Content vs. context for sentiment analysis: a comparative analysis over microblogs

Microblog content poses serious challenges to the applicability of traditional sentiment analysis and classification methods, due to its inherent characteristics. To tackle them, we introduce a method that relies on two orthogonal, but complementary sources of evidence: content-based features captured by n-gram graphs and context-based ones captured by polarity ratio. Both are language-neutral and noise-tolerant, guaranteeing high effectiveness and robustness in the settings we are considering. To ensure our approach can be integrated into practical applications with large volumes of data, we also aim at enhancing its time efficiency: we propose alternative sets of features with low extraction cost, explore dimensionality reduction and discretization techniques and experiment with multiple classification algorithms. We then evaluate our methods over a large, real-world data set extracted from Twitter, with the outcomes indicating significant improvements over the traditional techniques.