PaperlayerView exact revision on arXiv

Source-linked AI summary

Mitigating Demographic Disparities in Explanation-Human Alignment via Rationale Regularization

Suryansh Talukdar, Taha Sarwat, Jonathan Rusert

arXiv:2610.22586v1cs.CLcs.AI

TL;DR

In hate speech detection, strong predictive performance does not ensure equally plausible explanations across demographic groups, even though such explanations inform moderation. The paper introduces Plausibility-EFG and Fair-REF, and reports that ADV-RoBERTa improves classification and rationale alignment while reducing subgroup explanation disparities.

  • Problem

    Hate speech models can produce explanations with different plausibility across demographic groups, so predictive fairness alone does not establish explanation fairness.

    Source: §1 · p. 1

  • Method

    The paper introduces Plausibility-EFG to measure worst-case subgroup differences in attribution–rationale alignment and Fair-REF to mitigate disparities through rationale supervision and adversarial debiasing.

    Source: §1 · p. 1, §4.1 · p. 4

  • Results

    ADV-RoBERTa with Fair-REF improves classification performance and rationale alignment while reducing subgroup disparities in explanation quality.

    Source: §5.5 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    Explanation fairness is a distinct optimization objective in hate speech detection and should be evaluated separately from predictive fairness.

    Source: §1 · p. 1, §6 · p. 8

    When demographic attributes correlate with task-relevant features, sufficiently large λadv can suppress task-relevant signals and create a fairness–utility tradeoff.

    Source: Core Analytical Statements · p. 4

Abstract

from arXiv · show

We study explanation fairness in hate speech detection. High classification performance does not always mean that a model produces equally plausible explanations for different demographic groups. Using the HateXplain dataset, we introduce Plausibility-EFG, a metric that measures the largest difference in explanation plausibility across demographic groups by comparing model-generated token attributions with human-annotated rationales. We evaluate BERT, RoBERTa, and ADV-RoBERTa, an adversarially trained model, using Integrated Gradients and SHAP. Our results show that ADV-RoBERTa improves classification performance and alignment between explanations and human rationales while reducing explanation disparities across demographic groups. These improvements are observed with both attribution methods and remain consistent across training, validation, and test splits. We also show that reducing prediction-level bias alone does not eliminate disparities in explanations, suggesting that predictive fairness and explanation fairness are related but distinct objectives. Finally, our fairness-performance analysis shows that stronger adversarial regularization can further reduce Plausibility-EFG, although it may cause a small decrease in macro-F1. Overall, this work presents explanation fairness as an important objective in hate speech moderation and introduces Fair-REF, a framework designed to reduce disparities in model explanations across demographic groups.

1 Introduction

Hate speech classifiers are often judged by prediction quality even though their explanations inform moderation decisions, and explanation plausibility may differ across demographic groups. The paper introduces Plausibility-EFG to measure worst-case subgroup disparity in explanation–human alignment and Fair-REF to address it.

1 Introduction

  • Hate speech explanations matter for reviewer trust and borderline moderation decisions, yet classifiers are often evaluated primarily by predictive performance.Source: §1 · p. 1
  • Predictive fairness does not imply explanation fairness, so explanation quality should be audited independently of output-level performance.Source: §1 · p. 1
  • Plausibility-EFG measures worst-case group disparity in explanation–human alignment on HateXplain, while Fair-REF is a mitigation pipeline designed to improve explanation fairness while preserving strong task performance.Plausibility is measured by AUPRC over attribution scores and human rationales.Source: §1 · p. 1

2 Related Work

Prior work uses post-hoc attribution and human-rationale benchmarks to evaluate NLP explanations, while distinguishing whether explanations reflect model decisions from whether they match human judgments. This paper studies rationale plausibility across demographic groups and extends explanation-fairness work from auditing toward mitigation.

2 Related Work

  • Integrated Gradients and SHAP provide complementary token-attribution approaches that can be compared with human evidence spans in rationale benchmarks such as ERASER and HateXplain.Rationale extraction models such as REF instead include explanation generation in the training objective.Source: §2 · pp. 1–2
  • Faithfulness concerns whether an explanation reflects the model’s decision process, whereas plausibility concerns whether it matches human judgment or annotation.Prior work reports subgroup variation in explanation fidelity and finds that fairness and explainability do not necessarily co-improve in NLP.Source: §2 · p. 2
  • Unlike prior explanation-fairness studies focused on fidelity, this work measures plausibility on HateXplain and proposes Fair-REF as training-time mitigation of group-level rationale quality.The cited prior work motivates the paper’s orthogonality hypothesis.Source: §2 · p. 2

3 Theoretical Justification

The theoretical framing treats predictive performance, rationale alignment, and subgroup-consistent explanation quality as distinct but jointly relevant objectives. It defines a worst-case explanation fairness gap, motivates adversarial representation regularization as a surrogate, and states analytical and empirical qualifications on measurement and tradeoffs.

