Recognizing Contextual Polarity in Phrase-Level Sentiment Analysis

This paper introduces a new approach to phrase-level sentiment analysis centered on the notion of contextual polarity. Instead of assigning a fixed sentiment to words or phrases, the method proceeds in two stages: it first determines whether a given expression is neutral or polar in context, and then, for those judged polar, disambiguates whether the polarity is positive or negative. Separating the neutral-versus-polar decision from the polarity-disambiguation decision lets the system reason about how context shapes sentiment.

Using this two-step design, the system automatically identifies the contextual polarity for a large subset of sentiment expressions, and achieves results that are significantly better than the baseline. By highlighting that the same word can carry different polarity depending on context, the work helped establish contextual polarity as a core consideration in sentiment analysis research.