Sentiment classification of movie reviews using dual training and dual predition

BOW (Bag-of-Words) is the most common machine learning method used for sentiment classification. However, this model does not address the problem of polarity shift due to fundamental limitations. This is one of the reasons which affect its overall accuracy. In this paper, a method is discussed which uses dual training and dual prediction for sentiment classification while addressing polarity shift. The terms dual training and dual prediction refer to usage of both original review sample and opposite review sample for training and prediction. The opposite review samples are polarity opposite to original sample. These samples are artificially generated by polarity reversion. The overall method is evaluated on a movie review dataset and has shown noticeable improvements in comparison to machine learning methods based on bag of words model.