Abstract

Sentiment classification is a unique process of text categorization whose objective is to
categorize a text related to the sentimental polarities of opinions it consists of constructive or adverse,
positive or negative. Bag-of-words (BOW) is now the major famous method to form text in numerical
machine learning methods in Sentiment Analysis (SA). On the other hand, the accuracy of BOW
sometimes still remains lesser because of various basic disadvantages in handling the polarity shift
problem. To deal with this problem, Dual Sentiment Analysis (DSA) is proposed in the recent work and it
is used for SA classification. However in DSA is should consider more difficult polarity shift patterns are
middle, subjunctive and sentiment-inconsistent sentences in forming reversed reviews. To solve this
problem a Three-Stage Model (TSM) is combined to Dual Sentiment Analysis (DSA) classifier named as
DSA-TSM. Initially divide the each set of documents into a set of sub-sentences and build a hybrid
classifier which creates the rules and machine learning model to distinguish precise and hidden polarity
shifts, correspondingly. Secondly, corpus based method is introduced to build a pseudo-antonym
dictionary in the direction of polarity shift elimination technique, to remove polarity shift in negations.
Finally, dual training and dual prediction algorithm is proposed for learning a sentiment classifier which
classifiers the polarity into three major classes such as positive-negative-neutral by considering the neutral
reviews into consideration. The results of the proposed DSA-TSM model are significantly improved when
compared to DSA schema and other methods in terms of accuracy, precision, recall and f-measure.