Twitter Sentiment Classification Using Normalized Weighted Multinomial Naive Bayes

Twitter is a very popular micro-blogging website where individuals or group of people express their opinion and options about events, products and services leading to a pool of different contextual and unstructured messages. These messages hold different meaning or sentiments that can be useful in analysing how the public feels about certain products, events or services. Sentiment analysis involves the use of natural language processors or text analysers to extract and study subjective information in a word or group of words. The Bayesian model has been extensively used for determining sentiments in a sentence. However, the problem with the Bayesian model is its independent assumption. The independent assumption considers compound words which express a singular meaning as if each individual word of the compound word expresses a different meaning. This assumption causes a weight magnitude error during the classification and analysis of the sentiments in a message or tweet. This study addresses the weight magnitude error using the normalized weighted multinomial naïve Bayes (NWMNB). During the training of the model, the NWMNB evaluates the weight magnitude in each of the dataset, comparing and normalizing the weight (We) distribution before testing and classifying the tweet. The NWMNB was implemented and tested on benched marked data, results compared against results of other traditional naïve Bayes filters using the same dataset. The NWMNB out performs the traditional Naïve Bayes filters by 4.5 %.