Unsupervised Twitter Sentiment Classification
Mihaela Dînșoreanu, Andrei Bacu · 2014
Sentiment classification is not a new topic but data sources having different characteristics require customized methods to exploit the hidden existing semantic while minimizing the noise and irrelevant information. Twitter represents a huge pool of data having specific features. We propose therefore an unsupervised, domain-independent approach, for sentiment classification on Twitter. The proposed approach integrates NLP techniques, Word Sense Disambiguation and unsupervised rule-based classification. The method is able to differentiate between positive, negative, and objective (neutral) polarities for every word, given the context in which it occurs. Finally, the overall tweet polarity decision is taken by our proposed rule-based classifier. We performed a comparative evaluation of our method on four public datasets specialized for this task and the experimental results obtained are very good compared to other state-of-the-art methods, considering that our classifier does not use any training corpus.