Abusive Language on Social Media Through the Legal Looking Glass

Thales Bertaglia, Andreea Grigoriu, Michel Dumontier, Gijs van DijckACL 2021

TL;DR

This paper introduces ALYT, a new dataset of YouTube comments labeled for abusive language using a novel annotation framework inspired by legal definitions. This approach aims to provide a more nuanced and legally grounded understanding of online abuse, moving beyond simple binary classifications.

What Problem Does This Paper Address?

The proliferation of abusive language on social media poses a significant threat, contributing to mental and emotional distress for users. While automatic detection tools are crucial for combating this issue, their effectiveness hinges on high-quality, well-defined datasets. However, the field currently lacks standardized approaches for creating these datasets, leading to inconsistent definitions of abusive language, annotation guidelines, and reporting practices. This inconsistency hinders progress and leads to disagreements within the research community. The paper addresses this gap by proposing a legally-inspired annotation framework and a corresponding dataset that aims for greater consistency and a more nuanced understanding of abusive language.

Methods

The core of this paper lies in its novel annotation framework and dataset creation process. Instead of traditional binary labels (abusive/not abusive), the authors employ a 7-point Likert scale, inspired by legal concepts. This scale captures not only the presence of abuse but also its intensity and the annotator's confidence. The framework itself is derived from an analysis of European legal provisions and case law related to offensive language, hate speech, defamation, and incitement to hatred. Three key questions guide annotators:

  1. Does it incite hatred, violence, or discrimination against a minority/group?
  2. Does it cause distress or anxiety to the target?
  3. Does it cause degradation or belittlement in public opinion?

The dataset, ALYT, was created by collecting YouTube comments from videos on controversial topics (Gender Identity, Veganism, Workplace Diversity). Crucially, the authors avoided artificially balancing the data for abusive content, opting for random sampling to reflect real-world distributions. A team of law students was trained extensively on the annotation framework and manual, with iterative discussions to refine the process and improve inter-annotator agreement.

Datasets

The paper introduces the ALYT dataset, comprising 20,215 YouTube comments in English. These comments were collected from videos on three controversial topics: Gender Identity (GI), Veganism (VG), and Workplace Diversity (WD). The dataset is notable for its random sampling approach, aiming to reflect the natural distribution of abusive content on the platform, rather than artificially inflating it.

The authors also compare ALYT to existing datasets:

  • HSOL (Davidson et al., 2017): Labeled as "hateful" on Twitter.
  • HSHP (Waseem and Hovy, 2016): Labeled as "sexist or racist" on Twitter.
  • OLID (Zampieri et al., 2019): Labeled as "offensive and targeted" on Twitter.

Key Insights and Ideas

  • Legal Frameworks Offer Nuance: Adapting legal definitions provides a more robust and consistent foundation for identifying abusive language than ad-hoc or purely linguistic approaches. The legal perspective inherently considers intent, harm, and protected attributes, which are crucial for understanding abuse.
  • Likert Scale Captures Complexity: Moving beyond binary labels to a 7-point Likert scale allows for a more granular representation of abusive language, capturing varying degrees of severity and annotator confidence. This is particularly valuable for understanding subtle forms of abuse.
  • Real-World Distribution Matters: Random sampling, while potentially yielding less abusive content than targeted collection, provides a more realistic picture of online discourse. This is essential for building models that generalize well to actual platform data.
  • Training is Paramount: The extensive training phase with law students, including iterative feedback and discussion, was critical for achieving satisfactory inter-annotator agreement and ensuring data quality. This highlights the importance of human-in-the-loop processes in dataset creation.
  • Abuse is Diverse and Contextual: Lexical analysis reveals that abusive language isn't solely characterized by slurs or hateful keywords. It often involves targeted attacks on individuals, subtle forms of sexism, and context-dependent expressions that require deeper understanding.

Key Results

  • The ALYT dataset contains 11.42% abusive comments, a figure considered representative due to the random sampling approach.
  • Inter-annotator agreement, measured by Krippendorff's alpha, reached 73.8% for the overall dataset, indicating substantial consistency among annotators.
  • Classification experiments showed that a BERT-based model outperformed logistic regression by 8.6 points in macro F1, demonstrating BERT's ability to capture higher-level features and context, crucial for predicting abusive comments.
  • Logistic regression, while achieving higher precision, had significantly lower recall, suggesting it relies on surface-level lexical clues and struggles with context.

Limitations and Future Work

  • Scale vs. Binary Labels: The paper acknowledges that the implications of using a 7-point scale versus binary labels for downstream tasks were not explored. This comparison could be a relevant direction for future research.
  • Dataset Size and Balancing: While random sampling provides realism, it might lead to a lower prevalence of abusive content than desired for some applications. Future work could explore strategies for balancing datasets while maintaining realism.
  • Cross-Platform Generalization: The current dataset is specific to YouTube comments. Future work could investigate how well the annotation framework and models generalize to other social media platforms with different discourse styles.
  • Exploring Legal Nuances Further: The paper's legal inspiration is a strong starting point, but further exploration of specific legal doctrines and their operationalization for NLP tasks could yield even more refined annotation frameworks.

Summary updated

BibTeX
@inproceedings{bertaglia2021abusive,
  author = {Thales Bertaglia and Andreea Grigoriu and Michel Dumontier and Gijs van Dijck},
  title = {{Abusive Language on Social Media Through the Legal Looking Glass}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure