Techniques for Improving the Performance of Unsupervised Approach to Sentiment Analysis

Abstract

In this work, few techniques were proposed to enhance the performance of unsupervised sentiment analysis method to categorize review reports into sentiment orientations (positive and negative). In review reports, generally negations can change the polarity of other terms in a sentence. Therefore, a new technique for handling negations was proposed. As it is seen that, the positions of terms in a report are also important i.e. the same term appearing at different positions in a report may convey different amount of sentiments. Thus, a new technique was proposed to assign weights to the terms depending on their positions of occurrences within a review. Again, another technique was proposed to use the presence of exclamatory marks in the reviews as the effects of exclamatory marks are equally important in categorizing review reports. After incorporating all these concepts in the first phase of the proposed method, in the second phase, analysis of sentiment orientations was done using cluster ensemble method. The proposed method was tested on a state-of-the-art Movie review dataset and 91.75% accuracy was achieved. A significant improvement over some of the unsupervised and supervised methods in terms of accuracy was achieved with incorporation of the new techniques.