Combining a large sentiment lexicon and machine learning for subjectivity classification

Most previous work on subjectivity/sentiment classification bases on either machine learning techniques (such as SVM, Maximum Entropy, Naive Bayes, etc.) or general sentiment lexicons. This paper presents a novel approach to combine a large sentiment lexicon and machine learning techniques for opinion analysis: 1) a large sentiment lexicon is automatically adjusted according to training data; 2) machine learning techniques are used to learn models on training data; 3) the results given by machine learning classifiers and the supervised lexicon-based classifier are combined to get better results. The experiments with the NTCIR data show that our approach significantly outperforms the baselines on subjectivity classification, i.e. the adjusted large sentiment lexicon shows good performance and its combination with machine learning techniques shows further improvement.