HateBERT: Retraining BERT for Abusive Language Detection in English

TL;DR

HateBERT is a specialized BERT model retrained on Reddit comments from offensive communities, significantly improving performance in abusive language detection tasks across multiple datasets.

Contribution

The paper introduces HateBERT, a new BERT variant trained on abusive Reddit comments, demonstrating enhanced detection of offensive language over standard BERT.

Findings

HateBERT outperforms general BERT in abusive language detection.

Retraining on community-specific data improves model performance.

Portability of HateBERT varies with dataset annotation compatibility.

Abstract

In this paper, we introduce HateBERT, a re-trained BERT model for abusive language detection in English. The model was trained on RAL-E, a large-scale dataset of Reddit comments in English from communities banned for being offensive, abusive, or hateful that we have collected and made available to the public. We present the results of a detailed comparison between a general pre-trained language model and the abuse-inclined version obtained by retraining with posts from the banned communities on three English datasets for offensive, abusive language and hate speech detection tasks. In all datasets, HateBERT outperforms the corresponding general BERT model. We also discuss a battery of experiments comparing the portability of the generic pre-trained language model and its corresponding abusive language-inclined counterpart across the datasets, indicating that portability is affected by compatibility of the annotated phenomena.