A Computational Approach to Discovering Emergent Hate Speech on Social Media

The Problem

Automated hate-speech detection systems face a structural limit: they can only flag what they already know to look for. Systems trained on labeled datasets of known slurs will tend to miss whatever new vocabulary hate communities develop next — and the authors note that evading detection is itself a driver of linguistic creativity. When a moderation system learns to flag a term, that term tends to be replaced. New expressions emerge in community spaces, acquire meaning, and spread, sometimes long before any researcher or moderator adds them to a lexicon.

The authors also note that much contemporary hate speech is implicit rather than explicit. A study of 304 comments on Zeit Online and The Guardian that equated Israel's actions with Nazism found only one expressed in explicit terms. The remaining 303 used coded language, irony, analogies, and visual cues that keyword-based detection cannot catch — for example: "$oros" (the dollar sign replacing the S invokes an antisemitic stereotype linking Jewish people to greed); coded acronyms like "WWG1WGA" (a QAnon marker for "Where We Go One We Go All"); or "Operation Google," a 2016 effort to substitute innocuous-looking words for ethnic slurs (Google for Black, Yahoo for Mexican, Skittle for Muslim) so that hate could spread while evading automated filters.

Wu et al. propose to address this not by building a better classifier, but by building a discovery tool that surfaces novel vocabulary for researchers and practitioners to evaluate. The positioning aligns with Parker and Ruths (2023), who argue that computer science work in this area should be shaped by stakeholder needs.

Approach and Findings

The Hate Mitigation App (HMApp) is built around a seed-term pipeline. The user picks a hate category — antisemitism, anti-Black hatred, Islamophobia, anti-Asian hatred, or anti-LGBTQ+ hatred — and enters a small set of known expressions as seeds. The system then pulls posts from its corpus (currently the 4chan /pol/ board, updated daily) that contain those seeds and runs them through a two-stage filter to identify new candidate terms.

The first stage finds terms that appear unusually often in seed-containing posts compared to the broader corpus. The second stage uses BERT — a neural language model that represents words in a numerical space where similar meanings sit close together — to identify which of those terms are semantically close to known hate vocabulary. The authors illustrate the value of this approach with the term "ZOG," which is both an antisemitic acronym (Zionist Occupation Government) and the name of a children's-book dragon. BERT separates the two usages in its numerical space, so the offensive meaning can be surfaced without flagging the harmless one. Candidate terms are then compared against a curated database of anchor terms — verified hate expressions for each category.

Building these anchor lists is itself a meaningful methodological step. For each hate category, the authors drew on authoritative sources (the AJC Translate Hate Glossary; the Committee of 100 anti-Asian glossary; the Chew Glossary for LGBTQ+ hatred; Wikipedia for anti-Black hatred). They cross-checked candidate terms against Hatebase and the Racial Slur Database, keeping only those rated "Highly" or "Extremely Offensive." They then applied frequency filtering through Pyrra Technologies, focusing only on posts already classified as offensive, violent, or hateful.

The result is 22 anchor terms for anti-Black hatred, 20 for antisemitism, 16 for anti-Asian hatred, 24 for anti-LGBTQ+ hatred, and just 6 for Islamophobia. The authors note that the small Islamophobia list reflects thin source documentation, and that this asymmetry directly shapes the category-level results.

A representative output is the term "psyop" (short for "psychological operation"), which HMApp surfaces in antisemitic 4chan posts where it implies that Jewish people are covertly manipulating public opinion. The app also surfaces examples in which hate categories overlap in practice: one retrieved post contains derogatory terms targeting seven distinct groups in a single sentence.

Evaluation uses two complementary metrics. The hard score measures how many returned terms appear in established authoritative lexicons. The soft score measures how many can be confirmed as hate-associated through web searches and queries to general-purpose AI tools.

• Anti-Black hatred has the strongest hard score (0.44) — reflecting comparatively thorough existing documentation.
• Antisemitism shows a low hard score (0.16) alongside a moderate soft score (0.50).
• Anti-LGBTQ+ also performs strongly (hard 0.36, soft 0.63).
• Islamophobia (hard 0.16, soft 0.25) and anti-Asian hatred (hard 0.11, soft 0.49) produce the lowest hard scores.

The authors interpret the gap between hard and soft scores — particularly for antisemitism — as suggesting that some terminology circulates in online discourse before being incorporated into authoritative lexicons. This has direct implications for monitoring practice. A qualitative expert review also identified terms — "rootless" in anti-Black discourse, and "takfir" and "mushaf" in Islamophobic discourse — that neither metric captured but that carry clear context-dependent derogatory meaning.

The authors compare HMApp directly against general-purpose AI tools. When prompted for emerging antisemitic terminology, both Perplexity AI and Google Gemini returned lists of well-documented dog whistles and historical tropes; neither surfaced terms such as "psyop"  that HMApp identified in the live corpus. Unlike general-purpose AI systems, HMApp derives candidate terms directly from the corpus under analysis. Gemini did retrieve some Israel- and Zionism-related antisemitic terms that HMApp missed — a gap the authors attribute to the 4chan /pol/ user base's tendency toward traditional rather than anti-Zionist antisemitism, and a reminder that the corpus shapes the findings.

Implications

The authors emphasize that HMApp's design preserves what they call research agency: the analyst defines the conceptual frame and hate category, while the system surfaces candidates for human review. They are transparent that the app also returns some non-hateful terms as candidates, which is part of why they describe it as an analyst-assistant rather than a standalone classifier. The framing aligns with Kiritchenko et al.'s call for ethical considerations to be integrated across the detection pipeline, and with Parker and Ruths's stakeholder-centered argument.

For researchers and civil society organizations, the variation in hard and soft scores across categories is itself a finding: the authors take it to indicate where hate-speech documentation may lag online discourse, and where new lexicon development could be most useful. They identify additional downstream users beyond researchers and platform moderators, including law enforcement and crime prevention practitioners, mental-health services, and social and political scientists studying code-switching and dog whistles.

Planned extensions include data sources beyond 4chan /pol/ (complicated by platform API closures), additional hate categories (misogyny, anti-Hispanic/Latino hatred), an internal classification layer, an explainability module, and staged user studies with researchers and practitioners. The authors also note that the semantic-proximity approach could be adapted to discover new vocabulary in other domains. The authors describe HMApp in its current form as the first iteration of the tool, but argue that the underlying approach — using semantic proximity to known hate expressions to surface novel vocabulary from live social media data — contributes meaningfully to monitoring infrastructure for digital hate studies.