A HYBRID APPROACH FOR SUPERVISED TWITTER SENTIMENT CLASSIFICATION.

Abstract

Micro blogging Websites like Twitter, Facebook have become rich source of opinions. This information can be leveraged by different communities to perform sentiment analysis. There is a need for automatically detecting the polarity of Twitter messages.  A semantic sentiment mining system is proposed to determine the contextual polarity of a sentence. This hybrid approach uses three different machine learning models for classifying the sentiment as positive and negative. The system presents more significant approach towards the contextual information in the document which is one of the drawbacks of the systems which are available for determining contextual information. The first model uses rule-based classification based on compositional semantic rules that identifies expression level polarity. The second one performs sense-based classification based on WordNet senses as features to Support Vector Machine classifier. Further to provide a meaningful classification, semantics are incorporated as additional feature into the training data by the interpolation method. Thus, the third model performs entity-level analysis based on concepts obtained. The outputs of three models are handled by knowledge inference system to predict the polarity of sentence. This system is expected to produce better results when compared to the baseline system performance. The system aims to predict consumer moods and the attitude in real-time which can be efficiently utilized by the firms to increase productivity and revenue.