Aplicação de meta-atributos para detecção de discurso de ódio

Abstract

With the advent of online social networking, user interaction spaces are emerging. Such spaces can range from comments on a video site to the discussion section on a news site, for example. In these environments, people can chat, expose their opinions, discuss ideas, etc. However, groups are emerging and taking advantage of anonymity to spread hate speech, disseminating xenophobic and racist views, among others. Individuals who share such messages tend to hide their words with special characters or even sarcasm. It makes a difficult task for basic algorithms to block the content. Thus, it is necessary a way to detect these speeches automatically, since the amount of messages published daily is very large, such task becomes impracticable. Using preprocessing methods for features and meta-features extraction, this work obtain promising results in the area of hate speech detection in texts. The proposed method consists of stemization and removal of stopwords for data pre-processing and meta-features extraction with KNN algorithm. Using 10-fold cross validation and the SVM classifier, the proposed method achieves improvements of up to 9.67% over the baseline presented.