsusers express. This paper aims to summarize online debatescollected from a popular debate site called ‘convinceme.net’with the intention of capturing good topic information aswell as opinion rich sentences from the debate.Compared to generic summarization, opinion summariza-tion is a relatively novel area. Unlike traditional methods,two factors, the sentiment degree and the correlated events,play a major role in opinion summarization. Previous meth-ods [29, 15] have ef f ectively used these factors over news,blogs and conversation domain. However, online debates isa domain which is yet to be explored. An important aspectof online debates is that most of its sentences are sentimentrich and topic relevant. Thus, topic directed sentiment anal-ysis is an important feature to create ef f ective summaries.We have successfully used this feature and results validatethe ef f ectiveness of our approach.In our method, we analyzed factors governing importantsentences in debate summaries. We observed 3 importantfactors: informative sentences, sentiment rich sentences andsentences which describes topic related entities. Thus, topicrelated and sentiment carrying features are used in the pro-posed approach. We have also used positional features asthey have been ef f ective in generic summarization approaches.Document relevant features such as tf-idf scores are used tocapture content relevant sentences.Our system, DEBSumm generates extractive summariesusing the aforementioned features. ROUGE [18] scores havebeen used to evaluate the system summaries. Final resultsshow that sentiment words that are relevant to topic are themost important feature to create ef f ective summaries. Theresults also show that our system achieve better results thanprevious state-of-the-art and several baseline systems.The rest of the paper is organized as follows: Section 2describes related work; Section 3 gives a detailed descriptionof our approach. Section 4 describes experimental setup for
