Uncertainty Quantification of Text Classification in a Multi-Label Setting for Risk-Sensitive Systems

Proceedings of Recent Advances in Natural Language Processing, pages 41–48 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_005 41 Multilingual Racial Hate Speech Detection Using Transfer Learning Abinew Ali Ayele1,2, Skadi Dinter1, Seid Muhie Yimam1, Chris Biemann1 1Language Technology Group, Universität Hamburg, Germany, 2Faculty of Computing, BiT, Bahir Dar University, Ethiopia {abinew.ali.ayele, seid.muhie.yimam, chris.biemann}@uni-hamburg.de, skadi.dinter@posteo.de Abstract The rise of social media eases the spread of hateful content, especially racist content with severe consequences. In this paper, we analyze the tweets targeting the death of George Floyd in May 2020 as the event accelerated debates on racism globally. We focus on the tweets published in French for a period of one month since the death of Floyd. Using the Yandex Toloka platform, we annotate the tweets into categories as hate, offensive or normal. Tweets that are offensive or hateful are further anno- tated as racial or non-racial. We build French hate speech detection models based on the mul- tilingual BERT and CamemBERT and apply transfer learning by fine-tuning the HateXplain model. We compare different approaches to re- solve annotation ties and find that the detection model based on CamemBERT yields the best results in our experiments. 1 Introduction The rapid advancements of social media platforms like Facebook, Twitter, and YouTube during the last couple of years have enabled users to express and distribute their sentiments on events and ideas freely and conveniently. This eases the usage of hateful messages that can imply threats or harass- ment against minorities (Chiril et al., 2020). Since there are variations in defining hate speech globally, we took the following explanations as working def- initions in this research. Therefore, hate speech is defined as a public communication consisting of messages that may express threats, harassment, in- timidation, or disparagement of a person or a group on the basis of some characteristic such as race, color, ethnicity, gender, sexual orientation, nation- ality, religion, culture or other characteristic (Nock- leyby, 2000). Besides, offensive speech is also hurtful speech that is directed against another per- son. Compared with hate speech, offensive speech Figure 1: French language test example presented for performers has fewer legal implications since it does not attack people based on their group identity, rather it hurts individuals based on personal characteristics and makes them offended. More specifically, racism is a type of discrimination that makes up a large por- tion of hate speech and is usually directed against the perceived ethnicity, appearance, religion, or culture (Rzepnikowska, 2019). After the killing of George Floyd on May 25th, 2020, the number of racist comments on social me- dia platforms, especially on Twitter, has increased substantially (Carvalho et al., 2022). Social media platforms use mainly content moderation systems, which are human-machine collaborative systems to detect and handle hate speech as an automatic de- tection system in spite of the limitations that such systems have to control the problem (Horta Ribeiro et al., 2021). These days, the task of automatic hate speech detection in general and racial hate speech, in particular, has attracted …

Content truncated to 500 words