📄 Online antisemitism across platforms

Summary

The report addresses the difficulty of finding antisemitic content in the growing volume of social media data, where conspiracy narratives, partly driven by COVID-19, are spreading again. The author built a curated lexicon of more than 2,000 English and German words and multiword patterns. Each entry carries a 0–4 toxicity rating from two annotators and any number of labels covering rhetorical categories. On top of the lexicon, a transparent system gives each message a 0–100 score from the entries it matches, including variant word forms. It also highlights the matched spans so that a human can see the basis for the score. To check the system, the author ran it on five corpora where antisemitism was expected to be rare or common. The resulting ranking matched those expectations, from a mainstream 2015 baseline up to the unmoderated forum 8chan. The work matters because explainable, quickly updatable tools are needed for monitoring hate across platforms and for high-stakes moderation decisions.

Key points

• The lexicon combines several kinds of entries:

  • Nazi-era vocabulary such as Lügenpresse and Umvolkung.
  • Dehumanizing adjective plus target pairs, and violent verb plus target pairs.
  • Alt-right coinages such as holocough and Jewbacca.
  • Coded numerals such as 14/88 and HH.
  • Conspiracy terms such as blood libel and Vril.

• Entries can carry 10+ labels, for example RIDICULE, DEHUMANIZATION, VIOLENCE and CONSPIRACY. In the annotation tool these appear as four-letter codes such as FOOL, SCUM and KILL.

• Measured average toxicity and share of antisemitic posts per corpus:

  Corpus	Size	Avg. toxicity	Antisemitic posts
  Twitter/Facebook, 2015	100K	~2/100	<0.1%
  QAnon on Twitter, late 2020	100K	10/100	~3%
  QAnon on Twitter, early 2021	100K	5/100	~0.5%
  Gab, 2020	10K	20/100	8.5%
  8chan, 2020	100K	30/100	15%

• The author offers two explanations for the post-Capitol-riot drop:

  • Twitter suspended Trump and about 70,000 QAnon accounts.
  • Remaining users masked their language or moved to Gab and Telegram.

• Additional algorithms anonymize messages for GDPR compliance. The system processes thousands of messages in minutes and can be extended through weak supervision.

Interesting ideas

• I1. The system is validated by checking whether scores order corpora the way prior knowledge predicts, rather than against a hand-labelled test set; this is cheap and works where gold labels are scarce.
• I2. Lexicon entries are compositional patterns (a dehumanizing adjective or a violent verb joined to a target group) with graded scores and multiple labels, and the explanation is simply the matched span.
• I3. Annotators work with short mnemonic four-letter label codes, which speeds tagging and lowers cognitive load during curation.
• I4. A pseudo-random baseline sample was obtained from platform APIs by querying for punctuation marks or common vowels, giving a neutral reference corpus without topic bias.
• I5. The lexicon tracks coded and evolving vocabulary (number codes, blended neologisms) and is meant to be updated continually, because users mask language once they are watched.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Evaluation with negative controls could likewise rank corpora expected to differ (e.g., children's fables vs. news) and flag any component whose ordering contradicts expectations.
• I2 (frames): The same idea could be lifted from surface n-grams to frames, detecting a violence-class VerbNet predicate whose patient is a social group and citing that frame as the explanation for an aggression or hostility judgement.
• I3 (learn): Compact mnemonic tags could simplify the label sets given to human checkers or LLM teacher tiers when they review silver data.
• I4 (learn): A similar trick could produce unbiased background corpora for calibrating parser confidence or measuring domain drift.
• I5 (tokens): The tokenizer and morphology could recognize numeric codes and portmanteau blends as units, linking them to known lemmas instead of splitting them into opaque fragments.