Sentiment Classification Using a Sense Enriched Lexicon-based Approach

Abstract

The prominent approach in sentiment polarity classification is the Lexicon-based approach which relies on a dictionary to assign a score to subjective words. Most of the existing work use score of the most dominant sense in this process instead of using the contextually appropriate sense. The use of Word Sense Disambiguation (WSD) is less investigated in the sentiment classification tasks. This paper investigates the effect of integrating WSD into a Lexicon-based approach for Sentiment Polarity classification and compares it with the existing Lexicon-based approaches and the state-of-art supervised approaches. The lexicon used in this work is SentiWordNet v2.0. The proposed approach, called Sense Enriched Lexicon-based Approach (SELSA), uses a word sense disambiguation module to identify the correct sense of subjective words. Instead of using the score of the most frequent sense, it uses the score of the contextually appropriate sense only. For the purpose of comparison with the supervised approaches, the authors investigate Naïve Bayes (NB) and Support Vector Machines (SVM) classifiers which tend to perform better in earlier research. The performance of these classifiers is evaluated using Word2vec, Hashing Vectorizer, and bi-gram feature. The best-performing classifier-feature combination is used for comparison. All the evaluations are done on the Movie Review dataset. SELSA achieves an accuracy of 96.25% which is significantly better than the accuracy obtained by SentiWordNet-based approach without WSD on the same dataset. The performance of the proposed algorithm is also compared with the best-performing supervised classifier investigated in this work and earlier reported works on the same dataset. The results reveal that the SVM classifier performs better than SentiWordNet approach without WSD. However, after incorporating WSD the performance of the proposed Lexicon-based approach is significantly improved and it surpasses the best-performing supervised classifier (SVM with bi-gram features).