Word polarity classification by fine-tuning of query words

Analysis of the polarity of words is the basis of determining the affective orientation of a text. A novel word polarity classification method is proposed. We use the PLSA model to construct the similarity matrix of query words and benchmark words. Then the polarity value of each query word is estimated by the combination of polarities of related benchmark words. To improve the polarity values, we cluster query data according to their PLSA similarities and balance the polarities in one cluster. This fine-tuning step gives us a much more accurate estimation of polarity with a fixed size of corpus. Experimental results show the advantage of the proposed method.