Structural Opinion Mining for Graph-based Sentiment Representation

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1332–1341, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Structural Opinion Mining for Graph-based Sentiment Representation Yuanbin Wu, Qi Zhang, Xuanjing Huang, Lide Wu Fudan University School of Computer Science {ybwu,qz,xjhuang,ldwu}@fudan.edu.cn Abstract Based on analysis of on-line review corpus we observe that most sentences have compli- cated opinion structures and they cannot be well represented by existing methods, such as frame-based and feature-based ones. In this work, we propose a novel graph-based rep- resentation for sentence level sentiment. An integer linear programming-based structural learning method is then introduced to produce the graph representations of input sentences. Experimental evaluations on a manually la- beled Chinese corpus demonstrate the effec- tiveness of the proposed approach. 1 Introduction Sentiment analysis has received much attention in recent years. A number of automatic methods have been proposed to identify and extract opinions, emo- tions, and sentiments from text. Previous researches on sentiment analysis tackled the problem on vari- ous levels of granularity including document, sen- tence, phrase and word (Pang et al., 2002; Riloff et al., 2003; Dave et al., 2003; Takamura et al., 2005; Kim and Hovy, 2006; Somasundaran et al., 2008; Dasgupta and Ng, 2009; Hassan and Radev, 2010). They mainly focused on two directions: sentiment classification which detects the overall polarity of a text; sentiment related information extraction which tries to answer the questions like “who expresses what opinion on which target”. Most of the current studies on the second direc- tion assume that an opinion can be structured as a frame which is composed of a fixed number of slots. Typical slots include opinion holder, opinion expres- sion, and evaluation target. Under this representa- tion, they defined the task as a slots filling prob- lem for each of the opinions. Named entity recog- nition and relation extraction techniques are usually applied in this task (Hu and Liu, 2004; Kobayashi et al., 2007; Wu et al., 2009). However, through data analysis, we observe that 60.5% of sentences in our corpus do not follow the assumption used by them. A lot of important infor- mation about an opinion may be lost using those rep- resentation methods. Consider the following exam- ples, which are extracted from real online reviews: Example 1: The interior is a bit noisy on the free- way1. Example 2: Takes good pictures during the day- time. Very poor picture quality at night2. Based on the definition of opinion unit proposed by Hu and Liu (2004), from the first example, the information we can get is the author’s negative opin- ion about “interior” using an opinion expression “noisy”. However, the important restriction “on the freeway”, which narrows the scope of the opinion, is ignored. In fact, the tuple (“noisy”,“on the free- way”) cannot correctly express the original opinion: it is negative but under certain condition. The sec- ond example is similar. If the conditions “during the daytime” and “at night” are dropped, the extracted elements …

Content truncated to 500 words