Disagreeing Rationales: Rethinking Classification and Explainability Evaluation in Hate Speech Detection

The field of NLP has seen rapid advancements with the advent of large pre-trained models like BERT (Devlin et al., 2019), which significantly improved performance on various tasks. However, subjective NLP tasks—such as hate speech detection, offensive language identification, and sentiment analysis—pose unique challenges due to the inherent variability in human judgments. Traditional supervised learning approaches rely on single ground-truth labels, often obtained via majority voting, which oversimplifies the complex, context-dependent nature of human perception. Recent efforts, such as UMA et al. (2025), have begun to explore multi-annotator datasets and probabilistic labels, acknowledging that different individuals may interpret content differently based on cultural, personal, and contextual factors. Despite these advances, existing evaluation frameworks primarily focus on classification accuracy, neglecting the interpretability and diversity of explanations. As explainability becomes a critical component for deploying NLP models in sensitive applications, there is a pressing need for evaluation paradigms that incorporate human judgment variability, especially at the token-level rationales. This background sets the stage for the current work, which aims to systematically analyze how diverse annotations influence both model performance and interpretability in subjective tasks.

The core challenge addressed in this paper is the modeling and evaluation of human judgment diversity in subjective NLP tasks, particularly hate speech detection. Human annotators often disagree on whether a statement is hateful, and their rationales—highlighted tokens or explanations—also vary significantly. Traditional models and evaluation metrics assume a single 'truth,' which leads to biased or incomplete understanding, especially in socially sensitive contexts. This oversimplification hampers the development of fair, transparent, and robust systems capable of handling real-world complexity. Moreover, existing explainability methods, such as attention weights, do not adequately reflect the diversity of human rationales, raising questions about their fidelity and trustworthiness. The problem is compounded by the lack of standardized evaluation protocols that account for annotation disagreement at both label and rationale levels, limiting progress in building models that genuinely reflect human perspectives and promote fairness.

The key innovations introduced by this work include: • A multi-representation modeling framework that supports hard, soft, and intermediate label and rationale spaces, enabling models to learn from the full spectrum of human judgments. • A unified evaluation protocol that integrates classification and explainability metrics across diverse annotation spaces, providing a comprehensive assessment of model performance. • The adoption of multi-objective loss functions—cross-entropy, MSE, and KL divergence—that facilitate training models to handle annotation variability effectively. • Systematic re-implementation of existing rationale-based models (e.g., MRP, SRA) within this flexible framework, revealing the impact of rationale representation choices on performance. These innovations collectively address the limitations of traditional single-ground-truth approaches, offering a more nuanced understanding of human judgment diversity and its implications for model interpretability and fairness.

The methodology involves several interconnected steps:

• �� Designing multiple label and rationale spaces: including hard labels (single class), soft labels (probabilistic distributions), and intermediate representations that capture annotation disagreement.
• �� Collecting token-level rationales from multiple annotators across datasets like HateXplain and HateBRXplain, ensuring diverse perspectives.
• �� Developing multi-objective loss functions: combining cross-entropy for classification accuracy, MSE for rationale regression, and KL divergence for distribution alignment.
• �� Training Transformer-based models (e.g., BERT) with attention mechanisms to extract rationales, incorporating the multi-space labels and rationales.
• �� Implementing evaluation metrics: classification performance (macro F1, JSD), rationale plausibility (AUPRC, IoU), faithfulness (attention alignment), and complexity (entropy, Gini index).
• �� Conducting experiments across different configurations—hard, soft, and intermediate representations—and datasets, with statistical significance testing (paired t-test, FDR correction) to validate findings.
• �� Analyzing the sensitivity of metrics to representation choices and exploring correlations among evaluation dimensions to understand their interplay.

The experimental setup involves two multilingual datasets: HateXplain (English) and HateBRXplain (Portuguese), both containing token-level rationales and multiple annotations. Models are trained under various configurations—hard labels, soft labels, and intermediate representations—using the same hyperparameters for comparability. Baselines include original models like MRP and SRA, re-implemented within the proposed framework. The evaluation employs classification metrics such as macro F1-score and Jensen-Shannon divergence (JSD), as well as explainability metrics—plausibility (AUPRC, IoU), faithfulness (attention alignment), and complexity (entropy, Gini). Multiple runs with cross-validation ensure robustness, and statistical tests confirm significance. The analysis focuses on how different label and rationale spaces influence performance, stability, and interpretability, with particular attention to the benefits of soft representations in capturing annotation diversity.

Experimental results demonstrate that models trained with soft labels and rationales outperform traditional hard-label models across all metrics. In HateXplain, soft configurations improved macro F1-score by approximately 2.5% and AUPRC by 3.2%, indicating better classification and explanation quality. In HateBRXplain, the IoU score for rationales increased to 0.68, reflecting higher alignment with human rationales. Sensitivity analysis shows that models with soft rationales exhibit more stable performance across datasets and configurations, with higher correlations among evaluation metrics, especially in plausibility and faithfulness. The results confirm that softer representations better capture the diversity of human judgments, leading to explanations that are both more accurate and more trustworthy. These findings advocate for adopting flexible annotation spaces in subjective NLP tasks.

This framework is directly applicable to content moderation, hate speech detection, and social media analysis, where understanding diverse human perspectives is crucial. Organizations can leverage soft labels and rationales to develop more inclusive and fair moderation tools, reducing biases and improving transparency. The approach also benefits academic research by providing comprehensive evaluation tools for explainability, fostering the development of models that better reflect societal values. In the long term, integrating multi-modal data (images, videos) and natural language explanations can further enhance interpretability and user trust. The methodology supports building AI systems capable of nuanced understanding and respectful engagement in socially sensitive contexts, ultimately promoting responsible AI deployment.

The current study relies on limited datasets (HateXplain and HateBRXplain), which may not fully encompass the diversity of real-world social media content, potentially affecting generalizability. The models are based on BERT, and performance with larger or more recent architectures (e.g., GPT-4, T5) remains to be validated. The reliance on attention weights for rationale evaluation may introduce biases, and alternative interpretability methods should be explored. Handling multi-modal data and free-text explanations requires further methodological development. Computational costs associated with training models across multiple representation spaces pose scalability challenges. Additionally, the framework's effectiveness in dynamic, evolving social environments needs further investigation.