Identificação automática de comentários tóxicos em discussões online: uma análise de toxidade utilizando processamento de linguagem natural e aprendizado de máquina

Abstract

This work addresses the classification of toxic comments present on social media platforms and their impact, using the "Jigsaw Toxic Comment Classification Challenge" dataset from the Kaggle platform. The issue of toxic comments is discussed in relation to the online environment and its negative influence. The study aims to contribute to effective moderation of toxicity on digital platforms through machine learning classification models. The methodology involved initial training, where each model was assigned to a specific class, and those with the highest F1- score were selected for hyperparameter refinement. The algorithms SVM, RF, LR, and LSTM were used, along with the vectorizers TFIDF, CountVectorizer, and Tokenizer (exclusively for LSTM). The results showed satisfactory outcomes, surpassing, in some cases, the expectations of the literature, with cross-validation confirming the robustness of the models. It is concluded that the objectives of the work were achieved, and it is suggested for future analyses to incorporate techniques used in the competition and literature to enhance performance in the competition dataset.