A context-based algorithm for sentiment analysis

With netizens continuing to express a range of opinions and making assessments online, it has become a challenge to mine sentiments accurately from the ever-multiplying Big Data. We present a context-driven sentiment analysis scheme with the objective of refining the degree of subjectivity during sentiment analysis. The essence of our scheme is to capture in a stable manner, the mutual influence of the sentiments of neighbouring words on the sentiment of each word in a document. A parametric influence function combines the native sentiment score of each word with the context-derived sentiment score obtained from surrounding words. We apply a genetic algorithm to fine tune the parameters of the influence function so as to obtain the best possible accuracy for a given corpus. The experimental results on hotel reviews extracted from Tripadvisor.com show an average accuracy of 73.2% which is 3.6% more than the results obtained from the baseline sentiment analysis approach using native scores obtained from SentiWordNet. Though the improvement is small, it re-affirms our belief that contextual information provides valuable reinforcement of sentiment scores especially with regard to the borderline cases where words show near neutral sentiments. We also present a comparison with alternative sentiment analysis approaches that shows the strength of our proposed context-based sentiment analysis approach.