Multi-task learning for hate speech detection

Lee, Sooyong
Journal Title
Journal ISSN
Volume Title

Amidst the proliferation of social media and the accompanying explosion of information and content generation, the amount of online hate speech has grown rapidly. In efforts to build and train hate speech detection models to counter this, datasets have been annotated for hate speech. However, there exists incompatibility of categories of hate speech across different datasets, the lack of clear and ubiquitous definitions for hate speech, and generalization issues of models which depend highly on training data. To address this, we propose framing hate speech detection as multi-task learning (MTL) which provides a natural and principled way for a model to specialize on dataset-specific hate speech detection tasks while leveraging shared notions of hate speech across datasets to acquire more general notions of hate speech.