Modelo de detección de discurso de odio en Ecuador mediante clasificación supervisada de tweets y técnicas de NLP.

Abstract

Hate speech on social networks refers to the posting of offensive language, generally, directed at vulnerable groups in society. In this context, some governments have proposed legislative actions to require social media platforms to control content that may incite hatred and violence. In Ecuador, no governmental actions have been proposed to regulate content on social networks. On this basis, the present study develops a first approach to the development of a theoretical-practical framework to identify the hate speech of users of the Twitter platform in Ecuador. In addition, this research aims to generate empirical evidence for the debate and discussion of legal artifacts that would regulate the offensive content of posts generated in Ecuador. For this, three machine learning models based on NLP techniques are used. The results showed that the word embedding techniques are better adapted to the linguistic idiosyncrasy of Ecuador, and therefore, are more accurate in identifying Hate Speech with semantics typical of this geographical area. The transfer learning with the pretrained BERT (Bidirectional Encoder Representations from Transformers) model is better suited to identify Hate Speech in general linguistic contexts.