Uma abordagem de auto-ML para análise de sentimentos na língua portuguesa

Abstract

Sentiment analysis is a raising field in academia and industry. It is a type of data extraction from knowledge that makes use of diverse natural language processing tasks and data mining tech- niques to obtain its results. Finding the best combination of these tasks for a given data set is not a trivial task as there may be a substantial amount of combinations to evaluate. Besides, the evaluation of each combination may require considerable computational power that can restrict the number of possible evaluations. Therefore, in this work, we first evaluate the combination of five NLP tasks and three classifiers in the domain of sentiment analysis using texts written in Portuguese. The experimental results showed that different combinations of preprocessing tasks can significantly affect the predictive performance of a classifier for a given dataset. Con- sequently, it is clear the importance of performing the joint evaluation of preprocessing tasks with classifiers when choosing which preprocessing tasks and classifier should be used for a dataset. Therefore, this paper also proposes a Automated Machine Learning (Auto-ML) ap- proach to search for a good combination of a classifier with natural languages processing tasks without having to evaluate all possible combinations. The proposed approach uses evolutio- nary algorithms and Bayesian optimization with the Bootstrap Corrected Cross-Validation Bias (BBC-CV) bias correction technique to find such a combination. The approach presented in this paper, evaluated with Portuguese written text data sets, showed performance equivalent or better than another Auto-ML tool.