Implicitly abusive language refers to expressions that convey derogatory, offensive or harmful meaning without relying on overtly abusive words. Instead, abuse may be communicated indirectly through stereotypes, pragmatic inference, contextual assumptions or subtle forms of othering.

This talk examines why such language remains difficult to model in abusive language detection, focusing on challenges related to data, dataset biases and classification. It discusses why existing datasets often contain detrimental biases, such as identity-group bias, and argues for the need for more focused datasets targeting specific subproblems in implicit abuse detection.

The talk also considers how crowdsourcing and large language models can contribute to the creation and analysis of such datasets. Finally, it addresses the difficulty of modelling the perceived abusiveness of non-negative stereotypes, whose interpretation may depend on semantic variation, generalization, perspective framing and the identity group being described.

Michael Wiegand obtained his PhD from Saarland University in 2011. Until 2018, he was a postdoctoral researcher at the Department of Spoken Language Systems at Saarland University . In 2019, he served as a research group leader at the Leibniz ScienceCampus Empirical Linguistics and Computational Language Modeling (Leibniz Institute for the German Language, Mannheim/Heidelberg University). From 2020 to 2024, he held a temporary professorship in Computational Linguistics at the Digital Age Research Center (D!ARC) at the University of Klagenfurt.

Since 2024, he has been a senior scientist in Digital Philology at the Faculty of Philological and Cultural Studies at the University of Vienna.

Learn more about Michael Wiegand:

• Website