A comparative study on methods of sentiment analysis in tweets

Abstract

Twitter is a microblog on which users can post updates (tweets) to friends (followers). Sentiment Analysis has become an important field of study in this environment due to the sheer number of tweets available, which allows several applications such as monitoring of brands and products, forecasting political campaigns and even applications in the financial market. One of the great challenges in analyzing feelings is in the creation of predictive models that are able to classify tweets as positive, negative or neutral. The main models are based on natural language processing and machine learning. Given this context, this article aims to compare the performance of the following methods of sentiment analysis: machine learning, lexical dictionaries, emoticons, part-of-speech, ensembles and word embeddings. The objective is to indicate to the reader, among such approaches, which best suits the particularities of the tweets. The experiments were applied to two databases, Sanders and HCR. In both datasets, the procedures that obtained the best results were based on lexical dictionary and word embeddings with 79.09% and 79.36% of accuracy, respectively, for Sanders. While for HCR the result was 69.11% and 68.22% accuracy, respectively.

Keywords

• sentiment analysis
• machine learning
• lexical dictionaries
• part-of-speech
• ensembles
• word embeddings