Discovering the Discriminative Views: Measuring Term Weights for Sentiment Analysis

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 253–261, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Discovering the Discriminative Views: Measuring Term Weights for Sentiment Analysis Jungi Kim, Jin-Ji Li and Jong-Hyeok Lee Division of Electrical and Computer Engineering Pohang University of Science and Technology, Pohang, Republic of Korea {yangpa,ljj,jhlee}@postech.ac.kr Abstract This paper describes an approach to uti- lizing term weights for sentiment analysis tasks and shows how various term weight- ing schemes improve the performance of sentiment analysis systems. Previously, sentiment analysis was mostly studied un- der data-driven and lexicon-based frame- works. Such work generally exploits tex- tual features for fact-based analysis tasks or lexical indicators from a sentiment lexi- con. We propose to model term weighting into a sentiment analysis system utilizing collection statistics, contextual and topic- related characteristics as well as opinion- related properties. Experiments carried out on various datasets show that our approach effectively improves previous methods. 1 Introduction With the explosion in the amount of commentaries on current issues and personal views expressed in weblogs on the Internet, the field of studying how to analyze such remarks and sentiments has been increasing as well. The field of opinion mining and sentiment analysis involves extracting opin- ionated pieces of text, determining the polarities and strengths, and extracting holders and targets of the opinions. Much research has focused on creating testbeds for sentiment analysis tasks. Most notable and widely used are Multi-Perspective Question Answering (MPQA) and Movie-review datasets. MPQA is a collection of newspaper articles anno- tated with opinions and private states at the sub- sentence level (Wiebe et al., 2003). Movie-review dataset consists of positive and negative reviews from the Internet Movie Database (IMDb) archive (Pang et al., 2002). Evaluation workshops such as TREC and NT- CIR have recently joined in this new trend of re- search and organized a number of successful meet- ings. At the TREC Blog Track meetings, re- searchers have dealt with the problem of retriev- ing topically-relevant blog posts and identifying documents with opinionated contents (Ounis et al., 2008). NTCIR Multilingual Opinion Analy- sis Task (MOAT) shared a similar mission, where participants are provided with a number of topics and a set of relevant newspaper articles for each topic, and asked to extract opinion-related proper- ties from enclosed sentences (Seki et al., 2008). Previous studies for sentiment analysis belong to either the data-driven approach where an anno- tated corpus is used to train a machine learning (ML) classifier, or to the lexicon-based approach where a pre-compiled list of sentiment terms is uti- lized to build a sentiment score function. This paper introduces an approach to the senti- ment analysis tasks with an emphasis on how to represent and evaluate the weights of sentiment terms. We propose a number of characteristics of good sentiment terms from the perspectives of in- formativeness, prominence, topic–relevance, and semantic aspects using collection statistics, con- textual information, semantic associations as well as opinion–related properties of terms. …

Content truncated to 500 words