Human-Aligned Rationale Learning for Explainable Hate Speech Moderation: A Behavioral Evaluation Framework for Reliable Decision Support

1. Introduction

Online platforms increasingly rely on automated systems to moderate harmful content at scale. The rapid growth of user-generated content across social media platforms, discussion forums, and online gaming environments has made purely human moderation operationally infeasible. As a result, machine learning models are widely deployed to detect and filter abusive language, harassment, and hate speech [1,2]. While such systems enable large-scale moderation, their deployment introduces significant ethical and governance challenges, particularly when automated decisions affect public discourse, user participation, and the visibility of marginalized voices [3,4]. From an applied perspective, moderation systems must therefore support not only accurate detection but also interpretable and inspectable decision processes that can be integrated into human-in-the-loop workflows.

The use of artificial intelligence in content moderation raises broader questions at the intersection of technology and governance. Automated moderation systems function as de facto gatekeepers of online speech, shaping which content is amplified or suppressed within digital platforms [5]. Errors in these systems therefore carry important implications: false positives may suppress legitimate speech, while false negatives may allow harmful content to persist [6,7]. These challenges highlight the need for greater transparency and accountability in automated moderation systems.

Hate speech itself is a socially complex phenomenon that extends beyond explicit insults or slurs. Social science research highlights that harmful language often manifests through implicit hostility, coded expressions, or context-dependent forms of aggression [8,9]. The interpretation of such language is shaped by cultural context, community norms, and historical power relations [10]. Consequently, automated hate speech detection systems must navigate a challenging landscape where linguistic signals of harm may be subtle, ambiguous, or contested. Errors in this setting can have meaningful societal consequences, including the suppression of legitimate speech, the amplification of harassment, or uneven moderation outcomes across demographic groups [7,11].

A key challenge in automated moderation arises from the opacity of modern machine learning models. Many hate speech detection systems rely on deep neural architectures whose internal decision processes are difficult to interpret or audit [1]. This lack of transparency complicates accountability in platform governance, as users and moderators often cannot determine why particular pieces of content were flagged or removed [12]. From both technical and applied system perspectives, explainability has therefore emerged as an important requirement for responsible deployment of automated decision systems [13,14]. Transparent decision-making processes are particularly important in moderation contexts where affected users may seek justification or contest moderation outcomes.

Within the broader field of explainable artificial intelligence (XAI), a range of methods have been proposed to improve the interpretability of machine learning models. Post-hoc explanation techniques such as LIME and SHAP attempt to approximate local feature importance in complex models [15,16], while gradient-based approaches highlight input tokens contributing to predictions [17]. Although these techniques can provide insights into model behavior, they often remain external to the model training process and may not reliably capture the internal reasoning mechanisms of the underlying system [18]. This limitation has motivated growing interest in approaches that incorporate interpretability signals directly during model training.

Rationale-based learning represents one such paradigm. In these approaches, models are supervised not only with task labels but also with human-provided rationales—text spans that explain why a particular prediction should hold [19]. Subsequent work has explored neural architectures capable of generating or selecting rationales that support predictions, enabling models to produce explanations alongside classification outputs [20]. Benchmark datasets such as ERASER have further advanced this area by standardizing evaluation protocols for explanation quality in natural language processing tasks [21]. In principle, integrating human rationales into training can encourage models to ground predictions in linguistically meaningful evidence aligned with human reasoning.

Despite these advances, the implications of rationale supervision for automated moderation systems remain underexplored. Moderation tasks differ from many standard NLP benchmarks in that prediction errors carry normative and societal consequences. For instance, false positives may result in the removal or suppression of legitimate speech, while false negatives may allow abusive content to remain visible and cause harm to targeted communities. Furthermore, moderation models may behave differently when processing identity-linked content or linguistically diverse inputs, raising concerns around disproportionate error patterns [7,11]. These risks are particularly salient in multilingual or code-switched communication environments, where linguistic variation may degrade moderation performance or obscure harmful intent [22].

Addressing these challenges requires evaluation frameworks that extend beyond aggregate performance metrics such as overall accuracy or macro F1. In high-stakes moderation contexts, it is equally important to examine how models behave across socially salient content slices and under conditions of linguistic variability. Such analyses can surface potential risk concentrations or moderation instabilities that may not be visible in aggregate evaluation.

In this work, we introduce Human-Aligned Rationale Learning (HARL), a training framework that incorporates human-annotated rationales directly into hate speech detection model optimization. HARL encourages models to ground harmfulness predictions in human-identified indicators of abuse while maintaining competitive classification performance. We evaluate the framework across multiple moderation datasets, examining both detection reliability and explanation quality through plausibility and partial faithfulness analyses.

