Hate Speech Detection During the 2023 Chilean Plebiscite Constitutional Reform

This work presents a classification of hate speech using Natural Language Processing approaches, including collecting and labeling the data, text augmentation using the back-translation technique to address the imbalanced class problem, and data preprocessing. This led to the creation of a model capable of classifying hate in tweets in Spanish from platform X in the context of the 2023 Chilean Constitutional Plebiscite. Results show that approaches based on Convolutional Neural Networks (CNNs) in 1 dimension obtained better results than approaches based on Machine Learning because CNNs can identify patterns and relations between consecutive words, making them more accurate in understanding the context of the tweet. The CNN model achieved an overall accuracy of 86% on the testing dataset, while Machine Learning approaches achieved between 79% and 81% on the testing dataset. It is essential to consider that since the dataset presents an imbalance in the classes, other metrics were also presented, such as precision, F1-score, and recall, where, once again, the best results were obtained using CNN.