HatePrototypes: Interpretable and Transferable Representations for Implicit and Explicit Hate Speech Detection
This addresses the challenge of efficient and transferable hate speech detection for content moderation, particularly for implicit hate often overlooked in benchmarks, though it is incremental in improving existing methods.
The paper tackled the problem of detecting implicit and explicit hate speech by introducing HatePrototypes, class-level vector representations derived from language models, which enable cross-task transfer between these hate types with as few as 50 examples per class and support parameter-free early exiting.
Optimization of offensive content moderation models for different types of hateful messages is typically achieved through continued pre-training or fine-tuning on new hate speech benchmarks. However, existing benchmarks mainly address explicit hate toward protected groups and often overlook implicit or indirect hate, such as demeaning comparisons, calls for exclusion or violence, and subtle discriminatory language that still causes harm. While explicit hate can often be captured through surface features, implicit hate requires deeper, full-model semantic processing. In this work, we question the need for repeated fine-tuning and analyze the role of HatePrototypes, class-level vector representations derived from language models optimized for hate speech detection and safety moderation. We find that these prototypes, built from as few as 50 examples per class, enable cross-task transfer between explicit and implicit hate, with interchangeable prototypes across benchmarks. Moreover, we show that parameter-free early exiting with prototypes is effective for both hate types. We release the code, prototype resources, and evaluation scripts to support future research on efficient and transferable hate speech detection.