📄 Detecting Hate Speech on the World Wide Web

Summary

Warner and Hirschberg address automatic detection of hate speech, which they define as abusive language aimed at a group trait such as ethnicity, religion, gender or sexual orientation. Their premise is that each target group draws on a small set of frequent stereotype words. Because those words also appear in harmless text, they treat detection as a kind of word sense disambiguation. They combined offensive Yahoo! group comments with web pages reached through URLs supplied by the American Jewish Congress. Paragraphs mentioning Judaism or Israel were extracted, filtered and labelled with seven hate categories, and an author then corrected the labels to produce a gold corpus. A pilot anti-semitism classifier was trained on features from Yarowsky-style templates. It reached 94% accuracy, 68% precision and 60% recall (F1 ≈ 0.64), which roughly equals the F-measure of the human annotators when they were scored against the gold labels. This early work offers a usable definition of hate speech, an annotation scheme and a first baseline for a task that had no public classifier.

Key points

• Data sources:
  • several thousand Yahoo! group comments that readers flagged and administrators deleted (about 31 words on average, with no surrounding context);
  • text from 452 URLs originally gathered by Attenberg to find pages unsuitable for advertisers.
• Corpus construction:
  • a regular expression on Jewish/Israel-related terms produced about 9,000 paragraphs;
  • paragraphs were dropped if they were one word long, over 64 words, lacked a full sentence, or had runs of more than two unicode characters.
• Label set: anti-semitic, anti-black, anti-asian, anti-woman, anti-muslim, anti-immigrant and other-hate, with multiple labels allowed per paragraph.
• Human ceiling: two annotators scored against the gold corpus reached 59% precision and 68% recall, which the authors treat as an upper bound for the classifier.
• Classifier results: the best model scored 94% accuracy, 68% precision, 60% recall and F1 = 0.6375. Adding bigram and trigram templates made performance worse, and recall stayed low.

Interesting ideas

• I1. Hate detection as sense disambiguation: The same token, such as "jew", can be neutral or hateful depending on context. Framing the task as disambiguating each use, rather than matching a blocklist, fits any moderation problem where the vocabulary is shared but the intent differs.
• I2. One model per stereotype: Each target group comes with its own themes. Anti-semitic text leans on money, banking and media, while anti-hispanic text leans on borders and identity papers. Building a separate model for each stereotype first, then combining them, could transfer to other multi-target abuse or propaganda detection.
• I3. Evasion as a signal: Deliberate misspellings are evidence of offensive intent, not just noise to normalise away. Examples include one-letter swaps ("nagger"), homophones ("joo") and characters padded with separators ("j@e@w@").
• I4. Clear rules for borderline cases: The guidelines state that naming a hate group, boasting about one's own group, or discussing a slur as a word is not hate speech. Gratuitously tagging someone with their ethnicity, however, is. Writing such rules down explicitly helps any annotation project stay consistent.
• I5. Annotators as a reference system: Individual annotators are scored against the gold labels exactly as the classifier is. This gives a realistic performance ceiling and puts model scores in context.