Sentiment Analysis for Text Reviews

Abstract

The volume of user generated sentiment data on the web has massively increased. Systematic analysis of user reviews is very essential for extracting valuable user experience information that leads to decision making. Sentiment Analysis involves preprocessing, extracting, understanding, classifying and presenting the reviews expressed by users. It is the placement of words or syntactical structure that gives these reviews their meaning. Also the association among the words or semantic structure of a sentence adds to the meaning conveyed in a review. Bag-of-words (BoW) approach is popularly used for Sentiment analysis. BoW approach maps the terms in the reviews to structured term-document vectors but fails to handle the syntactic and semantic structure of sentences in reviews. This research work focuses on improving sentiment analysis accuracy by considering the syntactic and semantic structure of sentences in reviews. Sentiment Analysis accuracy depends majorly on the classifiers used. To improve sentiment classification accuracy, three classifiers based on relative frequency, average frequency and term frequency inverse document frequency (TFIDF) were proposed. First proposed classifies, Weighted Relative Term Frequency Sentiment Classifier (WRTFSC), classified the sentiment of terms on the basis of relative frequency of the term across positively tagged documents and negatively tagged documents. Second proposed classifier, Weighted Average Relative Term Frequency Sentiment Classifier (WARTFSC), overcame the drawback of WRTFSC s failure to handle biased reviews, containing terms with large frequencies. WARTFSC considered average relative frequency of terms. Third proposed classifier Weighted Sentiment-TFIDF (WSenti-TFIDF), classified terms on the basis of relative TFIDF across positively tagged documents and negatively tagged documents. Proposed classifiers are based on relative logarithmic differential term frequency and presence distribution.