3 Theoretical Justification

  • Fair-REF combines task, rationale, and adversarial demographic losses, using gradient reversal to suppress demographic information in a shared representation.The framework follows adversarial debiasing and rationale-alignment formulations.Source: §3 · p. 2
  • Integrated Gradients and SHAP provide attribution scores, while ranking- and overlap-based plausibility measures compare those scores with human rationales.The two explainers can preserve comparative model ordering while differing in attribution geometry and scale.Source: §3 · pp. 2–3
  • Plausibility-EFG is the worst-case inter-group difference in mean subgroup plausibility, operationalizing whether explanations are comparably plausible across demographic groups.The framework also uses significance testing and effect size to distinguish gap magnitude from distributional differences.Source: §3 · p. 3
  • Task-performance and explanation-fairness gradients need not align, so improving classification risk or output-level fairness does not guarantee more equitable explanations.Source: §3 · p. 3

Core Analytical Statements

  • Theoretical properties establish that EFG lies in [0, 1], is symmetric across groups, and is zero exactly when all group means match.Subgroup scarcity increases uncertainty in estimated group means, EFG, and effect sizes.Source: Core Analytical Statements · p. 4
  • Empirically, improved classification from RoBERTa over BERT leaves non-trivial subgroup gaps, while explanation fairness improves substantially under ADV-RoBERTa with Fair-REF.The paper treats this partial orthogonality claim as primarily empirical.Source: Core Analytical Statements · p. 4
  • When demographic attributes correlate with task-relevant features, sufficiently large adversarial regularization can suppress both group information and some task-relevant signals, creating a fairness–utility tradeoff.The authors describe this as an empirically observed conditional phenomenon, not a universal law; stronger adversarial pressure can reduce EFG while moderately degrading Macro-F1.Source: Core Analytical Statements · p. 4, §3 · p. 4
  • Normalized attribution methods can preserve model ordering while differing in absolute plausibility values, motivating use of both explainers in fairness analysis.Source: Core Analytical Statements · p. 4

4 Methodology

The methodology evaluates classifiers on HateXplain for prediction, rationale alignment, and subgroup explanation fairness. Fair-REF combines classification, rationale supervision, and adversarial debiasing, while attribution-based metrics audit plausibility and its demographic gap.

4.1 Problem Formulation

  • HateXplain supplies posts, majority-vote labels, human rationale masks, and demographic target groups for jointly evaluating prediction, explanation plausibility, and explanation fairness.Source: §4.1 · p. 4

4.2 Model Setup

  • The comparison includes BERT, RoBERTa, and ADV-RoBERTa, whose shared encoder feeds classification, demographic-adversary, and token-rationale heads.The adversarial head uses gradient reversal, and the rationale head predicts token-level scores.Source: §4.2 · p. 4

4.3 Training Objective

  • Training combines task, adversarial, and rationale losses; the adversarial loss is a differentiable surrogate because attribution-dependent Plausibility-EFG cannot be optimized directly.The surrogate is not a formal guarantee that EFG will be minimized; the primary setting selects λadv = 0.5 and λrat = 0.2 using validation-based balance of macro-F1 and Plausibility-EFG.Source: §4.3 · pp. 4–5

4.4 Attribution Methods

  • Integrated Gradients measures path-integrated gradients relative to a baseline, whereas SHAP estimates marginal feature contributions through cooperative game theory.Using both methods checks whether fairness conclusions depend on one explainer; differences in scale can affect fairness measurements.Source: §4.4 · p. 5

4.5 Plausibility and Fairness Metrics

  • AUPRC and IOU-F1 measure attribution–rationale plausibility, with AUPRC primary because rationales are sparse; Plausibility-EFG captures the largest subgroup disparity.The evaluation also reports Kruskal–Wallis significance, Cohen’s d, and bootstrap confidence intervals.Source: §4.5 · p. 5

4.6 Orthogonality and Mitigation

  • Fair-REF treats prediction and explanation fairness as distinct objectives, combining adversarial debiasing with rationale supervision and sweeping λadv to assess the fairness–performance frontier.Source: §4.6 · p. 5

5 Experiments and Results

Across HateXplain experiments, ADV-RoBERTa pairs stronger classification and rationale alignment with smaller subgroup explanation gaps. Ablations and adversarial-regularization sweeps show that explanation fairness requires explicit objectives and can trade off against predictive performance.

