Leveraging Argument Structure to Predict Content Hatefulness

The key methodological bet here is that hateful rhetoric follows predictable logical scaffolding, meaning the relationship between a premise and its conclusion can itself be a signal of harm, independent of the specific words used. That framing separates this from conventional keyword or embedding-based classifiers, which struggle with coded language and plausible-deniability phrasing common in white supremacy forums.

This connects directly to the work we covered from early May on Directed Social Regard, which tackled a related problem from a different angle: mapping coexisting positive and negative attitudes toward specific targets within a single message. Both papers are responding to the same core failure mode in content moderation, which is that flat polarity scores miss the structural logic of how harmful rhetoric is actually constructed. Together they suggest a convergence toward richer, more compositional NLP representations for harm detection, moving away from token-level signals toward relational and argumentative ones.

The real test is whether argument-structure features generalize beyond the WSF-ARG+ white supremacy corpus to other hate speech domains, such as misogyny or antisemitism benchmarks. If the approach holds accuracy gains on a cross-domain evaluation, it becomes a credible candidate for integration into moderation pipelines; if it degrades sharply, the method may be encoding domain-specific rhetorical patterns rather than anything structurally general.