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.
HateBERT: Retraining BERT for Abusive Language Detection in English
HateBERT, a retrained BERT model for abusive language detection, outperforms the general BERT model across multiple English datasets and shows varying portability depending on dataset compatibility.
- Year
- 2020
- Venue
- ACL (WOAH) 2021 8
- Authors
- 4
- Hosting
- Abstract onlyARXIV-DEFAULT
Cite
Notes
Only stored in your browser.
Attribution
- Abstract & full text
- arxiv.org/abs/2010.12472v2ARXIV-DEFAULT
- TL;DR
- Semantic Scholar