Discourse Level Explanatory Relation Extraction from Product Reviews Using First-Order Logic

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 946–957, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Discourse Level Explanatory Relation Extraction from Product Reviews Using First-order Logic Qi Zhang, Jin Qian, Huan Chen, Jihua Kang, Xuanjing Huang School of Computer Science Fudan University Shanghai, P.R. China {qz, 12110240030, 12210240054, 12210240059, xjhuang}@fudan.edu.cn Abstract Explanatory sentences are employed to clarify reasons, details, facts, and so on. High quality online product reviews usually include not only positive or negative opinions, but also a variety of explanations of why these opinions were given. These explanations can help readers get easily comprehensible informa- tion of the discussed products and aspect- s. Moreover, explanatory relations can also benefit sentiment analysis applications. In this work, we focus on the task of identi- fying subjective text segments and extracting their corresponding explanations from prod- uct reviews in discourse level. We propose a novel joint extraction method using first- order logic to model rich linguistic features and long distance constraints. Experimental results demonstrate the effectiveness of the proposed method. 1 Introduction Through analyzing product reviews with high help- fulness ratings assigned by readers, we find that a large number of explanatory sentences are used to clarify the causes, details, or consequences of opin- ions. According to the statistic based on the dataset we crawled from a popular product review website, more than 56.1% opinion expressions are further explained by other sentences. Since most consumers are not experts, these explanations would bring lots of helpful and easy comprehension information for them. Suggestions about writing a product review also advise authors to include not only whether they like or dislike a product, but also why.1 1http://www.reviewpips.com/ http://www.amazon.com/gp/community-help/customer- For example, let us consider the following snip- pets extracted from online reviews: Example 1: TVs with lower refresh rates may suffer from motion blur. If you’re watching a fast-paced football game, for example, you may notice a bit of blurring as the players run around the field. Example 2: The LED screen is highly reflec- tive. The reflection of my own face makes it very hard to see the subject I am trying to shoot. The first sentence of example 1 expresses negative opinion about refresh rate, which is one of the most important attributes of TV. The second sentence describes the consequence of it through an example. In example 2, detail descriptions are used to explain the reflection problem of the camera screen. Although, explanations provide valuable infor- mation, to the best of our knowledge, there is no existing work that deals with explanation extraction for opinions in discourse level. We think that if explanatory relations can be automatically identified from reviews, sentiment analysis applications may benefit from it. Existing opinion mining approaches mainly focus on subjective text. They try to de- termine the subjectivity and polarity of fragments of documents (e.g. a paragraph, a sentence, a phrase and a word) (Pang et al., 2002; Riloff et al., 2003; Takamura et al., 2005; …

Content truncated to 500 words