A System for Multilingual Sentiment Learning On Large Data Sets

Proceedings of COLING 2012: Technical Papers, pages 577–592, COLING 2012, Mumbai, December 2012. A System For Multilingual Sentiment Learning On Large Data Sets Alex CHENG1 Oles ZHU LY N 1 (1) Department of Computer Science, University of Toronto, Canada hyc@cs.toronto.edu, oles@cs.toronto.edu Abstract Classifying documents according to the sentiment they convey (whether positive or neg- ative) is an important problem in computational linguistics. There has not been much work done in this area on general techniques that can be applied effectively to multiple languages, nor have very large data sets been used in empirical studies of sentiment clas- sifiers. We present an empirical study of the effectiveness of several sentiment classification al- gorithms when applied to nine languages (including Germanic, Romance, and East Asian languages). The algorithms are implemented as part of a system that can be applied to multilingual data. We trained and tested the system on a data set that is substantially larger than that typically encountered in the literature. We also consider a generalization of the n-gram model and a variant that reduces memory consumption, and evaluate their effectiveness. Keywords: sentiment, classification, multilingual, empirical verification. 577 1 Introduction Classifying text documents according to the sentiment they convey is an important problem in computational linguistics. Sentiment reflects the emotional content in the document or the attitude of the speaker to the subject matter in the document, and can be positive or negative. For example, “Thank you for the pleasant time we spent together” conveys a positive sentiment, while “I was devastated when you left” conveys a negative sentiment. Sentiment classifiers that can process massive amounts of data quickly and accurately have applications in many segments of society. Marketing and brand management firms that are interested in how consumers generally feel about particular companies and their products can apply sentiment classifiers to social media documents containing relevant keywords. Government agencies that monitor electronic communications in order to identify and locate dissidents can use sentiment classifiers to find subversive messages. To the best of our knowledge, there has not been much work done in this area on general techniques that can be applied effectively to multiple languages, nor have very large data sets been used in empirical studies of sentiment classifiers. In this paper, we present an empirical study of two sentiment classification algorithms applied to nine languages (including Germanic, Romance, and East Asian languages). One of these algorithms is a naive Bayes classifier, and the other is an algorithm that boosts a naive Bayes classifier with a logistic regression classifier, using majority vote. These algorithms are implemented as part of a system that can be applied to multilingual data. Our implementation is fast, allowing a large number of documents to be classified in a short amount of time, with high accuracy. Automatic sentiment classification of text documents requires that the documents be mod- eled in a way that is amenable to the algorithm being used. The typical approach is to model the documents using n-grams. In this paper, we consider …

Content truncated to 500 words