💬 NLP

Self-Explaining Hate Speech Detection with Moral Rationales

This paper introduces Supervised Moral Rationale Attention (SMRA), a novel self-explaining framework that aligns model attention with expert-annotated moral rationales based on Moral Foundations Theory to improve the robustness and interpretability of hate speech detection, supported by the newly released HateBRMoralXplain dataset in Brazilian Portuguese.

Original authors: Francielle Vargas, Jackson Trager, Diego Alves, Surendrabikram Thapa, Matteo Guida, Berk Atil, Daryna Dementieva, Andrew Smart, Ameeta Agrawal

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Francielle Vargas, Jackson Trager, Diego Alves, Surendrabikram Thapa, Matteo Guida, Berk Atil, Daryna Dementieva, Andrew Smart, Ameeta Agrawal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a giant, noisy digital town square where people shout, joke, argue, and sometimes say terrible things. In the world of computer science, specifically in a field called Natural Language Processing (which is just a fancy way of saying "teaching computers to understand human language"), scientists are trying to build digital bouncers. These bouncers are algorithms designed to spot "hate speech"—the mean, harmful, and dangerous shouting that targets people based on who they are. But here's the tricky part: computers are often terrible at understanding why something is mean. They are like kids who memorized a list of "bad words" but don't understand the difference between a playful insult among friends and a genuine threat. If a computer only looks for specific bad words, it might get confused by sarcasm, cultural jokes, or cleverly disguised hate. This is a big problem because if the digital bouncer makes mistakes, it might silence innocent people or let dangerous hate slip through the cracks. To fix this, researchers are trying to teach computers not just what to flag, but why they are flagging it, using a concept called "moral reasoning"—basically, asking the computer to think about fairness, harm, and loyalty, just like a human does.

Enter a new team of researchers who decided to stop guessing and start teaching computers to think like moral philosophers. They introduced a clever new system called Supervised Moral Rationale Attention (SMRA). Think of SMRA as a training program for a digital detective. Instead of just showing the detective a list of "bad words" to look for, the trainers (human experts) show them specific sentences and say, "Look here! This part is bad because it hurts someone's feelings," or "This part is bad because it breaks the rules of fairness." The computer learns to pay attention to these specific "moral clues" rather than just scanning for surface-level keywords. To make this training possible, the team also created a massive new library of examples called HateBRMoralXplain. This isn't just a list of bad comments; it's a collection of 7,000 real Instagram comments from Brazil, carefully labeled by experts who explained exactly which moral rule was broken and why. It's like giving the detective a textbook filled with real-life cases and the expert's handwritten notes on the margins.

When they put this new detective to the test, the results were promising. The SMRA system didn't just get slightly better at spotting hate speech; it got better at explaining why it spotted it. In tests, the system improved its accuracy in finding hate speech and, more importantly, its ability to point to the exact words that made the comment harmful. The paper suggests that by focusing on these deeper moral reasons, the computer becomes less likely to be tricked by spurious correlations (like thinking a word is bad just because it often appears near other bad words) and more likely to understand the actual context. Interestingly, the team found that while the computer got better at finding the "moral" reasons, it didn't necessarily become "fairer" in every single way, suggesting that teaching a computer to be moral is a complex journey, not a magic switch. They also tried using giant, powerful AI models (like LLMs) to do the job, but found that even these super-smart models struggled to understand the moral nuances unless they were given very specific instructions and the right context.

The researchers also discovered something fascinating about the "why." When they compared their moral-focused detective to one that only looked for "hate" words, the moral-focused one was much better at aligning with human judgment. However, they noted that the explanations became more concise—shorter and punchier—which is good, but sometimes meant the computer missed a few tiny details (a slight drop in "comprehensiveness"). The paper suggests that this trade-off is worth it because the explanations are more faithful to the model's actual thinking process. In the end, the team concludes that while their new method isn't a perfect solution that solves all problems, it is a significant step forward. It shows that if we want computers to be good digital citizens, we need to teach them to care about the moral weight of words, not just the words themselves. The paper ends with a note of caution: this is still early research, and while the results are encouraging, there is still a lot to learn about how to make these systems work across different cultures and languages without introducing new biases.

Technical Summary: Self-Explaining Hate Speech Detection with Moral Rationales

Problem Statement

Existing hate speech (HS) detection models suffer from opacity and over-reliance on surface-level lexical cues. This reliance leads to vulnerability against spurious correlations, reduced robustness, and a lack of cultural contextualization. Furthermore, standard models often inherit biases from training data and annotation processes, resulting in the unfair targeting of marginalized groups. While explainable AI approaches exist, they are largely categorized as either post-hoc (generating explanations after prediction, often unfaithful or computationally expensive) or self-explaining (integrating explanation into the learning process). However, current self-explaining methods often fail to capture the normative and moral reasoning essential for distinguishing culturally contingent hate speech from harmless language. Recent studies suggest that moral values constitute a transferable latent representation across languages, yet standard models do not explicitly leverage this.

Methodology: Supervised Moral Rationale Attention (SMRA)

The authors propose Supervised Moral Rationale Attention (SMRA), a self-explaining framework designed to align neural attention with expert-annotated moral rationales.

Core Mechanism

SMRA modifies standard transformer-based classifiers (e.g., BERT) by introducing a normative inductive regularization into the training objective. Instead of relying solely on lexical markers, the model is guided to attend to tokens identified as morally salient by human experts.

