Creating Domain-Specific Sentiment Lexicons via Text Mining

Sentiment analysis aims to identify and categorize customer’s opinion and judgments using either traditional supervised learning techniques or unsupervised approaches. Traditionally, Sentiment Analysis is performed using machine learning techniques such as a naive Bayes classification or support vector machines (SVM), or could make use of a sentiment lexicon, that is, a list of words that are mapped to a sentiment score. Our work focuses on generating a domain-specific lexicon using probabilities and information theoretic techniques. By employing text mining, we overcome the poor performance of transferred supervised machine learning techniques and remove the need to adapt an existing lexicon while maintaining accuracy. We show that text mining techniques performs as well as traditional approaches and we demonstrate that domain specific lexicons perform better than general lexicons in a sentiment analysis task. We further review and compare the generated lexicons.