Aprendizado profundo para classificação de sentimentos em microblogs

Abstract

The Internet today represents one of the leading means of communication and information sharing. In Brazil and in several other countries, the vast majority of the population uses it for a variety of activities. Examples include social networks and microblogs, environments where there are frequent manifestation of opinions and discussions about products, services or other subjects of general interest. In this context, the task of sentiment analysis may label opinions expressed in such vehicles as positive, negative or neutral, using algorithms of natural language processing (NLP) and machine learning. The objective of this study was to investigate the application of deep learning techniques to sentiment classification on a Twitter corpus, considering subjects of interest to the population. Specifically, this study considered the manifestation of opinions about the Brazilian presidential elections of 2018. Data collection was done through Tweepy, an application programming interface (API) provided by Twitter, on all days of televised debates. The database was preprocessed using NLP techniques. To perform the experiments, a relevant set of data classification algorithms was selected from the literature (naive Bayes, decision tree, logistic regression and support vector machines), and compared with deep neural networks. The results obtained showed that the logistic regression classifier achieved the best performance among traditional algorithms, with an averaged accuracy of 54%. The predictive performance of the deep neural network was equivalent, achieving the same 54% of averaged accuracy in a given parameter setting. This result is promising because there is a large space for improving the performance of such a technique as a small amount of parameters was studied. In addition, the poor performance of other techniques, such as the naive Bayes, show that the number of labeled tweets (Twitter posts) on the database is still small and that the overall performance of the classification techniques can be improved by labeling more of them.