Towards Opinion Summarization from Online Forums

Proceedings of Recent Advances in Natural Language Processing, pages 138–146, Hissar, Bulgaria, Sep 7–9 2015. Towards Opinion Summarization from Online Forums Ying Ding School of Information Systems Singapore Management University ying.ding.2011@smu.edu.sg Jing Jiang School of Information Systems Singapore Management University jingjiang@smu.edu.sg Abstract Summarizing opinions expressed in online forums can potentially benefit many peo- ple. However, special characteristics of this problem may require changes to stan- dard text summarization techniques. In this work, we present our initial attempt at extractive summarization of opinionated online forum threads. Given the nature of user generated content in online dis- cussion forums, we hypothesize that be- sides relevance, text quality and subjectiv- ity also play important roles in deciding which sentences are good summary sen- tences. We therefore construct an anno- tated corpus to facilitate our study of ex- tractive summarization of online discus- sion forums. We define a set of features to capture relevance, text quality and sub- jectivity, and empirically test their useful- ness in choosing summary sentences. Us- ing unpaired Student’s t-test, we find that sentence length and number of sentiment words have high correlations with good summary sentences. Finally we propose some simple modifications to a standard Integer Linear Programming based sum- marization framework to incorporate these features. 1 Introduction With the growing popularity of social media, people often share their experience and opinions openly on the Internet. Especially when a con- troversial event happens, there are many differ- ent opinions expressed in online forum threads, including judgement of the people and organiza- tions involved in the event and suggestions for future changes. Since it is too time consuming to go through all the posts of a thread to under- stand every individual’s opinion, summarizing on- line discussion forums becomes an important task that may benefit people including government pol- icy makers and social scientists. While text sum- marization has been extensively studied, summa- rizing noisy and subjective user-generated content is still an under-explored area. A vast body of work has been done on summarizing online prod- uct reviews, but because of the special proper- ties of product reviews, opinion summarization of product reviews tends to focus on product aspect identification and sentiment polarity classification. When it comes to summarizing general online dis- cussions, particularly discussions on controversial topics such as a policy or a social issue, the chal- lenges we face can be very different from summa- rizing product reviews. Table 1 shows a set of summary sen- tences selected by a state-of-the-art summarization method (Gillick and Favre, 2009) from a forum thread on criticizing parliament ministers sleeping in a meeting. We can see that the summary con- tains low-quality sentences and some sentences do not express opinions. This result shows that tradi- tonal text summarization techniques, which only consider text representativeness, may not be able to summarize opinions from online forums very well. In particular, we hypothesize that two important factors should be considered for summarizing on- line discussions. First, because forum posts are often noisy, with misspelling, broken …

Content truncated to 500 words