Incorporating lexicon knowledge into SVM learning to improve sentiment classification

Two typical approaches to sentiment analysis are lexicon look up and machine learning. Even though recent studies have shown that machine learning approaches in general outperform the lexicon look up approaches, completely ignoring the knowledge encoded in sentiment lexicons may not be optimal. We present an alternative method that incorporates sentiment lexicons as prior knowledge with machine learning approaches such as SVM to improve the accuracy of sentiment analysis. This paper also describes a method to automatically generate domain specific sentiment lexicons for this learning purpose. Our experiment results show that the domain specific lexicons we constructed lead to a significant accuracy improvement for our sentiment analysis task. We also offer a theoretical proof on why our approach works.