Disentangling Subjectivity and Uncertainty for Hate Speech Annotation and Modeling using Gaze

Özge Alaçam, Sanne Hoeken, Andreas Säuberli, Hannes Gröner, Diego Frassinelli, Sina Zarrieß, Barbara Plank2025 · DOI

Abstract

Özge Alacam, Sanne Hoeken, Andreas Säuberli, Hannes Gröner, Diego Frassinelli, Sina Zarrieß, Barbara Plank. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.

2025.

This paper introduces a framework to distinguish between genuine subjective disagreement and annotator uncertainty in hate speech datasets. By combining self-reported confidence scores with eye-tracking (gaze) data, the researchers demonstrate that these two types of label variation affect machine learning models in fundamentally different ways, suggesting that "noise" in datasets often contains valuable signal.

Introduction: Beyond "Noise" in Hate Speech Detection

Hate speech detection is inherently subjective. When annotators disagree on whether a sentence is "hateful," standard NLP pipelines typically treat this as noise to be smoothed over via majority voting or discarded. However, this research argues that disagreement stems from two distinct sources: genuine subjectivity (valid differences in opinion based on background) and uncertainty (confusion due to ambiguity or borderline cases). Distinguishing between a "controversial" case and a "hard" case is crucial for building human-aligned models.

The Quadrant-based Framework

To disentangle these dimensions, the authors propose a framework that categorizes data instances based on Individual Variation (disagreement levels) and Confidence (self-reported certainty).

The Quadrant-based Framework for categorizing label variation based on confidence and individual variation.

The framework defines four quadrants:

  • Q-I: Subjective Opinion (High Confidence, High Variation): Clear personal stance despite lack of consensus.
  • Q-II: Subjective Uncertainty (Low Confidence, High Variation): General confusion and split opinions.
  • Q-III: Collective Uncertainty (Low Confidence, Low Variation): Consensus on being unsure.
  • Q-IV: Collective Opinion (High Confidence, Low Variation): Clear-cut "ground truth" cases.

Methodology: Gaze Infusion and Datasets

The study utilizes the GAZE4HATE dataset, consisting of 90 German sentences with annotations from 43 participants. Crucially, the dataset includes high-resolution eye-tracking data recorded during reading. The researchers incorporated gaze features—such as pupil size and fixation counts—into BERT-based classifiers to see if behavioral signals could help navigate subjective complexity.

They also employed a diagnostic "weight reduction" tool, systematically down-weighting specific quadrants during training to observe changes in model behavior.

Key Insights and Results

The analysis revealed that gaze patterns are strong indicators of both hatefulness and uncertainty. For instance, total fixation count significantly predicts hatefulness ($p < .0001$), while pupil size variation correlates with annotator confidence.

Model Performance and Weighting Effects

Models pre-trained specifically on hate speech (hsd-text) significantly outperformed base German BERT models. Interestingly, the research found that "Subjective Opinion" (Q-I) instances are informative; down-weighting them often hurt performance, suggesting these controversial cases help define the boundaries of hate speech.

Model performance (F1) comparing equal weighting against fixed and distance-based reduction strategies across quadrants.

As shown in the results table, applying a "distance-based reduction" (penalizing extreme outliers) improved the base BERT model's F1 from .574 to .628. This indicates that while some variation is beneficial, extreme subjective outliers can confuse simpler models that lack behavioral context.

Conclusion

This work demonstrates that label variation in hate speech is not merely noise but a structured signal that can be disentangled using confidence and gaze data. Future work aims to scale these insights to larger datasets and explore their application in preference optimization for Large Language Models (LLMs), helping reward models better understand why humans disagree.

Summary updated

BibTeX
@misc{alacam2025disentangling,
  author = {Özge Alaçam and Sanne Hoeken and Andreas Säuberli and Hannes Gröner and Diego Frassinelli and Sina Zarrieß and Barbara Plank},
  title = {{Disentangling Subjectivity and Uncertainty for Hate Speech Annotation and Modeling using Gaze}},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.1460},
}
Figure