Exploiting Structured Ontology to Organize Scattered Online Opinions

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 734–742, Beijing, August 2010 Exploiting Structured Ontology to Organize Scattered Online Opinions Yue Lu, Huizhong Duan, Hongning Wang, ChengXiang Zhai Department of Computer Science University of Illinois at Urbana-Champaign {yuelu2,duan9,wang296,czhai}@illinois.edu Abstract We study the problem of integrating scat- tered online opinions. For this purpose, we propose to exploit structured ontology to obtain well-formed relevant aspects to a topic and use them to organize scattered opinions to generate a structured sum- mary. Particularly, we focus on two main challenges in implementing this idea, (1) how to select the most useful aspects from a large number of aspects in the ontology and (2) how to order the selected aspects to optimize the readability of the struc- tured summary. We propose and explore several methods for solving these chal- lenges. Experimental results on two dif- ferent data sets (US Presidents and Digital Cameras) show that the proposed methods are effective for selecting aspects that can represent the major opinions and for gen- erating coherent ordering of aspects. 1 Introduction The explosive growth of online opinions raises in- teresting challenges for opinion integration and summarization. It is especially interesting to in- tegrate and summarize scattered opinions in blog articles and forums as they tend to represent the general opinions of a large number of people and get refreshed quickly as people dynamically gen- erate new content, making them valuable for un- derstanding the current views of a topic. However, opinions in blogs and forums are usually fragmental, scattered around, and buried among other off-topic content, so it is quite chal- lenging to organize them in a meaningful way. Traditional text summarization techniques gener- ate an unstructured list of sentences as a sum- mary, which cannot reveal representative opinions on different aspects of a topic or effectively facil- itate navigation into the huge opinion space. To address this limitation, recent work has shown the usefulness of generating a structured summary of opinions, in which related opinions are grouped into topical aspects with explicit labeling of all the aspects. A major challenge in producing such a structured summary is how to generate these as- pects for an arbitrary topic (e.g., products, politi- cal figures, policies, etc.). Intuitively, the aspects should be concise phrases that can both be easily interpreted in the context of the topic under con- sideration and capture the major opinions. How- ever, where can we find such phrases and which phrases should we select as aspects? Furthermore, once we selected aspects, how should we order them to improve the readability of a structured summary? One way to generate aspects is to clus- ter all the opinion sentences and then identify rep- resentative phrases in each cluster. Although as- pects selected in this way can effectively capture the major opinions, a major limitation is that it is generally hard to ensure that the selected phrases are well connected with the given topic (Chen and Dumais, 2000). In this paper, we propose a …

Content truncated to 500 words