Source-linked AI summary
Understanding Abuse: A Typology of Abusive Language Detection Subtasks
Zeerak Waseem, Thomas Davidson, Dana Warmsley, Ingmar Weber
TL;DR
Abusive-language subtasks overlap while differing in their targets and explicitness, creating inconsistent labels and annotation guidelines. The paper proposes a two-axis typology and derives implications for annotation and feature construction. It concludes that researchers should align research design, annotators, and features with the abuse phenomenon under study, while recognizing that implicit abuse is harder to annotate and model.
Problem
Abusive-language subtasks address partially overlapping phenomena, but limited examination of their relationships and inconsistent labels have produced contradictory annotation guidelines.
Method
The paper synthesizes work on hate speech, cyberbullying, and online abuse into a typology based on target specificity and abusive-language explicitness.
Results
The typology identifies analytical distinctions among abusive-language subtasks and connects them to implications for annotation and feature construction.
Takeaways & Limitations
Researchers should select annotation guidelines, annotators, and features according to the abuse phenomenon they aim to measure.
Takeaways & Limitations
Implicit abuse is considerably more difficult to annotate and is expected to be more difficult to detect and model than explicit abuse.
Abstract
from arXiv · showhide
As the body of research on abusive language detection and analysis grows, there is a need for critical consideration of the relationships between different subtasks that have been grouped under this label. Based on work on hate speech, cyberbullying, and online abuse we propose a typology that captures central similarities and differences between subtasks and we discuss its implications for data annotation and feature construction. We emphasize the practical actions that can be taken by researchers to best approach their abusive language detection subtask of interest.
1 Introduction
Abusive-language subtasks overlap but address partially distinct phenomena, and inconsistent labels have produced contradictory annotation guidelines. The paper proposes a typology to synthesize these subtasks and guide research strategies.
- Research on abusive language, hate speech, cyberbullying, and trolling has grown, while their relationships remain insufficiently examined.
- Prior studies use overlapping labels differently, classifying similar remarks as insults, hate speech, derogatory language, or offensive language.
- This lack of consensus produces contradictory annotation guidelines for similar messages.
- The proposed typology synthesizes abusive-language subtasks around whether language targets a specific individual or entity or a generalized group.
2. Is the abusive content explicit or implicit?
The typology distinguishes abusive-language subtasks partly by whether abuse is directed at a specific target or a generalized group.
- Researchers distinguish abuse directed towards a specific individual or entity from abuse directed towards a generalized group.
2 A typology of abusive language
The typology organizes abusive language along two dimensions: target specificity and explicitness. Abuse may address a named or otherwise specific target, a generalized group, and may be explicit or implicit.
- The typology considers whether abuse targets a specific target and the degree to which it is explicit.
- Directed abuse targets a named, tagged, or otherwise specifically referenced individual or entity, while generalized abuse targets categories such as racial or sexual groups.
- Cyberbullying and trolling are presented as directed abuse aimed at individuals and online communities, respectively.
- Explicit abuse is unambiguous in its potential to be abusive, whereas implicit abuse obscures abusive meaning through ambiguity, sarcasm, lack of profanity, or hateful terms.
- Implicit abuse may still carry extremely abusive connotations despite not immediately denoting abuse.
3 Implications for future research
The typology links abusive-language subtypes to different annotation and modeling strategies. It emphasizes matching annotation, annotator expertise, and features to the abuse phenomenon being measured.
- The typology is intended to clarify how researchers can understand, measure, and model each subtype of abuse.
- 3.1 Implications for annotation: Cyberbullying has relatively consistent annotation guidelines and high inter-annotator agreement of 93%.
- 3.1 Implications for annotation: Annotation is more difficult for implicit abuse because its connotations may require domain-specific knowledge.
- 3.1 Implications for annotation: Annotation strategy should follow the target abuse type, with nuanced distinctions for subtasks occupying multiple typology cells.
- 3.2 Implications for modeling: Directed-abuse detection benefits from target-identifying features such as mentions, proper nouns, named entities, and coreference resolution.
- 3.2 Implications for modeling: Dictionary-based approaches may help identify explicit abuse, but abusive-word presence should not be the sole criterion because terms have varied uses.
- 3.2 Implications for modeling: Implicit-abuse detection may require semi-supervised lexicon expansion and character n-grams to capture abusive-word variation.
4 Discussion
The typology has implications for cross-subtask learning, fine-grained modeling, annotation design, feature selection, and recognizing differences in abuse effects and detectability.
- Researchers should share advances across subtasks because hate speech and cyberbullying both substantially involve identifying targeted abuse.
- Fine-grained distinctions along the typology’s axes may produce more focused systems that better identify particular types of abuse.
- Annotation guidelines and annotator selection should reflect the abuse phenomenon, and published studies should transparently describe their annotation processes.
- Feature sets should be matched to individual subtasks, with future work clarifying when different feature types are appropriate.
- Explicit and implicit, targeted and generalized abuse differ in effects, reporting, and detectability, so stakeholders may prioritize them differently.
5 Conclusion
The paper presents a typology that synthesizes abusive-language detection subtasks and highlights overlooked analytical distinctions. It recommends using these distinctions to guide research design, annotation, feature creation, modeling, and empirical comparison.
- The typology synthesizes different abusive-language detection subtasks to clarify key aspects of abuse detection.
- Researchers should focus on the phenomena they want to measure and choose appropriate research designs rather than seeking perfectly bounded subtask definitions.
- The typology is intended to inform data collection, annotation, feature creation, and modeling while encouraging transparent reporting and empirical examination of subtask similarities and differences.