Summary

Hate speech spread through social media such as Twitter deserves special attention, as its increase may be related to the rise in hate crimes. Of the 11 categories of discrimination contemplated by the Spanish Ministry of Internal Affairs, the second in which the most hate crimes are registered per year is political ideology. However, this category falls outside of most action plans to study and combat hate crimes. The same occurs in the case of academic works since most focus on analyzing and detecting hate in English and at a general level. The few authors who have targeted their studies to a single type of hate to improve accuracy, have focused on racism, xenophobia, or gender discrimination, but never on political ideology. Furthermore, the detection prototypes developed so far have not used databases generated manually by various coders. This paper aims to overcome these limitations, developing and evaluating an automatic hate speech detector on Twitter in Spanish for reasons of ideological discrimination, using supervised machine learning techniques. For this, we developed a total of eight predictive models from an ad-hoc generated training corpus, and making use of shallow modelling, but also deep learning, which has allowed to improve the final performance of the prototype. In addition, the development of the corpus allowed us to observe that 16.2% of the sample, collected in autumn 2019 and manually analyzed, included some type of ideological hatred.