Effective sentiment classification based on words and word senses

Sentiment analysis is an area which has gained a lot of attention in recent years, mainly due to the many practical applications it supports and a growing demand for such applications. Previous work has shown that word senses carry potentially useful information for sentiment analysis. However due to limitations in the existing methods to assign senses to words in open-domain texts, this word sense based approach has not demonstrated significant advantages over the traditional term-based approaches. Also most sentiment lexicons consider limited polarity values (usually 2 or 3), therefore may convey very little information for distinguishing between a sentence that is positive and one that is negative. This paper proposes to address these limitations by making use of term-based sentiment lexicons, by increasing the number of possible polarity values considered in lexicons, and by processing the sentences for local negation. Finally we propose a novel sentiment representation generated by merging the word features with the developed sentiment features in a single sequence. Our evaluations show that and our proposed improvements effectively increased the quality of the sentiment representations. However, these sentiment representations are still behind the best factored representations which exclude the sentiment features.