5.1 Dataset and Evaluation Pipeline

  • HateXplain’s 20,148 posts provide labels, token rationales, and demographic identities; evaluation covers post-level classification and exploded subgroup explanation fairness.The pipeline evaluates classification, attribution, plausibility, subgroup benchmarks, explanation fairness, ablations, and fairness–performance tradeoffs.Source: §5.1 · pp. 5–6

5.2 Hate Speech Classification

  • ADV-RoBERTa improves accuracy by 0.0891 and macro-F1 by 0.0675 over BERT, and leads both baselines on accuracy, macro-F1, and weighted-F1.The paper reports that fairness-aware optimization does not degrade classification performance.Source: §5.2 · p. 6, Table 1 · p. 7

5.3 Explanation Plausibility

  • ADV-RoBERTa ranks highest on all reported IG plausibility metrics; mean AUPRC rises from 0.52 to 0.59 post-level and reaches 0.60 exploded-level.Train, validation, and test values remain close, and SHAP shows the same model ordering as IG.Source: §5.3 · p. 6, Table 2 · p. 7

5.5 Component Ablation of Fair-REF

  • Rationale supervision mainly improves plausibility, while adversarial debiasing mainly reduces explanation gaps; combining both yields the strongest overall classification, alignment, and Plausibility-EFG results.Fair-REF combines task, rationale, and adversarial losses.Source: §5.5 · p. 7

5.4 Quantifying Explanation Fairness

  • ADV-RoBERTa achieves the highest IG and SHAP plausibility and lowest Plausibility-EFG, while reducing subgroup disparities across both explainers.The EFG reduction is accompanied by lower Cohen’s d and non-significant Kruskal–Wallis statistics under SHAP.Source: §5.4 · p. 7, Figure 2 · p. 7, Table 4 · p. 8

5.6 Fairness–Performance Tradeoff

  • Stronger adversarial regularization monotonically reduces EFG, but excessive strength can lower classification performance; λadv = 0.3 and 0.5 are practical moderate settings.The λadv = 1.0 setting gives the lowest EFG but a larger performance drop than λadv = 0.1.Source: §5.6 · p. 8

5.7 Discussion

  • Adversarial debiasing suppresses decodable demographic information in shared representations, reducing the role of subgroup-specific artifacts in attribution patterns and explanation disparities.The authors argue explanation fairness requires dedicated objectives rather than being assumed from prediction-level debiasing.Source: §5.7 · p. 8

6 Conclusion

On HateXplain, ADV-RoBERTa improves classification and rationale alignment while reducing subgroup disparities, supporting explanation fairness as a distinct objective. Fair-REF explicitly measures and optimizes this objective.

6 Conclusion

  • ADV-RoBERTa improves classification performance and rationale alignment while reducing subgroup disparities in explanation quality on HateXplain.These findings support treating predictive fairness and explanation fairness as related but distinct objectives that require separate evaluation.Source: §6 · p. 8
  • Fair-REF demonstrates that explanation fairness can be explicitly measured and optimized as a practical framework for hate speech moderation.Source: §6 · p. 8

A Limitations

The limitations concern a fairness–performance tradeoff, attribution-method sensitivity, subgroup scarcity, and ambiguity or inconsistency in human rationales. These factors constrain interpretation of alignment and fairness estimates.

A Limitations

  • As λadv rises from 0.1 to 1.0, Mean EFG falls from 0.048 to 0.024 while Macro-F1 changes from 0.7448 to 0.7327.The authors caution that stronger adversarial regularization can overregularize the shared representation, so fairness should be constrained by predictive utility.Source: A Limitations · p. 10
  • IG and SHAP agree on model ranking but yield different absolute plausibility scores and EFG values.For ADV-RoBERTa, post-level plausibility is 0.59 under IG and 0.52 under SHAP, while fairness gaps are 0.03 and 0.02, respectively; the authors recommend reporting multiple explainers.Source: A Limitations · p. 10
  • Smaller groups such as Refugee and Caucasian have fewer instances, increasing variance in group-level AUPRC and EFG estimates.The authors therefore consider fairness claims stronger for well-represented groups such as Women and African than for underrepresented ones.Source: A Limitations · p. 10
  • Mixed or inconsistent human rationales make perfect alignment difficult, while low rationale density increases IOU-F1 sensitivity to thresholding.Consequently, low scores may reflect annotation ambiguity rather than a true explanation failure.Source: A Limitations · p. 10
  • Higher rationale alignment reflects improved human-aligned explanation quality, not guaranteed causal interpretability.The authors also note that attribution methods are not ground-truth explanations and plausibility should not be conflated with faithfulness to a model’s decision process.Source: A Limitations · p. 10

