Improving Document-Level Sentiment Classification Using Contextual Valence Shifters

Traditional sentiment feature extraction methods in document-level sentiment classification either count the frequencies of sentiment words as features, or the frequencies of modified and unmodified instances of each of these words. However, these methods do not represent the sentiment words' linguistic context efficiently. We propose a novel method and feature set to handle the contextual polarity of sentiment words efficiently. Our experiments on both movie and product reviews show a significant improvement in the classifier's performance (an overall accuracy increase of 2%), in addition to statistical significance of our feature set over the traditional feature set. Also, compared with other widely-used feature sets, most of our features are among the key features for sentiment classification.