Técnicas de aprendizado de máquina e mineração de textos aplicadas para a previsão da avaliação de consumidores de e-commerce

Abstract

E-commerce has grown in recent years in Brazil, becoming increasingly present in the daily lives of millions of people. In an ever-competitive market, companies seek to differentiate themselves by providing better customer service, making the analysis of data generated by their consumers vital, regardless of its formats. Sentiment analysis combines concepts of natural language processing and machine learning to extract opinions contained in texts through classification, traditionally associated with a "positive" or "negative" sentiment. This thesis aims to analyze the sentiments of customers for a specific product on Amazon using techniques of natural language processing and machine learning to analyze consumer reviews/opinions, classifying their sentiment regarding the product. Different pre-processing techniques were tested to assess the impact on the obtained results. Additionally, five machine learning techniques were used to classify the texts: Decision Tree, Naïve Bayes Bernoulli and Naïve Bayes Multinomial, K-nearest Neighbors, and Support Vector Machines. The results of applying different pre-processing scenarios showed that the techniques directly and differently affect each algorithm. The algorithms achieved performances considered adequate in the four scenarios tested, with the overall best result obtained by Naïve Bayes Multinomial, achieving 90%, 92%, 91%, and 89% accuracy in each of the scenarios.