Classification of Inconsistent Sentiment Words using Syntactic Constructions

An important problem in sentiment analysis are inconsistent words. We define an inconsistent word as a sentiment word whose dictionary polarity is reversed by the sentence context in which it occurs. We present a supervised machine learning approach to the problem of inconsistency classification, the problem of automatically distinguishing inconsistent from consistent sentiment words in context. Our first contribution to inconsistency classification is that we take into account sentence structure and use syntactic constructions as features – in contrast to previous work that has only used word-level features. Our second contribution is a method for learning polarity reversing constructions from sentences annotated with polarity. We show that when we integrate inconsistency classification results into sentence-level polarity classification, performance is significantly increased.