Understanding Abuse: A Typology of Abusive Language Detection Subtasks

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/W17-3012.pdf

Summary

Research on detecting harmful online content is split into subtasks such as hate speech, cyberbullying, trolling, offensive language and personal attacks. Their labels and guidelines overlap and sometimes contradict each other: a message counted as hate speech in one dataset may be marked merely offensive in another. Waseem, Davidson, Warmsley and Weber offer a unifying two-dimensional typology rather than a new dataset or model. The first dimension is the target: abuse is either directed at a specific person or entity, or aimed at a generalized group such as an ethnicity or sexual orientation. The second is form: abuse is either explicit (for example, slurs) or implicit (relying on sarcasm, euphemisms or code words). Each subtask occupies one or more cells, and examples from earlier corpora illustrate all four. Drawing on prior literature, the authors derive recommendations for annotation and for choosing features in each cell. The framework helps researchers write consistent guidelines, pick suitable features, and share insights across subtasks that have developed in parallel.

Key points

• Typology and examples. The typology crosses directed vs. generalized with explicit vs. implicit. Table 1 fills each cell with examples from prior work, including Van Hee et al., Dinakar et al., Davidson et al. (2017), Nobata et al. (2016), Burnap and Williams (2015) and Magu et al. (2017). The examples show euphemisms like "skypes" and "Google" standing in for slurs.
• Annotator agreement.
  • Agreement is high for cyberbullying, a directed phenomenon: 93% (Dadvar et al., 2013).
  • It is lower for hate speech, harassment and offensive language: κ ≈ 0.60–0.80.
  • The authors link this gap to whether the target is easy to identify.
• Annotation bias and expertise.
  • Davidson et al. (2017) found annotators label racism as hate speech more often than sexism.
  • Expert annotators produce more accurate labels than non-experts (Waseem, 2016a).
  • Detailed guidelines did not make non-expert labels more reliable (Ross et al., 2016).
• Suggested features by cell:
  • Directed abuse: mentions, named entities, coreference, POS sequences and typed dependencies.
  • Explicit abuse: lexicons and sentiment.
  • Implicit abuse: character n-grams, word embeddings, semi-supervised keyword expansion (King et al., 2017) and extra-textual user features.
• Paper type. This is a position paper with no new corpus or experiments. It was published at ACL 2017's inaugural workshop devoted to online abusive language, appearing on pp. 78–84 of that event's proceedings.