Table 2. Comparison of studies that looked at cross-dataset (“cross-domain”) generalisation, by datasets and models used.

Dataset types: H: hate speech, O: other offensive language, *: contains subtypes. Most studies carried out cross-dataset experiments, training and testing each model on all datasets. The exceptions are: Gröndahl et al. (2018) and Nejadgholi & Kiritchenko (2020) used different datasets for training and testing; Fortuna, Soler & Wanner (2020) compared datasets through class vector representations.

		Study
Dataset name	Type	Karan & Šnajder (2018)	Gröndahl et al. (2018)	Waseem (2016)	Wiegand, Ruppenhofer & Kleinbauer (2019)	Swamy, Jamatia & Gambäck (2019)	Pamungkas & Patti (2019), Pamungkas, Basile & Patti (2020)	Arango, Prez & Poblete (2020)	Fortuna, Soler & Wanner (2020)	Caselli et al. (2020)	Nejadgholi & Kiritchenko (2020)	Glavaš, Karan & Vulić (2020)	Fortuna, Soler-Company & Wanner (2021)
Waseem	H*	✓	✓	✓	✓	✓	✓	✓	✓		✓		✓
Davidson	H,O		✓	✓		✓			✓				✓
Founta	H,O				✓	✓					✓		✓
HatEval	H*						✓	✓	✓	✓			✓
Kaggle	H,O*	✓			✓				✓				✓
Gao	H	✓										✓
AMI	H*						✓		✓				✓
Warner	H				✓
Zhang	H		✓
Stromfront	H												✓
TRAC	O	✓			✓				✓			✓	✓
Wulczyn	O	✓	✓								✓	✓
OLID	O					✓	✓			✓			✓
AbuseEval	O*									✓
Kolhatkar	O	✓
Razavi	O				✓
Golbeck	O						✓
model		SVM	LR, MLP, LSTM, CNN-GRU	MLP	FastText	BERT	SVM, LSTM, GRU, BERT	LSTM, GBDT	N/A	BERT	BERT	BERT, RoBERTa	BERT, ALBERT, fastText, SVM