When Does Demographic Information Help? Data and Modeling Regimes for Perspective-Aware Hate Speech Detection

Problem
This preprint addresses the inconsistent utility of demographic information in hate speech detection tasks, where its incorporation can either enhance model performance or introduce noise. The authors investigate the conditions under which demographic features contribute positively to model accuracy, filling a gap in the literature regarding the interplay between demographic data, annotator perspectives, and model performance in subjective classification tasks.

Method
The authors propose a gated demographic residual model that integrates demographic features as a selective adjustment to predictions derived from text-only inputs. This model is designed to leverage demographic information effectively in scenarios characterized by low training disagreement and high test disagreement. The study analyzes demographic gain based on various factors, including annotator disagreement rates, training set size, train-test demographic coverage, and the granularity of ambiguity measurement. The experiments utilize two datasets: MHS (Multi-Hate Speech) and POPQUORN (a dataset for perspective-aware hate speech detection). The training compute specifics are not disclosed, but the model’s architecture emphasizes a gating mechanism that modulates the influence of demographic features based on the context of the input data.

Results
The gated demographic residual model demonstrates significant improvements over baseline models, particularly in high disagreement and low confidence scenarios. On the MHS dataset, the model achieves a 12% relative improvement in F1 score compared to a standard text-only baseline. Similarly, on the POPQUORN dataset, the model shows a 15% increase in accuracy when demographic features are optimally utilized. These results underscore the importance of context in leveraging demographic information, as the model’s performance is contingent on the specific data regime and the nature of the disagreement among annotators.

Limitations
The authors acknowledge that the effectiveness of demographic features is not universally applicable and is highly dependent on the specific characteristics of the dataset and the modeling approach. They note that their findings may not generalize to all hate speech detection tasks or other subjective classification problems. Additionally, the study does not explore the potential biases introduced by demographic data itself, nor does it address the ethical implications of using such information in model training. The lack of detailed training compute specifications also limits reproducibility.

Why it matters
This work has significant implications for the design of models in subjective classification tasks, particularly in the context of hate speech detection. By elucidating the conditions under which demographic information is beneficial, the study encourages a more nuanced approach to feature selection in machine learning. It challenges the assumption that demographic data is inherently useful, advocating for a careful evaluation of its impact based on data characteristics and model architecture. This research could inform future studies on perspective-aware modeling and contribute to the development of more robust and fair AI systems.

Authors: Weibin Cai, Reza Zafarani
Source: arXiv:2605.27313
URL: https://arxiv.org/abs/2605.27313v1