Improving Term Weight with Normalized Class Mutual Information for Sentiment Classification

The relevance of keywords in a document is assessed by assigning a higher term weight to them. The conventional use of Tf and Idf in short documents may introduce biases. Various term weighting strategies have been proposed to enhance the meaning of term weights by addressing each component individually, aiming for a more nuanced and refined representation. This study enhances the efficiency of Tf-Idf term weighting for sentiment classification by incorporating Normalized Class Mutual Information (NCMI). In our proposed method, the NCMI value is computed to assess the discriminative power of each term in distinguishing between positive and negative classes. The experiment utilizes three datasets, Amazon, IMDB, Magazine, Books, DVD, Electronics, and Kitchen_housewares, along with two sentiment classifications, NB and Random forest. The results demonstrate that the NCMI can improve the classification accuracy and F1-measure.