Polarity reinforcement: Sentiment polarity identification by means of social semantics

We propose a combination of machine learning and socially constructed concepts for the task of sentiment polarity identification. Detecting words with polarity is difficult not only due to limitations in current sentiment dictionaries but also due to the colloquial terms that are often used. Current approaches disregard the dynamics of language, i.e. that new words are often created comprising different polarities. In fact, the online community is very creative in coining terms about certain subjects such as ldquotweetuprdquo (a request by a user to meet with friends via Twitter) or ldquowhackrdquo (Street slang, meaning bad). Our approach utilizes a user generated dictionary of urban term definitions as a resource for polarity concepts. Therefore, we are not only able to map newly created words to their respective polarity but also enhance common expressions with additional features and reinforce the polarity, strengthening our initial finding. We empirically show that the use of polarity reinforcement improves the sentiment classification.