Task Definition: The framework addresses two tasks:
- Binary Hate Speech Classification: Predicting if content is hate speech.
- Multi-label Moral Sentiment Classification: Identifying which Moral Foundations Theory (MFT) categories (Care/Harm, Fairness/Cheating, Loyalty/Betrayal, Authority/Subversion, Purity/Degradation) are present.
Attention Alignment Loss: For training instances with available moral rationales, SMRA minimizes the Mean Squared Error (MSE) between the model's normalized attention distribution ( $a$ ) and a binary rationale mask ( $r$ ) derived from human annotations.
Objective Function: The total loss ( $\ell_{total}$ ) combines the standard cross-entropy loss ( $\ell_{CE}$ ) for classification with the supervised attention alignment loss ( $\ell_{MSE}$ ):
$\ell_{total} = \ell_{CE} + \alpha \cdot \ell_{MSE}$
Where $\alpha$ controls the strength of the attention supervision. This loss is applied only to samples with annotated moral rationales.

Dataset: HateBRMoralXplain

To support this framework, the authors introduce HateBRMoralXplain, an extension of the Brazilian Portuguese HateBR benchmark.

Composition: 7,000 Instagram comments (3,500 offensive, 3,500 non-offensive) annotated by three experts.
Annotations: Includes hate labels, MFT categories, and token-level moral rationales (minimal text spans justifying the moral label).
Metadata: Enriched with socio-political metadata (political party, gender of the original post author) to facilitate bias analysis.
Inter-annotator Agreement: Substantial to almost perfect agreement for primary moral labels ( $\kappa = 0.811$ ), with moderate to substantial agreement for secondary and tertiary labels.

Key Contributions

SMRA Framework: The first self-explaining hate speech detection framework that aligns model attention directly with expert-annotated moral rationales, aiming to improve interpretability, robustness, and cultural contextualization.
HateBRMoralXplain Dataset: A large-scale, expert-annotated corpus for low-resource Brazilian Portuguese, uniquely enriched with moral categories, human-annotated rationales, and socio-political metadata.
Empirical Evaluation: A comprehensive evaluation of SMRA against baselines (standard fine-tuning, post-hoc methods like LIME/SHAP, and other self-explaining approaches like SRA) and Large Language Models (LLMs) using various prompting strategies.

Results

Experiments were conducted on binary hate speech detection and multi-label moral sentiment classification using mBERT and BERTimbau.

Performance Improvements

Predictive Performance: SMRA consistently improved performance over baselines. In binary classification, it increased Accuracy (+0.0004) and Macro F1 (+0.0082). In multi-label moral classification, it improved Macro F1 (+0.015) and AUROC (+0.002).
Explanation Quality (Plausibility): SMRA significantly enhanced alignment with human rationales.
- IoU F1: +0.0743 (binary) and +0.241 (moral) improvements over baselines.
- Token F1: +0.0503 (binary) and +0.241 (moral) improvements.
Explanation Quality (Faithfulness):
- Comprehensiveness: Slightly decreased (−0.0203), indicating explanations became more concise.
- Sufficiency: Improved (−0.0231, noting the metric definition where lower is better for the drop in probability, or interpreted as the paper's context of "improved" meaning more compact yet faithful rationales). The paper notes SMRA produces more concise explanations that remain faithful.
Fairness: Fairness metrics (GMB-Sub, GMB-BPSN, GMB-BNSP) remained stable, suggesting that improvements in explanation quality did not introduce significant bias trade-offs.

Comparative Analysis

SMRA vs. SRA (Supervised Rational Attention): SMRA achieved substantially higher plausibility (IoU F1 +0.1195, Token F1 +0.1508) compared to SRA, which relies on lexical rationales. SRA showed unstable faithfulness (negative Sufficiency), whereas SMRA maintained positive Sufficiency, indicating more stable causal alignment between rationales and predictions.
LLM Performance: LLMs (GPT-4o-mini, Llama-3.1-70B) showed notable improvements in hate speech classification (approx. 2–3%) when prompted with moral rationales and definitions. However, LLMs performed poorly on moral sentiment classification (max F1 of 0.38). Translation of prompts to English resulted in performance drops, highlighting the importance of language-specific context.

Significance and Claims

The paper claims that SMRA represents a significant step toward more interpretable and culturally aware hate speech detection. By integrating moral reasoning directly into the training objective, the framework moves beyond surface-level lexical cues to capture the normative structures underlying hate speech.

The authors assert that:

Moral Rationales are Superior to Lexical Cues: Aligning attention with moral rationales reduces reliance on spurious correlations and improves cross-cultural generalization compared to methods using only lexical rationales.
Inherent Interpretability: Unlike post-hoc methods, SMRA produces inherently interpretable explanations that are faithful to the model's decision process.
No Bias Trade-off: Enhancing explanation quality and robustness does not necessarily degrade fairness; the model maintained stable fairness metrics.
Contextual Necessity: The study highlights that hate speech detection requires cultural and moral context, particularly for low-resource languages and nuanced cases involving sarcasm or political context, which standard models often miss.

The authors conclude that while SMRA improves robustness and interpretability, it relies on costly human annotation and is currently evaluated primarily on Brazilian Portuguese data, necessitating future work on scalability and cross-lingual generalization.