Decoding coded language in video-games communications

Hate speech manifests in various forms, ranging from overt slurs to more subtle, coded messages known as dogwhistles. These dogwhistles allow speakers to convey apparently innocuous messages that include an additional controversial layer only recognizable to a specific subgroup. This duality enables plausible deniability for the speaker while reinforcing group identity among those who understand the hidden message. Computational methods have emerged as a promising approach to detect these subtle cues. By leveraging deep contextual representations and network-based analysis, researchers can uncover patterns and associations within language that hint at underlying negative connotations. For instance, certain words may not inherently carry hateful meanings but are associated with negativity due to their usage in specific contexts. The integration of transformer-based architectures with social network signals further enhances the detection of covert hate speech, enabling systems to distinguish it from merely offensive language. These multifaceted approaches are essential for developing robust moderation systems capable of addressing toxicity. Overall, detecting hate speech requires a comprehensive understanding of linguistic nuances and social contexts, combined with advanced computational techniques like deep learning and network analysis. As such, our integrated approach will effectively tackle the challenge of identifying both overt and subtle forms of hate speech.

Faculty Supervisor:

Grégoire Winterstein

Student:

Partner:

Ubisoft Divertissement

Discipline:

Sociology

Sector:

Information and cultural industries

University:

Université du Québec à Montréal

Program: