Topic Refinement in Multi-level Hate Speech Detection

Hate speech detection is quite a hot topic in NLP and various annotated datasets have been proposed, most of them using binary generic (hateful vs. non-hateful) or finer-grained specific (sexism/racism/etc.) annotations, to account for particular manifestations of hate. We explore in this paper how to transfer knowledge across both different manifestations, and different granularity or levels of hate speech annotations from existing datasets, relying for the first time on a multilevel learning approach which we can use to refine generically labelled instances with specific hate speech labels. We experiment with an easily extensible Text-to-Text approach, based on the T5 architecture, as well as a combination of transfer and multitask learning. Our results are encouraging and constitute a first step towards automatic annotation of hate speech datasets, for which only some or no fine-grained annotations are available.