HateBERT: Retraining BERT for Abusive Language Detection in English

Abstract

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 curated and made available to the public. We present the results of a detailed comparison between a general pre-trained language model and the retrained version 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 fine-tuned models across the datasets, suggesting that portability is affected by compatibility of the annotated phenomena.

Fight Fire with Fire: Fine-tuning Hate Detectors using Large Samples of Generated Hate Speech

“…Our results are favorable, ranging from 0.68-0.80 in macro-F1. Our results also outperform a variant of BERT that has been pretrained using hateful texts (Caselli et al, 2020): we achieved 0.61 in hate-F1 using the generic BERT finetuned on the original SE dataset vs. their 0.65, and improved this result to 0.77 with data augmentation. Compared with the CNN-GRU results (Zhang et al, 2018) reported in Wullach et al (2021), we obtain better results both prior and post augmentation in most cases.…”

ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Abuse Detection in Conversational AI

“…We show that training on labelled in-domain data leads to better performance than similarly sized out-of-domain datasets, confirming the differences between the domains and highlighting the need for conversational data. While performance using general domain pretrained models leaves room for improvement, in future work, we hope to experiment with different initialisation settings, using models trained on data and tasks more similar to those of ConvAbuse, such as HateBERT (Caselli et al, 2021) or HurtBERT (Koufakou et al, 2020).…”

Cyberbullying Classifiers are Sensitive to Model-Agnostic Perturbations

“…For all BERT models, we use the pre-trained bertlarge-cased, for BART we use bart-large. 8 We also report experiments where we use a fine-tuned toxicity version of pre-trained BERT (Caselli et al, 2021, GroNLP/hateBERT) for f cnd in Dropout BERT (here referred to as Hate BERT, and when using a finetuned substitute classifier: Hate BERT+). The idea here is similar to that of using Dropout BERT+; domainspecific vocabularies will likely result in better and more varied substitutions.…”

“…Our work combines multiple sizeable-to the extent that they respectively produced several surveys (Fortuna and Nunes, 2018;Gunasekara and Nejadgholi, 2018;Mishra et al, 2019;Banko et al, 2020;Madukwe et al, 2020;Muneer and Fati, 2020;Salawu et al, 2020;Jahan and Oussalah, 2021;Mladenovic et al, 2021) Recent cyberbullying work (Reynolds et al, 2011;Xu et al, 2012;Nitta et al, 2013;Bretschneider et al, 2014;Dadvar et al, 2014;Van Hee et al, 2015, e.g., are seminal work) has primarily focused on deploying Transformer-based models (Vaswani et al, 2017); by and large fine-tuning (Swamy et al, 2019;Paul and Saha, 2020;Gencoglu, 2021, e.g. ), or re-training (Caselli et al, 2020) BERT. It is worth noting that Elsafoury et al (2021a;Elsafoury et al (2021b) show that although fine-tuning BERT achieves state-of-theart performance in classification, its attention scores do not correlate with cyberbullying features, and they expect generalization of such models to be subpar.…”