Predicting Opinion Dependency Relations for Opinion Analysis

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 345–353, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Predicting Opinion Dependency Relations for Opinion Analysis Lun-Wei Ku National Yunlin University of Science and Technology Douliou, Yunlin 64002, Taiwan lwku@yuntech.edu.tw Ting-Hao Kenneth Huang National Taiwan University No.1, Sec.4, Roosevelt Rd., Taipei, Taiwan tinghaoh@andrew.cmu.edu Hsin-Hsi Chen National Taiwan University No.1, Sec.4, Roosevelt Rd., Taipei, Taiwan hhchen@ntu.edu.tw Abstract Syntactic structures have been good fea- tures for opinion analysis, but it is not easy to use them. To find these features by su- pervised learning methods, correct syntac- tic labels are indispensible. Two possible sources to acquire syntactic structures are parsing trees and dependency trees. For the annotation processing, parsing trees are more readable for annotators, while depen- dency trees are easier to use by programs. To use syntactic structures as features, this paper tried to annotate on human friendly materials and transform these annotations to the corresponding machine friendly ma- terials. We annotated the gold answers of opinion syntactic structures on the parsing tree from Chinese Treebank, and then pro- posed methods to find their corresponding dependency relations on the dependency trees generated from the same sentence. With these relations, we could train a mod- el to annotate opinion dependency relations automatically to provide an opinion depen- dency parser, which is language indepen- dent if language resources are incorporated. Experiment results show that the annotated syntactic structures and their corresponding dependency relations improve at least 8% of the performance of opinion analysis. 1 Introduction Opinion analysis has drawn much attention in research communities of machine learning and natural language processing. In the early stages, words in documents were used as the main features (Pang et al., 2002). Some opinion dic- tionaries were created for this demand (Ku et al., 2007). However, researchers soon realized that word features were not sufficient for ac- quiring good performances, so they started to include syntactic structures and semantic in- formation (Qiu et al., 2008). Their researches showed that linguistic knowledge is helpful in determining opinions. For various applications related to opinions, syntactic structures have become powerful tools for extracting useful clues. To find opi- nions in product reviews, modification rela- tions were used to identify the product and their features (Lu et al., 2009), e.g., a good price (feature) of this camera (product). To find opinion holders and targets, templates and linguistic rules were adopted (Breck et al., 2007). To find more opinion words, dependen- cy relations were utilized (Qiu et al., 2011). Even when applying the basic negation rule that flips opinion polarity over, we need to find its modified word first by syntactic clues. However, we will show that syntactic relations do not directly suggest opinions. Syntactic relations are obtained usually from all kinds of syntax trees. Parsing trees (phrase structured) and dependency trees (grammatical) are the most commonly seen ones. Parsing trees are in-order trees which keep the order of words in sentences, so they are more readable for …

Content truncated to 500 words