Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Operationalizing content moderation "accuracy" in the Digital Services Act

Johnny Tian-Zheng Wei, Frederike Zufall, Robin Jia

TL;DR

This paper addresses how to operationalize the EU's DSA reporting requirement for automated content moderation accuracy by reframing 'accuracy' as precision and recall rather than raw accuracy. It couples a legal-interpretive analysis with a statistical estimation framework, proposing stratified sampling to efficiently estimate recall and reduce annotation burden. Through a Reddit case study and a CivilComments simulation, it demonstrates that recall can be reported with unbiased estimates and practical data requirements, while highlighting several underspecification areas needing regulatory clarification. The work offers concrete methodological guidance for regulators and platforms to enable meaningful, rights-respecting transparency and informs future policy design for algorithmic accountability in online moderation.

Abstract

The Digital Services Act, recently adopted by the EU, requires social media platforms to report the "accuracy" of their automated content moderation systems. The colloquial term is vague, or open-textured -- the literal accuracy (number of correct predictions divided by the total) is not suitable for problems with large class imbalance, and the ground truth and dataset to measure accuracy against is unspecified. Without further specification, the regulatory requirement allows for deficient reporting. In this interdisciplinary work, we operationalize "accuracy" reporting by refining legal concepts and relating them to technical implementation. We start by elucidating the legislative purpose of the Act to legally justify an interpretation of "accuracy" as precision and recall. These metrics remain informative in class imbalanced settings, and reflect the proportional balancing of Fundamental Rights of the EU Charter. We then focus on the estimation of recall, as its naive estimation can incur extremely high annotation costs and disproportionately interfere with the platform's right to conduct business. Through a simulation study, we show that recall can be efficiently estimated using stratified sampling with trained classifiers, and provide concrete recommendations for its application. Finally, we present a case study of recall reporting for a subset of Reddit under the Act. Based on the language in the Act, we identify a number of ways recall could be reported due to underspecification. We report on one possibility using our improved estimator, and discuss the implications and areas for further legal clarification.

Operationalizing content moderation "accuracy" in the Digital Services Act

TL;DR

This paper addresses how to operationalize the EU's DSA reporting requirement for automated content moderation accuracy by reframing 'accuracy' as precision and recall rather than raw accuracy. It couples a legal-interpretive analysis with a statistical estimation framework, proposing stratified sampling to efficiently estimate recall and reduce annotation burden. Through a Reddit case study and a CivilComments simulation, it demonstrates that recall can be reported with unbiased estimates and practical data requirements, while highlighting several underspecification areas needing regulatory clarification. The work offers concrete methodological guidance for regulators and platforms to enable meaningful, rights-respecting transparency and informs future policy design for algorithmic accountability in online moderation.

Abstract

The Digital Services Act, recently adopted by the EU, requires social media platforms to report the "accuracy" of their automated content moderation systems. The colloquial term is vague, or open-textured -- the literal accuracy (number of correct predictions divided by the total) is not suitable for problems with large class imbalance, and the ground truth and dataset to measure accuracy against is unspecified. Without further specification, the regulatory requirement allows for deficient reporting. In this interdisciplinary work, we operationalize "accuracy" reporting by refining legal concepts and relating them to technical implementation. We start by elucidating the legislative purpose of the Act to legally justify an interpretation of "accuracy" as precision and recall. These metrics remain informative in class imbalanced settings, and reflect the proportional balancing of Fundamental Rights of the EU Charter. We then focus on the estimation of recall, as its naive estimation can incur extremely high annotation costs and disproportionately interfere with the platform's right to conduct business. Through a simulation study, we show that recall can be efficiently estimated using stratified sampling with trained classifiers, and provide concrete recommendations for its application. Finally, we present a case study of recall reporting for a subset of Reddit under the Act. Based on the language in the Act, we identify a number of ways recall could be reported due to underspecification. We report on one possibility using our improved estimator, and discuss the implications and areas for further legal clarification.
Paper Structure (75 sections, 10 equations, 2 figures, 3 tables)

This paper contains 75 sections, 10 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and related work
  3. Count-based metrics
  4. Prevalence metrics
  5. Recall
  6. Metrics required by regulation
  7. The Digital Services Act
  8. Territorial scope
  9. Considerations for future implementation
  10. Scope of content moderation
  11. Definition of "content moderation"
  12. Relation to Fundamental Rights
  13. "Accuracy" as a transparency requirement
  14. Transparency requirements in the Act
  15. Troublingly, the DSA leaves the term "accuracy" underspecified.
  16. ...and 60 more sections

Figures (2)

  • Figure 1: Number of samples required to estimate the prevalence of toxicity (4.1%) within 20% in the unfiltered CivilComments, with different number of strata. Binning is based on predicted scores from a finetuned Roberta model. All pilot results are averaged over 30 trials. Dotted line represents random sampling. As the number of strata grows, estimators using optimal allocation require fewer samples. Pilot methods outperform the equal allocation baseline, but when there are too many strata, annotation effort is wasted on the pilot samples.
  • Figure 2: Number of samples required to estimate the prevalence of toxicity within 20%, for different prevalences. All pilot results are averaged over 100 trials. The numbers associated with the pilot is the total number of pilot samples across all bins. By choosing appropriate powers of two, each bin always has an integer number of pilot samples. Dotted lined represents random sampling. For small prevalences, more pilot samples are better, because more samples estimate the standard deviations within each bin better. In contrast, more bins fares worse because each bin has less pilot samples.