aclanthology.org web signal

Röttger EMNLP paper: minimal data expands hate detection

By Published Updated

TL;DR

  • Röttger and co-authors find a small amount of target-language fine-tuning data is enough to achieve strong hate speech detection performance.
  • The benefits of adding more target-language annotation decrease exponentially, the EMNLP 2022 paper reports across five non-English languages.
  • Initial fine-tuning on English data can partially substitute for target-language labels and improve model generalisability, per the study.

Most hate speech datasets are English-only. Annotating hateful content in other languages is, as the authors put it, "expensive, time-consuming and potentially harmful to annotators."

In a 2022 EMNLP paper, Paul Röttger, Debora Nozza, Federico Bianchi and Dirk Hovy run mono- and multilingual models across five non-English languages and land on three findings. First: "a small amount of target-language fine-tuning data is needed to achieve strong performance." Second: "the benefits of using more such data decrease exponentially." Third: "initial fine-tuning on readily-available English data can partially substitute target-language data and improve model generalisability."

The abstract does not name which five languages were tested, nor the exact dataset sizes or specific model architectures behind those curves. The authors frame the results as "actionable recommendations for hate speech detection in low-resource language settings."

Shared on Bluesky by 2 AI experts

Originally reported by aclanthology.org

Read the original article →

Original headline: Data-Efficient Strategies for Expanding Hate Speech Detection into Under-Resourced Languages