SENTIMENT ANALYSIS OF ONLINE USER REVIEWS
Divya Kumari, Himmat Singh, Aditya Mishra · International Journal of Modern Trends in Engineering and Research · 2016
The project aims to improve the existing methods that are being employed in the field of sentiment analysis of online user reviews. The proposed method is a dual training algorithm that analyses both the original and reversed training reviews in pairs for learning a sentiment analyzer. The algorithm, in addition to the positive and negative aspects of a review also recognizes the neutral aspect (a three-class classification). The model called dual sentiment analysis addresses the polarity shift problem in sentiment classification. Keywords - Dual sentiment analysis, Bag-of-words, Original, Reversed, Sentiment classifier I. INTRODUCTION Sentiment analysis - It refers the use of natural language processing techniques, text analysis and computational linguistics to check and extract subjective information in the source materials. In recent years, the volume of online reviews available on the Internet has grown by leaps and bounds. Sentiment analysis and opinion mining is becoming a hotspot in the field of data mining and natural language processing. Sentiment classification is a basic task in sentiment analysis, with its aim to classify the sentiments (either positive or negative) of a given text or data. The general practice in sentiment classification involves the techniques and methods in traditional topic-based text classification, where the Bag-of words (BOW) model is generally used for text representation. The statistical machine learning algorithms (like naive baye's, maximum entropy classifier, and support vector machines) are then applied to train a sentiment classifier. Still a large number of problems exist while using the conventional BOW model. A large number of researches in sentiment analysis aimed at improving BOW by including linguistic. However, due to the fundamental deficiencies in BOW, most of the efforts showed very little effects in enhancing the classification accuracy as the two sentimentally opposite texts are considered very similar by the Bag-of-Words representation. This is the main reason why standard machine learning algorithms often fail while facing the polarity shift problem. We present a very simple but efficient model, called Dual Sentiment Analysis (DSA), to mark the problem of polarity shift in sentiment classification technique. By considering the property, that, sentiment classification has two opposite class tags (i.e., positive and negative), we will first suggest a data expansion technique by creating sentiment reversed reviews. The original and reversed reviews are in a one-to-one correlation. Thereafter, we suggest a dual training algorithm (DT) and a dual prediction algorithm (DP) respectively, to bring in use the original and reversed illustrative in pairs to train a statistical classifier and make predictions. In dual training algorithm, the classifier is learnt by augmenting a combination of likelihoods or probabilities of the original and reversed training data set. In DP, predictions are made by taking in consideration the two sides of one review. That is, we measure not only how positive or negative the original review is, but we also check as to how negative or positive the reversed review is. We further extend our DSA frame from polarity (positive vs. negative) classification to 3-class, i.e. positive vs. negative vs. Neutral, sentiment classification, by taking the neutral reviews into consideration in both dual training algorithm and dual prediction algorithm. To reduce DSA's dependency on an external antonym dictionary, we finally develop a corpus-based method for constructing a pseudo-antonym dictionary. The pseudo-antonym dictionary is language-independent and domain-adaptive. It thus makes the DSA model possible to be applied into a wide range of applications.