Naive Bayes with Negation Handling for Sentiment Analysis of Twitter Data

This paper proposes an enhanced negation handling technique for sentiment analysis of Twitter data using the Naive Bayes algorithm and Part-of-Speech (POS) tagging. Negation handling detects negated content in text and can thus improve sentiment prediction. The proposed technique focuses on the detection of direct negation words such as "not" and "no", and implicitly negated content such as "could have been" and "should have been". The paper compares the proposed negation handling technique with an existing negation handling technique. The Sentiment140 dataset is used in the experiments. Naive Bayes with the proposed negation handling technique gave an accuracy of 77.57% while the accuracy of the Naive Bayes with the existing negation handling was 76.93% and the accuracy of the standard Naive Bayes was 76.12 % for a dataset of 1,000,000 tweets. Of these 1,000,000 tweets 197,381 contained one or more negations. Taking these negated tweets alone, the proposed technique showed an improvement over the existing technique and standard Naive Bayes with accuracies respectively of 76.51%, 75.98%, and 75.09%. The improvements and shortcomings of the proposed technique are discussed.