B Additional Analysis

Additional analyses find that ADV-RoBERTa improves explanation plausibility across demographic groups under both IG and SHAP, with subgroup gaps narrowing. Increasing adversarial regularization reduces explanation disparity while trading off some classification performance.

B Additional Analysis

  • Across experiments, explanation plausibility improves across demographic groups under both Integrated Gradients and SHAP, with mitigation following a controlled fairness–performance tradeoff.Source: B Additional Analysis · p. 10
  • Under IG, lower-performing groups move closer to the top group, while SHAP shows the same directional improvement across identities.The group-wise tables and figures indicate reduced subgroup variance rather than gains confined to the strongest or most frequent groups.Source: B Additional Analysis · p. 10, Table 7 · p. 11, Figure 4 · p. 11
  • Increasing λadv steadily reduces explanation disparity with only a moderate classification-performance cost, and the reported tradeoff curve is smooth.The analysis presents fairness regularization as tunable rather than an all-or-nothing tradeoff.Source: B Additional Analysis · p. 11, Table 9 · p. 12, Figure 6 · p. 12
  • The demographic distribution is imbalanced, so aggregate plausibility alone could hide meaningful variation across groups.The appendix points to subgroup-level evaluation as necessary for assessing explanation fairness.Source: B Additional Analysis · p. 11
  • ADV-RoBERTa remains strongest across classification, plausibility, and fairness metrics in the consolidated model comparison.The appendix characterizes the improvements as systematic and stable across attribution methods.Source: B Additional Analysis · pp. 11–12

C.1 Proof of Proposition 1 (Basic properties of EFG)

The proposition establishes basic properties of EFG: it is bounded, symmetric, and zero exactly when all subgroup means are equal.

C.1 Proof of Proposition 1 (Basic properties of EFG)

  • Because each plausibility score lies in [0, 1], every subgroup mean also lies in [0, 1].Source: §C.1 · p. 12
  • Taking the maximum of pairwise subgroup differences preserves the bound, and absolute value makes EFG symmetric.Source: §C.1 · p. 12
  • EFG equals zero if and only if every pair of subgroup means is equal.Source: §C.1 · p. 12

C.2 Proof of Proposition 2 (Subgroup-scarcity variance effect)

Subgroup estimates become less stable as group support shrinks, while prediction and explanation-fairness objectives are structurally distinct. Improving prediction therefore need not reduce explanation disparities.

C.2 Proof of Proposition 2 (Subgroup-scarcity variance effect)

  • As subgroup size falls, subgroup-mean variance decays only at rate 1/|D_g|, making sparse-group fairness estimates more sensitive and motivating bootstrap intervals and cautious interpretation.The same instability applies to Cohen’s d when one subgroup has limited support.Source: §C.2 · p. 13
  • Toxicity-label composition varies considerably across demographic identities, motivating subgroup-level fairness evaluation.Source: Figure 8 · p. 13
  • Across IG and SHAP, ADV-RoBERTa achieves the strongest overall balance of classification performance, rationale alignment, and subgroup fairness among the three compared models.Source: Figure 9 · p. 13

C.3 Formal argument for Observation 1 (Partial orthogonality)

  • Prediction and explanation-fairness objectives depend on distinct features and attribution-rationale alignments, so prediction improvement need not reduce explanation gaps.Their gradients are not generally proportional, and prediction loss can decrease while explanation fairness remains unchanged or worsens.Source: §C.3 · p. 14

C.4 Proof sketch for Proposition 3 (Conditional tradeoff under fairness regularization)

The appendix frames adversarial fairness regularization as a conditional fairness–utility tradeoff and compares explanation rankings across attribution methods. It then describes the selected baselines, Fair-REF architecture, and joint training objectives.

