Twitter sentiment analysis using deep learning

Abstract

Sentiment Analysis, also known as opinion mining, is a natural language processing (NLP) technique that helps in analyzing pieces of texts, or an entire text to determine the emotional attitude of the author behind writing the particular message, review, or tweet. It is, basically, a text classification method that helps in identifying whether an online writing carries a positive, negative, or neutral connotation to it. Sentiment analysis finds its significance in a lot of domains these days, the most popular among them being brand and social media monitoring. Most businesses are mindful of their customers’ opinions on their products, thereby working on their strengths, and enhancing customer experience, which in turn, benefits them. It even helps in finance and stock monitoring with the correct analysis of customer sentiments for a more beneficial investment into it. Market research and analyzing competitors in it is also a noticeably big advantage that can be achieved via sentiment analysis. In this article, we propose various methods to mine the twitter data, and eventually classify it as positive, negative, or neutral. Amongst the traditional machine learning methods, we have implemented the Support Vector Machine (SVM) algorithm, the Multinomial Naive Bayes, and the Gaussian Naive Bayes methods. For the deep learning methods, we have implemented a Multichannel CNN method, a combination of LSTM and CNN algorithm. We have also used BERT in combination with other deep learning algorithms. Among the lexicon-based approaches, we have used a polarity dictionary in combination with BERT to get superior results on our approach. To implement these models, the datasets used were the US airline twitter dataset which gave the best result at an accuracy of 80.12%, and the SemEval 2013 task A dataset which gave the best result at an accuracy of 65.54%.