Mining Opinion Words and Opinion Targets in a Two-Stage Framework

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1764–1773, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Mining Opinion Words and Opinion Targets in a Two-Stage Framework Liheng Xu, Kang Liu, Siwei Lai, Yubo Chen and Jun Zhao National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China {lhxu, kliu, swlai, ybchen, jzhao}@nlpr.ia.ac.cn Abstract This paper proposes a novel two-stage method for mining opinion words and opinion targets. In the first stage, we propose a Sentiment Graph Walking algo- rithm, which naturally incorporates syn- tactic patterns in a Sentiment Graph to ex- tract opinion word/target candidates. Then random walking is employed to estimate confidence of candidates, which improves extraction accuracy by considering confi- dence of patterns. In the second stage, we adopt a self-learning strategy to refine the results from the first stage, especially for filtering out high-frequency noise terms and capturing the long-tail terms, which are not investigated by previous meth- ods. The experimental results on three real world datasets demonstrate the effective- ness of our approach compared with state- of-the-art unsupervised methods. 1 Introduction Opinion mining not only assists users to make in- formed purchase decisions, but also helps busi- ness organizations understand and act upon cus- tomer feedbacks on their products or services in real-time. Extracting opinion words and opinion targets are two key tasks in opinion mining. Opin- ion words refer to those terms indicating positive or negative sentiment. Opinion targets represent aspects or attributes of objects toward which opin- ions are expressed. Mining these terms from re- views of a specific domain allows a more thorough understanding of customers’ opinions. Opinion words and opinion targets often co- occur in reviews and there exist modified relations (called opinion relation in this paper) between them. For example, in the sentence “It has a clear screen”, “clear” is an opinion word and “screen” is an opinion target, and there is an opinion relation between the two words. It is natural to identify such opinion relations through common syntactic patterns (also called opinion patterns in this pa- per) between opinion words and targets. For ex- ample, we can extract “clear” and “screen” by us- ing a syntactic pattern “Adj-{mod}-Noun”, which captures the opinion relation between them. Al- though previous works have shown the effective- ness of syntactic patterns for this task (Qiu et al., 2009; Zhang et al., 2010), they still have some lim- itations as follows. False Opinion Relations: As an example, the phrase “everyday at school” can be matched by a pattern “Adj-{mod}-(Prep)-{pcomp-n}-Noun”, but it doesn’t bear any sentiment orientation. We call such relations that match opinion patterns but express no opinion false opinion relations. Pre- vious pattern learning algorithms (Zhuang et al., 2006; Kessler and Nicolov, 2009; Jijkoun et al., 2010) often extract opinion patterns by frequency. However, some high-frequency syntactic patterns can have very poor precision (Kessler and Nicolov, 2009). False Opinion Targets: In another case, the phrase “wonderful time” can be matched …

Content truncated to 500 words