Low-Quality Product Review Detection in Opinion Summarization

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 334–342, Prague, June 2007. c⃝2007 Association for Computational Linguistics Low-Quality Product Review Detection in Opinion Summarization Jingjing Liu Nankai University Tianjin, China v-jingil@microsoft.com Yunbo Cao Microsoft Research Asia Beijing, China yucao@microsoft.com Chin-Yew Lin Microsoft Research Asia Beijing, China cyl@microsoft.com Yalou Huang Nankai University Tianjin, China huangyl@nankai.edu.cn Ming Zhou Microsoft Research Asia Beijing, China mingzhou@microsoft.com Abstract Product reviews posted at online shopping sites vary greatly in quality. This paper ad- dresses the problem of detecting low- quality product reviews. Three types of bi- ases in the existing evaluation standard of product reviews are discovered. To assess the quality of product reviews, a set of spe- cifications for judging the quality of re- views is first defined. A classification- based approach is proposed to detect the low-quality reviews. We apply the pro- posed approach to enhance opinion sum- marization in a two-stage framework. Ex- perimental results show that the proposed approach effectively (1) discriminates low- quality reviews from high-quality ones and (2) enhances the task of opinion summari- zation by detecting and filtering low- quality reviews. 1 Introduction In the past few years, there has been an increasing interest in mining opinions from product reviews (Pang, et al, 2002; Liu, et al, 2004; Popescu and Etzioni, 2005). However, due to the lack of editorial and quality control, reviews on products vary greatly in quality. Thus, it is crucial to have a mechanism capable of assessing the quality of reviews and detecting low-quality/noisy reviews. Some shopping sites already provide a function of assessing the quality of reviews. For example, Amazon1 allows users to vote for the helpfulness of each review and then ranks the reviews based on the accumulated votes. However, according to our survey in Section 3, users’ votes at Amazon have three kinds of biases as follows: (1) imbalance vote bias, (2) winner circle bias, and (3) early bird bias. Existing studies (Kim et al, 2006; Zhang and Va- radarajan, 2006) used these users’ votes for train- ing ranking models to assess the quality of reviews, which therefore are subject to these biases. In this paper, we demonstrate the aforemen- tioned biases and define a standard specification to measure the quality of product reviews. We then manually annotate a set of ground-truth with real world product review data conforming to the speci- fication. To automatically detect low-quality product re- views, we propose a classification-based approach learned from the annotated ground-truth. The pro- posed approach explores three aspects of product reviews, namely informativeness, readability, and subjectiveness. We apply the proposed approach to opinion summarization, a typical opinion mining task. The proposed approach enhances the existing work in a two-stage framework, where the low-quality re- view detection is applied right before the summari- zation stage. Experimental results show that the proposed ap- proach can discriminate low-quality reviews from high-quality ones effectively. In addition, the task of opinion summarization can be enhanced by de- tecting …

Content truncated to 500 words