Connecting peace studies and natural language processing to rethink hate speech detection as hostile narrative analysis.

Abstract

In response to limitations with current computational methods of hate speech detection, this research connects Peace Research and Natural Language Processing (NLP) to propose the idea of hostile narrative analysis. The corpus guiding this research contrasts Hitler's Mein Kampf and texts from the ‘War on Terror’ era with speeches from Martin Luther King, who advocated for non-violent change. Experiments using this corpus find the current computational methods of hate speech detection are unconnected to a defining theory, which questions their explanatory rigour. Hate speech itself is a polysemous term, and using the computational method of text classification skews an orator’s intended meaning. The response to this finding with hostile narrative analysis draws upon Galtung’s theory of cultural violence from Peace Research to detect the ‘Self-other gradient’. This gradient refers to processes of violence legitimisation by elevating the Self while deflating or debasing the value of the Other. As a broad hypothesis, the steeper the gradient between the Self and the other, the more legitimate violence becomes. The computational methods for detecting the Self-Other gradient then draw upon pattern-based methods in NLP. As a general observation, problems with current computational methods arise from a technical first approach before applying theory; this paper begins with cultural violence theory to guide technological development. This paper seeks to contribute to the field of Web Science and, to the best of my knowledge, constitutes the first attempt to connect cultural violence with NLP to analyse hostile narratives.

More information

Identifiers

ORCID for Stephen Paul Anning: orcid.org/0000-0003-4911-7907

ORCID for Craig Webber: orcid.org/0000-0003-3900-7579

Catalogue record

Export record