Extracting Aspect-Evaluation and Aspect-Of Relations in Opinion Mining

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 1065–1074, Prague, June 2007. c⃝2007 Association for Computational Linguistics Extracting Aspect-Evaluation and Aspect-of Relations in Opinion Mining Nozomi Kobayashi ∗ Kentaro Inui, and Yuji Matsumoto Graduate School of Information Science, Nara Institute of Science and Technology 8916-5 Takayama, Ikoma, Nara, 630-0192, Japan {nozomi-k,inui,matsu}@is.naist.jp Abstract The technology of opinion extraction allows users to retrieve and analyze people’s opin- ions scattered over Web documents. We de- fine an opinion unit as a quadruple consist- ing of the opinion holder, the subject being evaluated, the part or the attribute in which the subject is evaluated, and the value of the evaluation that expresses a positive or neg- ative assessment. We use this definition as the basis for our opinion extraction task. We focus on two important subtasks of opinion extraction: (a) extracting aspect-evaluation relations, and (b) extracting aspect-of re- lations, and we approach each task using methods which combine contextual and sta- tistical clues. Our experiments on Japanese weblog posts show that the use of contex- tual clues improve the performance for both tasks. 1 Introduction The explosive increase in Web communication has attracted increasing interest in technologies for auto- matically mining personal opinions from Web doc- uments such as product reviews and weblogs. Such technologies would benefit users who seek reviews on certain consumer products of interest. Previous approaches to the task of mining a large- scale document collection of customer opinions (or ∗ Currently, NTT Cyber Space Laboratories, 1-1, Hikarinooka, Yokosuka, Kanagawa, 239-0847 Japan reviews) can be classified into two approaches: Doc- ument classification and information extraction. The former is the task of classifying documents or pas- sages according to their semantic orientation such as positive vs. negative. This direction has been form- ing the mainstream of research on opinion-sensitive text processing (Pang et al., 2002; Turney, 2002, etc.). The latter, on the other hand, focuses on the task of extracting opinions consisting of information about, for example, ⟨who feels how about which as- pect of what product⟩ from unstructured text data. In this paper, we refer to this information extraction- oriented task as opinion extraction. In contrast to sentiment classification, opinion extraction aims at producing richer information and requires an in- depth analysis of opinions, which has only recently been attempted by a growing but still relatively small research community (Yi et al., 2003; Hu and Liu, 2004; Popescu and Etzioni, 2005, etc.). Most previous work on customer opinion ex- traction assumes the source of information to be customer reviews collected from customer review sites (Popescu and Etzioni, 2005; Hu and Liu, 2004; Liu et al., 2005). In contrast, in this paper, we con- sider the task of extracting customer opinions from unstructured weblog posts. Compared with extrac- tion from review articles, extraction from weblogs is more challenging because weblog posts tend to exhibit greater diversity in topics, goals, vocabu- lary, style, etc. and are much more likely to in- clude …

Content truncated to 500 words