Wikipedia based semantic smoothing for twitter sentiment classification

Abstract

Sentiment classification is one of the important and popular application areas for text classification in which texts are labeled as positive and negative. Moreover, Naïve Bayes (NB) is one of the mostly used algorithms in this area. NB having several advantages on lower complexity and simpler training procedure, it suffers from sparsity. Smoothing can be a solution for this problem, mostly Laplace Smoothing is used; however in this paper we propose Wikipedia based semantic smoothing approach. In our study we extend semantic approach by using Wikipedia article titles that exist in training documents, categories and redirects of these articles as topic signatures. Results of the extensive experiments show that our approach improves the performance of NB and even can exceed the accuracy of SVM on Twitter Sentiment 140 dataset.