Hatevolution: What Static Benchmarks Don't Tell Us

TL;DR

This paper highlights the importance of time-sensitive benchmarks for evaluating hate speech detection models, revealing that static benchmarks may not accurately reflect models' robustness over evolving language.

Contribution

It empirically evaluates 20 language models on evolving hate speech data, demonstrating the limitations of static benchmarks and advocating for temporal evaluation methods.

Findings

Static benchmarks do not capture language evolution effects.

Models show decreased robustness over time in hate speech detection.

Time-sensitive benchmarks improve evaluation reliability.

Abstract

Language changes over time, including in the hate speech domain, which evolves quickly following social dynamics and cultural shifts. While NLP research has investigated the impact of language evolution on model training and has proposed several solutions for it, its impact on model benchmarking remains under-explored. Yet, hate speech benchmarks play a crucial role to ensure model safety. In this paper, we empirically evaluate the robustness of 20 language models across two evolving hate speech experiments, and we show the temporal misalignment between static and time-sensitive evaluations. Our findings call for time-sensitive linguistic benchmarks in order to correctly and reliably evaluate language models in the hate speech domain.