An empirical study of sentence features for subjectivity and polarity classification

• We extend and combine the body of empirical evidence regarding sentence subjectivity and polarity classification.
• We provide a comprehensive analysis of different sentence features using data from multiple benchmarks.
• Our analysis suggests possible use cases of the features considered and gives insights into their potential applicability.
• Some features claimed as effective in the literature have little value once the models incorporate other elemental features.

While a number of isolated studies have analysed how different sentence features are beneficial in Sentiment Analysis, a complete picture of their effectiveness is still lacking. In this paper we extend and combine the body of empirical evidence regarding sentence subjectivity classification and sentence polarity classification, and provide a comprehensive analysis of the relative importance of each set of features using data from multiple benchmarks. To the best of our knowledge, this is the first study that evaluates a highly diversified set of sentence features for the two main sentiment classification tasks.