Rather than introducing a new model architecture, HARL provides a controlled framework for integrating explanation signals into model training and studying their effects on moderation behavior. By keeping the underlying architecture fixed and varying only the explanation-guided supervision, the framework enables a clearer analysis of how explanation alignment influences prediction quality, explanation grounding, and moderation outcomes under different data conditions.

Beyond standard evaluation, we further conduct a deployment-oriented behavioral analysis to examine moderation behavior across socially salient and linguistically variable contexts. Using demographic annotations available in the HateXplain dataset, we analyze model performance across identity-linked content slices and examine variations in error patterns such as false positive and false negative rates. In parallel, we assess moderation stability under varying degrees of code-switching to understand how linguistic variability influences model performance and explanation behavior. Rather than proposing a new detection model, this study focuses on understanding how explanation-guided supervision affects the behavior of existing moderation systems.

Most existing work on automated hate speech moderation evaluates models primarily through aggregate performance metrics such as accuracy or Macro-F1 [23,24], while explanation methods are often applied only as post-hoc interpretability tools [15,16]. Although these approaches provide useful diagnostic insights, they rarely examine how explanation signals influence model behavior during training or how moderation systems behave under socially and linguistically meaningful conditions [25]. In contrast, this work adopts a behavioral evaluation perspective in which explanation signals are integrated directly into the training process and used to analyze moderation outcomes across identity-linked and linguistically variable inputs. This perspective positions explanation-guided learning as a practical mechanism for improving interpretability and enabling structured analysis of moderation system behavior in deployment settings.

Our contributions are as follows:

• We present HARL, an explanation-aware moderation framework that integrates human rationale supervision into model training through a controlled alignment objective, enabling systematic analysis of explanation-guided learning in hate speech detection.
• We introduce a unified evaluation protocol that jointly examines predictive performance, explanation plausibility, partial faithfulness, and runtime efficiency, providing a comprehensive view of explanation behavior in moderation systems.
• We conduct deployment-oriented behavioral analysis across identity-linked content and varying levels of code-switching, highlighting how explanation-guided supervision influences moderation outcomes under socially and linguistically variable conditions.
• We analyze trade-offs between attribution methods in terms of plausibility, faithfulness, and computational cost, and discuss their implications for real-time moderation, auditing, and decision-support workflows.

Importantly, this work builds on previous efforts [26] and examines these dimensions jointly, enabling a more comprehensive evaluation of moderation systems than studies that focus solely on predictive accuracy or explanation plausibility in isolation.

Positioning of this work.

This work does not propose a new rationale-learning paradigm, a fairness intervention, or a causal explanation method. Instead, it investigates whether explanation-guided supervision can serve as a practical design mechanism for automated moderation systems, and whether its effects can be systematically evaluated through a deployment-oriented behavioral analysis framework. Accordingly, HARL is positioned as an applied framework for explanation-aware training and moderation auditing, rather than as a standalone algorithmic contribution.

2. Related Work

Research on automated hate speech moderation sits at the intersection of social science, machine learning, and explainable artificial intelligence. Prior work has examined the societal dynamics of harmful language online, the development of automated detection systems, and emerging approaches to improve transparency and accountability in AI-driven moderation systems. This section reviews relevant literature across these domains.

2.1. Hate Speech and Moderation from a Social Science Perspective

Hate speech is widely understood as a socially embedded phenomenon shaped by cultural, political, and historical contexts. Social science research has shown that harmful language often extends beyond explicit slurs to include implicit hostility, coded language, and context-dependent expressions of exclusion or harassment [6,8,10]. Network-based studies further demonstrate that hate speech frequently spreads through community dynamics and coordinated interactions rather than isolated statements [9].

Scholars have also highlighted the governance challenges faced by online platforms attempting to regulate harmful content. Platforms function as private regulators of digital speech, shaping the visibility of content and influencing online discourse [5]. Human moderation systems themselves are labor-intensive and expose moderators to psychological harm due to repeated exposure to disturbing content [4]. As a result, many platforms increasingly rely on automated moderation technologies, raising concerns regarding transparency, accountability, and potential impacts on marginalized communities [3].

2.2. Machine Learning Approaches to Hate Speech Detection

Machine learning techniques have become central to automated hate speech detection due to the scale of online content. Early approaches relied on lexical features and traditional classifiers such as logistic regression or support vector machines [1,2]. More recent work has leveraged deep neural architectures, including convolutional neural networks, recurrent neural networks, and transformer-based language models capable of capturing contextual semantic representations [27,28].

