Multi-aspects Review Summarization Based on Identification of Important Opinions and their Similarity

Multi-aspects Review Summarization Based on Identification of Important Opinions and their Similarity ⋆ Ryosuke Tadano, Kazutaka Shimada, and Tsutomu Endo Department of Artificial Intelligence, Kyushu Institute of Technology, 680-4 Kawazu, Iizuka, Fukuoka 820-8502, Japan {r tadano, shimada, endo}@pluto.ai.kyutech.ac.jp Abstract. The development of the Web services lets many users easily provide their opinions recently. Automatic summarization of enormous sentiments has been expected. Intuitively, we can summarize a review with traditional document summarization methods. However, such methods have not well-discussed “aspects”. Basically, a review consists of sentiments with various aspects. We summarize reviews for each aspect so that the summary presents information without biasing to a specific topic. In this paper, we propose a method for multi- aspects review summarization based on evaluative sentence extraction. We handle three fea- tures; ratings of aspects, the tf-idf value, and the number of mentions with a similar topic. For estimating the number of mentions, we apply a clustering algorithm. By integrating these features, we generate a more appropriate summary. The experiment results show the effectiveness of our method. Keywords: sentiment analysis, multi-aspects review summarization, ratings of aspects, im- portant sentence extraction, opinion integration. 1 Introduction As Web services like CGMs have become widely used, people can easily post reviews for products or services. Although handling these information (evaluative information) has become necessary, there exists too much information on the Web. Therefore, extracting information that users want and summarizing them have been expected recently. Intuitively, we can summarize a review with traditional document summarization methods. For instance, Brandow et al. (1995) have summa- rized a document by extracting sentences with some features such as the presence of signature words and the location in the document. For sentiment summarization, Pang and Lee (2004) have extracted all subjective sentences. They suggested that these extracted sentences could be used as summaries. However, a review basically consists of sentiments with various aspects (i.e., “image quality” and “usability” of a camera). Therefore, we need to extract information for each aspect in the case of review summarization. Aspect summarization can present information without biasing to a specific topic. We focus on multi-aspects review summarization in this research. We generate a summary by extracting important sentences and arranging them. Since we ex- tract summary sentences, we need to discuss which sentences are important and how to extract important sentences. In the case of treating multi-reviews, we need to handle the redundant in- formation. Pang and Lee (2008) have reported that while in traditional summarization redun- dant information is discarded, in sentiment summarization redundancy indicates the importance of opinions. Therefore, we treat redundancy as a feature for decision of important sentences. We assume that reviews we treat in this research have multiple aspects and a reviewer gives ratings of the aspects (i.e., 0 to 5 stars). Reviewers also write free comments about the target. We leverage ⋆ This work was supported by Kayamori Foundation of Informational Science Advancement Copyright 2010 by Ryosuke Tadano, Kazutaka Shimada, and Tsutomu Endo PACLIC 24 Proceedings 685 Figure 1: …

Content truncated to 500 words