LCCT: A Semi-supervised Model for Sentiment Classification

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 546–555, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics LCCT: A Semi-supervised Model for Sentiment Classification Min Yang1 Wenting Tu1 Ziyu Lu1 Wenpeng Yin2 Kam-Pui Chow1 1Department of Computer Science, The University of Hong Kong, Hong Kong {myang,wttu,zylu,chow}@cs.hku.hk 2Center for Information and Language Processing, University of Munich, Germany wenpeng@cis.lmu.de Abstract Analyzing public opinions towards products, services and social events is an important but challenging task. An accurate sentiment an- alyzer should take both lexicon-level infor- mation and corpus-level information into ac- count. It also needs to exploit the domain- specific knowledge and utilize the common knowledge shared across domains. In addi- tion, we want the algorithm being able to deal with missing labels and learning from incom- plete sentiment lexicons. This paper presents a LCCT (Lexicon-based and Corpus-based, Co-Training) model for semi-supervised sen- timent classification. The proposed method combines the idea of lexicon-based learning and corpus-based learning in a unified co- training framework. It is capable of incor- porating both domain-specific and domain- independent knowledge. Extensive experi- ments show that it achieves very competitive classification accuracy, even with a small por- tion of labeled data. Comparing to state-of- the-art sentiment classification methods, the LCCT approach exhibits significantly better performances on a variety of datasets in both English and Chinese. 1 Introduction Due to the popularity of opinion-rich resources (e.g., online review sites, forums, blogs and the mi- croblogging websites), people express their opinions all over the Internet. Motivated by the demand of gleaning insights from such valuable data, a flurry of research devotes to the task of extracting people’s opinions from online reviews. Such opinions could be expressed on products, services or policies, etc (Pang and Lee, 2008). Existing sentiment analysis approaches can be divided into two categories based on the source of information they use: the lexicon- based approach (Turney, 2002; Dave et al., 2003) and the corpus-based approach (Pang et al., 2002; Blitzer et al., 2007; Wan, 2009). The lexicon-based approach counts positive and negative terms in a re- view based on the sentiment dictionary and classi- fies the document as positive if it contains more pos- itive terms than negative ones. On the contrary, the corpus-based approach uses supervised learning al- gorithms to train a sentiment classifier. Further study (Kennedy and Inkpen, 2006; Andreevskaia and Bergler, 2008; Qiu et al., 2009) shows that corpus-based and lexicon-based approaches have complementary performances. Specifically, the corpus-based approach has high precision but low recall on positive instances, while the lexicon-based approach has high recall but low precision on positive instances. In fact, corpus- based approaches are over conservative in classify- ing instances as positive, because positive reviews usually contain many neutral statements. In contrast, the lexicon-based approaches tend to classify nega- tive or neutral instances as positive when there are a few positive words appear in the document. It mo- tivates us to develop a new approach that achieves good …

Content truncated to 500 words