Sentiment analysis for online reviews using conditional random fields and support vector machines

Abstract

Sentiment analysis of online reviews is an important way of mining useful information from the Internet. Despite several advantages, the accuracy of sentiment analysis based on a domain dictionary relies on the comprehensiveness and accuracy of the dictionary. Instead of creating a domain dictionary, we propose an approach for online review sentiment classification, which uses a conditional random field algorithm to extract the emotional characteristics from fragments of the review. The characteristic (feature) words are then weighted asymmetrically before a support vector machine classifier is used to obtain the sentiment orientation of the review. In our experiments, the average accuracy reached 90%, showing that using sentiment feature fragments instead of whole reviews and weighting the characteristic words asymmetrically can improve the sentiment classification accuracy.