Tâches Auxiliaires Multilingues pour le Transfert de Modèles de Détection de Discours Haineux

Detecting hateful content is a challenging task, as it requires extensive cultural and contextual knowledge from the model; the necessary knowledge varies depending on the speaker’s language or the target of the content. However, annotated data for specific domains and languages are often inexistant or limited. In that case, annotated data in other languages can be exploited ; but the crosslingual transfer is often difficult due to these cultural and contextual variations. In this paper, we highlight this limitation for several domains and languages and show the positive impact of learning multilingual auxiliary tasks - sentiment analysis, recognition, and tasks based on morpho-syntactic information - on the cross-lingual zero-shot transfer of hate speech detection models in order to bridge this cultural gap.