Table 1. English datasets used in cross-dataset generalisation studies.

Positive labels are listed with their original wording. Expert annotation type include authors and experts in social science and related fields. ?: Type of annotations not available in original paper, the found descriptions are thus included. Note that only datasets used in generalisation studies are listed—for comprehensive lists of hate speech datasets, see Vidgen & Derczynski (2020) and Poletto et al. (2020).

Dataset name	Publication	Source	Positive labels	Annotator type
Waseem	Waseem & Hovy (2016) Waseem (2016)	Twitter	Racism Sexism	Expert; Expert and crowdsourcing
Davidson	Davidson et al. (2017)	Twitter	Hate speech Offensive	Crowdsourcing
Founta	Founta et al. (2018)	Twitter	Hate speech Offensive	Crowdsourcing
HatEval	Basile et al. (2019)	Twitter	Hateful	Expert and crowdsourcing
Kaggle	Jigsaw (2018)	Wikipedia	Toxic Severe toxic Obscene Threat Insult Identity hate	Crowdsourcing
Gao	Gao & Huang (2017)	Fox News	Hateful	? (Native speakers)
AMI	Fersini, Nozza & Rosso (2018) Fersini, Rosso & Anzovino (2018)	Twitter	Misogynous	Expert
Warner	Warner & Hirschberg (2012)	Yahoo! American Jewish Congress	Anti-semitic Anti-black Anti-asian Anti-woman Anti-muslim Anti-immigrant Other-hate	? (Volunteer)
Zhang	Zhang, Robinson & Tepper (2018)	Twitter	Hate	Expert
Stromfront	De Gibert et al. (2018)	Stormfront	Hate	Expert
Kumar	Kumar et al. (2018b)	Facebook, Twitter	Overtly aggressive Covertly aggressive	Expert
Wulczyn	Wulczyn, Thain & Dixon (2017)	Wikipedia	Attacking	Crowdsourcing
OLID (OffensEval)	Zampieri et al. (2019a)	Twitter	Offensive	Crowdsourcing
AbuseEval	Caselli et al. (2020)	Twitter	Explicit (abuse) Implicit (abuse)	Expert
Kolhatkar	Kolhatkar et al. (2019)	The Globe and Mail	Very toxic Toxic Mildly toxic	Crowdsourcing
Razavi	Razavi et al. (2010)	Natural Semantic Module Usenet	Flame	Expert
Golbeck	Golbeck et al. (2017)	Twitter	Harassing	Expert