Harnessing WordNet Senses for Supervised Sentiment Classification

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1081–1091, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Harnessing WordNet Senses for Supervised Sentiment Classification Balamurali A R1,2 Aditya Joshi2 Pushpak Bhattacharyya2 1 IITB-Monash Research Academy, IIT Bombay 2Dept. of Computer Science and Engineering, IIT Bombay Mumbai, India - 400076 {balamurali,adityaj,pb}@cse.iitb.ac.in Abstract Traditional approaches to sentiment classifica- tion rely on lexical features, syntax-based fea- tures or a combination of the two. We pro- pose semantic features using word senses for a supervised document-level sentiment classi- fier. To highlight the benefit of sense-based features, we compare word-based representa- tion of documents with a sense-based repre- sentation where WordNet senses of the words are used as features. In addition, we highlight the benefit of senses by presenting a part-of- speech-wise effect on sentiment classification. Finally, we show that even if a WSD engine disambiguates between a limited set of words in a document, a sentiment classifier still per- forms better than what it does in absence of sense annotation. Since word senses used as features show promise, we also examine the possibility of using similarity metrics defined on WordNet to address the problem of not finding a sense in the training corpus. We per- form experiments using three popular similar- ity metrics to mitigate the effect of unknown synsets in a test corpus by replacing them with similar synsets from the training corpus. The results show promising improvement with re- spect to the baseline. 1 Introduction Sentiment Analysis (SA) is the task of prediction of opinion in text. Sentiment classification deals with tagging text as positive, negative or neutral from the perspective of the speaker/writer with respect to a topic. In this work, we follow the definition of Pang et al. (2002) & Turney (2002) and consider a binary classification task for output labels as positive and negative. Traditional supervised approaches for SA have explored lexeme and syntax-level units as features. Approaches using lexeme-based features use bag- of-words (Pang and Lee, 2008) or identify the roles of different parts-of-speech (POS) like adjec- tives (Pang et al., 2002; Whitelaw et al., 2005). Approaches using syntax-based features construct parse trees (Matsumoto et al., 2005) or use text parsers to model valence shifters (Kennedy and Inkpen, 2006). Our work explores incorporation of semantics in a supervised sentiment classifier. We use the synsets in Wordnet as the feature space to represent word senses. Thus, a document consisting of words gets mapped to a document consisting of corresponding word senses. Harnessing WordNet senses as features helps us address two issues: 1. Impact of WordNet sense-based features on the performance of supervised SA 2. Use of WordNet similarity metrics to solve the problem of features unseen in the training cor- pus The first points deals with evaluating sense-based features against word-based features. The second is- sue that we address is in fact an opportunity to im- prove the performance of SA that opens up because of the choice of sense space. …

Content truncated to 500 words