Language-Independent Sentiment Analysis Using Subjectivity and Positional Information

International Conference RANLP 2009 - Borovets, Bulgaria, pages 360–364 Language-Independent Sentiment Analysis Using Subjectivity and Positional Information Veselin Raychev and Preslav Nakov∗ Department of Mathematics and Informatics Sofia University “St Kliment Ohridski” 5, James Bourchier Blvd.,1164 Sofia, Bulgaria {veselin.raychev, preslav.nakov}@fmi.uni-sofia.bg Abstract We describe a novel language-independent ap- proach to the task of determining the polarity, positive or negative, of the author’s opinion on a specific topic in natural language text. In partic- ular, weights are assigned to attributes, individ- ual words or word bi-grams, based on their posi- tion and on their likelihood of being subjective. The subjectivity of each attribute is estimated in a two-step process, where first the probability of being subjective is calculated for each sentence containing the attribute, and then these proba- bilities are used to alter the attribute’s weights for polarity classification. The evaluation results on a standard dataset of movie reviews shows 89.85% classification accuracy, which rivals the best previously published results for this dataset for systems that use no additional linguistic in- formation nor external resources. Keywords Sentiment analysis, subjectivity identification, polarity classifi- cation, text categorization. 1 Introduction Recently, there has been growing research interest in determining the polarity, positive or negative, of the author’s opinion on a specific topic in natural lan- guage texts. Such analysis has various potential ap- plications ranging from components for web sites to business and government intelligence [6]. Previous re- search on document sentiment classification has shown that machine learning based classifiers perform much better compared to rule-based systems [7]. However, the task remains challenging since opinions are typi- cally expressed in a specific manner, using many rare words and language expressions. As previous research has shown [8], even words with a single occurrence on training can turn out to be good predictors on testing. As a result, the classification accuracy for sentiment analysis using machine learning approaches tends to be much lower compared to that for other text classi- fication tasks like topic identification. ∗Also: Department of Computer Science, National Uni- versity of Singapore, 13 Computing Drive, Singapore 117417, nakov@comp.nus.edu.sg 2 Related work Pang & al. [7] pioneered the field of sentiment analy- sis. They worked on a sentiment polarity classification task, choosing between a positive and negative label using Na¨ıve Bayes and support vector machines (svm), where each text document was represented as a bag- of-words with weights for word presence. They fur- ther tried to use negation, word positions and part-of- speech (pos) information without much success, and found that many techniques that typically help for topic classification negatively affected the accuracy for sentiment polarity. The experiments were carried out on a set of 2,000 movie reviews mined from the web, 1,000 positive and 1,000 negative, without explicit in- formation about polarity, i.e., without ranks, scores, or number of stars. The dataset was made publicly available1 and has since become the de-facto standard for training and evaluation in most of the subsequent research. In the case of movie reviews, sentiment polarity clas- sification has …

Content truncated to 500 words