Abstract
Opinion detection and opinion analysis is a challenging but important task. Such sentiment analysis can be done using traditional supervised learning methods such as naive Bayes classification and support vector machines (SVM) or unsupervised approaches based on a lexicon may be employed. Because lexicon-based sentiment analysis methods make use of an opinion dictionary that is a list of opinion-bearing or sentiment words, sentiment lexicons play a key role. Our work focuses on the task of generating such a lexicon. We propose several novel methods to automatically generate a general-purpose sentiment lexicon using a corpus-based approach. While most existing methods generate a lexicon using a list of seed sentiment words and a domain corpus, our work differs from these by generating a lexicon from scratch using probabilistic techniques and information theoretical text mining techniques on a large diverse corpus. We conclude by presenting an ensemble method that combines the two approaches. We evaluate and demonstrate the effectiveness of our methods by utilizing the various automatically-generated lexicons during sentiment analysis. When used for sentiment analysis, our best single lexicon achieves an accuracy of 87.60% and the ensemble approach achieves 88.75% accuracy, both statistically significant improvements over 81.60% with a widely-used sentiment lexicon.