C.4 Proof sketch for Proposition 3 (Conditional tradeoff under fairness regularization)

  • When demographic information is entangled with task-relevant semantics, sufficiently strong adversarial suppression can reduce EFG while degrading predictive utility.The resulting fairness–utility Pareto frontier is conditional optimization intuition, not a universal theorem.Source: §C.4 · p. 14

C.5 Lemma and formal argument for Observation 2 (Explainer-level ranking stability)

  • IG and SHAP agree on the model ordering BERT < RoBERTa < ADV-RoBERTa, but differ in absolute plausibility scale and subgroup spread.The methods are not guaranteed to be strict monotonic transforms of each other, so agreement concerns ranking rather than identical attribution magnitudes.Source: §C.5 · pp. 14–15

D.1 Baseline Justification

  • BERT, RoBERTa, and ADV-RoBERTa provide a controlled comparison because ADV-RoBERTa keeps the RoBERTa architecture fixed while changing the training objective.DeBERTa and HateBERT are excluded for architectural or pretraining confounds, and decoder-only LLMs require methodological changes for comparable attribution evaluation.Source: §D.1 · p. 15

D.2 Detailed Model Architectures

  • Fair-REF’s ADV-RoBERTa combines a shared encoder with classification, demographic-adversary, and rationale-prediction branches.The rationale head uses human-annotated masks to provide an explicit alignment signal [Mathew et al., 2021], alongside adversarial invariance learning [Zhang et al., 2018].Source: §D.2 · p. 15

D.3 Forward Pass and Loss Construction

  • Fair-REF jointly optimizes task, rationale-supervision, and adversarial losses to encourage accurate prediction, rationale generation, and explanation fairness.Rationale loss directly supervises salient tokens against human evidence spans, and instances without valid rationale annotations do not contribute to that term.Source: Figure 10 · p. 16, §D.3 · p. 16

D.4 Gradient Reversal and Fairness Ablation

The paper combines adversarial invariance with rationale supervision and evaluates attribution-based plausibility and subgroup disparity. Its central finding is that explanation fairness is not guaranteed by prediction fairness, while stronger regularization can trade predictive utility for lower explanation disparity.

D.4 Gradient Reversal and Fairness Ablation

  • Gradient reversal trains the encoder to minimize task and rationale losses while maximizing demographic-adversary error, encouraging task-discriminative, rationale-aligned, demographically invariant representations.The fairness-strength sweep varies λfair while holding λrat fixed; with λfair = 0, the adversarial branch is disabled.Source: §D.4 · pp. 16–17

D.5 Training Dynamics and Stability

  • To stabilize adversarial training, the protocol warms up λadv, clips gradients, applies [CLS] dropout, selects checkpoints by validation macro-F1, and uses early stopping.The safeguards address instability early in training.Source: §D.5 · p. 17

D.6 Attribution Details

  • IG integrates gradients from a zero-vector baseline, while SHAP estimates token importance with Shapley values; both attribution sets are normalized and used for plausibility and fairness metrics.IG negative values and SHAP negative values are excluded; the SHAP evaluation budget is 256 calls per sample.Source: §D.6 · p. 17

D.7 Plausibility and Fairness Metrics

  • AUPRC measures attribution-score ranking against human rationale tokens, while Plausibility-EFG measures worst-case inter-group disparity.The evaluation also reports Kruskal–Wallis H-statistics, Cohen’s d, and bootstrap 95% confidence intervals.Source: §D.7 · p. 17

D.8 Training and Hardware Configuration

  • All models use the same preprocessing and optimization settings, with validation macro-F1 determining checkpoint selection and λfair swept across five values.Training uses a maximum sequence length of 128 tokens and a single NVIDIA A100 40GB GPU.Source: §D.8 · pp. 17–18

D.9 Dataset Construction Pipeline

  • The fairness-evaluation set filters HateXplain to 27,566 rows covering eight canonical identity groups with nonempty rationale masks.Multi-target posts contribute attribution scores to multiple groups, but the taxonomy does not construct explicit intersectional categories and induces dependence between subgroup estimates.Source: §D.9 · p. 18

D.10 Orthogonality and Mitigation Framing

  • Prediction fairness and explanation fairness are related but only partially orthogonal; improving one does not guarantee improvement in the other.Fair-REF directly lowers Plausibility-EFG through rationale alignment and adversarial invariance, while higher λfair values introduce a measurable performance cost.Source: §D.10 · p. 18

E Ethical Considerations

  • Human rationales are annotator-provided evidence, not absolute ground truth, so plausibility scores indicate annotation agreement rather than proof of explanation correctness.The ethical discussion also cautions against overinterpreting small groups and calls for human oversight in deployment.Source: E Ethical Considerations · pp. 18–19
Loading 2610.22586v1…