Measuring the Effect of Discourse Relations on Blog Summarization

International Joint Conference on Natural Language Processing, pages 1401–1409, Nagoya, Japan, 14-18 October 2013. Measuring the Effect of Discourse Relations on Blog Summarization Shamima Mithun Concordia University Montreal, Quebec, Canada shamima.mithun@gmail.com Leila Kosseim Concordia University Montreal, Quebec, Canada kosseim@encs.concordia.ca Abstract The work presented in this paper attempts to evaluate and quantify the use of dis- course relations in the context of blog summarization and compare their use to more traditional and factual texts. Specif- ically, we measured the usefulness of 6 discourse relations - namely compari- son, contingency, illustration, attribution, topic-opinion, and attributive for the task of text summarization from blogs. We have evaluated the effect of each relation using the TAC 2008 opinion summariza- tion dataset and compared them with the results with the DUC 2007 dataset. The re- sults show that in both textual genres, con- tingency, comparison, and illustration re- lations provide a significant improvement on summarization content; while attribu- tion, topic-opinion, and attributive rela- tions do not provide a consistent and sig- nificant improvement. These results indi- cate that, at least for summarization, dis- course relations are just as useful for in- formal and affective texts as for more tra- ditional news articles. 1 Introduction It is widely accepted that in a coherent text, units should not be understood in isolation but in rela- tion with each other through discourse relations that may or may not be explicitly marked. A text is not a linear combination of textual units but a hi- erarchial organized group of units placed together based on informational and intentional relations to one another. According to (Taboada, 2006), “Dis- course relations - relations that hold together dif- ferent parts (i.e. proposition, sentence, or para- graph) of the discourse - are partly responsible for the perceived coherence of a text”. For example, in the sentence “If you want the full Vista expe- rience, you’ll want a heavy system and graphics hardware, and lots of memory”, the first and sec- ond clauses do not bear much meaning indepen- dently; but become more meaningful when we re- alize that they are related through the discourse re- lation condition. Discourse relations have been found useful in many NLP applications such as natural language generation (e.g. (McKeown, 1985)) and news summarization (e.g. (Blair-Goldensohn and McK- eown, 2006; Bosma, 2004)) to improve coherence and better simulate human writing. However, most of these work have been developed for formal, well-written and factual documents. Text avail- able in the social media are typically written in a more casual style, are opinionated and speculative (Andreevskaia et al., 2007). Because of this, tech- niques developed for formal texts, such as news articles, often do not behave as well when deal- ing with informal documents. In particular, news articles are more uniform in style and structure; whereas blogs often do not exhibit a stereotypi- cal discourse structure. As a result, for blogs, it is usually more difficult to identify and rank rel- evant units for summarization compared to news articles. Several work have …

Content truncated to 500 words