Sentence Compression for Target-Polarity Word Collocation Extraction

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1360–1369, Dublin, Ireland, August 23-29 2014. Sentence Compression for Target-Polarity Word Collocation Extraction Yanyan Zhao1, Wanxiang Che2, Honglei Guo3, Bing Qin2, Zhong Su3 and Ting Liu2∗ 1: Department of Media Technology and Art, Harbin Institute of Technology 2: Department of Computer Science and Technology, Harbin Institute of Technology 3: IBM Research-China {yyzhao, bqin, tliu}@ir.hit.edu.cn, {guohl, suzhong}@cn.ibm.com Abstract Target-polarity word (T-P) collocation extraction, a basic sentiment analysis task, relies primarily on syntactic features to identify the relationships between targets and polarity words. A major problem of current research is that this task focuses on customer reviews, which are natural or spontaneous, thus posing a challenge to syntactic parsers. We address this problem by proposing a framework of adding a sentiment sentence compression (Sent Comp) step before performing T-P collocation extraction. Sent Comp seeks to remove the unnecessary information for senti- ment analysis, thereby compressing a complicated sentence into one that is shorter and easier to parse. We apply a discriminative conditional random field model, with some special sentiment- related features, in order to automatically compress sentiment sentences. Experiments show that Sent Comp significantly improves the performance of T-P collocation extraction. 1 Introduction Sentiment analysis deals with the computational treatment of opinion, sentiment and subjectivity in tex- t (Pang and Lee, 2008), and has received considerable attention in recent years (Liu, 2012). Target- Polarity word (T-P) collocation extraction, which aims to extract the collocation of a target and its cor- responding polarity word in a sentiment sentence, is a basic task in sentiment analysis. For example, in a sentiment sentence “这款相机拥有新颖的外形” (The camera has a novel appearance), “外形” (appearance) is the target, and “新颖” (novel) is the polarity word that modifies “外形” (appearance). According, ⟨外形, 新颖⟩ (⟨appearance, novel⟩) is the T-P collocation. Generally, T-P collocation is a basic and complete sentiment unit, thus is very useful for many sentiment analysis applications. Features derived from syntactic parse trees are particularly useful for T-P collocation extraction (Ab- basi et al., 2008; Duric and Song, 2012). For example, the syntactic relation “Adj ATT ↶ Noun”, where the ATT denotes an attributive syntactic relation, can be used as an important evidence to extract the T-P collocation ⟨外形, 新颖⟩ (⟨appearance, novel⟩) in the above sentiment sentence (Bloom et al., 2007; Qiu et al., 2011; Xu et al., 2013). However, one major problem of these approaches is the “naturalness” of sentiment sentences, that is, such sentences are more natural or spontaneous compared with normal sentences, thus posing a challenge to syntactic parsers. Accordingly, many wrong syntactic features have been produced and these can further result in the poor performance of T-P collocation extraction. Taking the sentence in Figure 1(a) as an example, because the word “多亏” (fortunately) is so chatty,1 the parsing result is wrong. Thus, are unable to extract the T-P collocation ⟨键盘, 好⟩ (⟨keyboard, good⟩). To solve the “naturalness” problem, we can train a parser on sentiment sentences. Unfortunately, an- notating such …

Content truncated to 500 words