Generalized Sentiment-Bearing Expression Features for Sentiment Analysis

In this work, we propose a novel approach to extract sentiment-bearing expression features derived from dependency structures. Rather than directly use dependency relations generated by a parser, we propose a set of heuristic rules to detect both explicit and implicit negations in the text. Then, three patterns are defined to support generalized sentiment-bearing expressions. By altering existing dependency features with detected negations and generalized sentimentbearing expressions we are able to achieve more accurate sentiment polarity classification. We evaluate the proposed approach on three labeled collections of different lengths, and measure the gain from the generalized dependency features when used in addition to the bag-of-words features. Our results demonstrate that generalized dependency-based features are more effective when compared to standard features. Using these we are able to surpass the state-of-the-art in sentiment classification.