The development of benchmark datasets such as HateXplain and other annotated moderation corpora has enabled systematic evaluation of these systems [23,29]. However, despite improvements in predictive accuracy, several studies have identified problematic behaviors in hate speech classifiers, including sensitivity to identity terms and uneven error patterns across demographic groups [7,11]. These findings highlight the importance of evaluating moderation systems beyond aggregate performance metrics.

2.3. Explainable AI for Hate Speech Detection

The opacity of modern machine learning models has motivated growing interest in explainable artificial intelligence (XAI) for content moderation systems. XAI techniques aim to provide interpretable insights into model predictions, enabling users and system designers to better understand the reasoning behind automated decisions [13,14].

Post-hoc explanation approaches such as LIME and SHAP estimate feature importance for individual predictions [15,16], while gradient-based attribution methods highlight tokens that contribute to model outputs [30]. These techniques have been applied to hate speech detection models to reveal linguistic features influencing classification decisions [23,31]. However, post-hoc explanations may not faithfully reflect the internal reasoning of complex models, leading to concerns regarding explanation reliability [18].

Recent research has explored integrating explainability more directly into hate speech detection models. For example, recent frameworks combine deep learning architectures with explanation techniques to provide interpretable predictions [32]. More recent work also investigates the use of large language models to generate interpretable rationales for hate speech detection, enabling classifiers to maintain performance while providing explanation signals [33]. Other approaches use structured reasoning or step-by-step explanations to improve both detection accuracy and explanation quality [34]. Recently, works have utilised LLMs as a stand-in for traditional explainer methods as well[35]. These developments highlight a growing interest in explanation-aware moderation systems[25].

2.4. Rationale-Supervised Learning and Human-in-the-Loop Approaches

Rationale-supervised learning represents an approach to explainability that incorporates human explanations directly during model training. In these methods, models are supervised not only with task labels but also with human-annotated rationales indicating which parts of the input text justify a prediction [19]. Neural architectures capable of selecting or generating rationales have been proposed to produce explanations alongside classification outputs, encouraging models to ground predictions in linguistically meaningful evidence [20].

Datasets such as ERASER and HateXplain provide human-annotated explanations that enable systematic evaluation of rationale-based models [21,29]. Recent work has explored how such rationales can improve both interpretability and predictive performance in hate speech detection systems by aligning model reasoning with human judgments.

Rationale supervision also connects closely with the broader paradigm of human-in-the-loop machine learning. Human-in-the-loop approaches incorporate human knowledge and feedback into the learning process to improve system reliability, transparency, and alignment with human expectations [36,37]. In moderation contexts, human rationales act as structured signals capturing human reasoning about harmful language. Integrating these signals during model training can guide model attention toward relevant linguistic cues while maintaining human oversight in the design and evaluation of automated moderation systems.

2.5. Evaluation of Explanations and Fairness in Moderation Systems

Evaluating explanations remains an open challenge in explainable AI research. Explanation quality is typically assessed along two complementary dimensions: plausibility and faithfulness. Plausibility measures whether explanations align with human reasoning, often through token-level agreement with human annotations [21]. Faithfulness evaluates whether highlighted features genuinely influence model predictions and is often tested through perturbation-based methods such as prediction-drop or feature removal experiments [38].

Recent work has further examined the relationship between these dimensions, demonstrating that plausibility and faithfulness may not always align and may require complementary evaluation methods [21,39].

In parallel, research on fairness in hate speech detection has shown that models may disproportionately associate identity-related terms with toxicity or exhibit uneven error patterns across demographic groups [7,11]. From an evaluation perspective, these concerns highlight the need for frameworks that examine both explanation quality and moderation behavior across socially relevant contexts.

Building on these insights, the present work investigates how rationale-supervised learning influences moderation reliability, explanation grounding, and moderation behavior across identity-linked and linguistically diverse content.

2.6. AI Ethics and Responsible Moderation Systems

The use of automated moderation systems also raises broader questions within AI ethics. As algorithmic systems increasingly participate in the governance of online speech, concerns have emerged regarding transparency, accountability, and fairness in automated decision-making [40,41]. Moderation algorithms can influence which voices are amplified or suppressed in online spaces, making their behavior socially consequential [5].

Recent work in responsible AI emphasizes the importance of interpretable and auditable models in high-stakes decision contexts [12,18]. In moderation settings, explainability can support transparency by helping moderators and users understand the reasoning behind automated decisions. At the same time, studies have shown that hate speech detection systems may exhibit uneven behavior across identity-linked language [7,11]. These concerns motivate evaluation frameworks that examine both explanation quality and moderation behavior across socially relevant contexts.

Research gap

While prior work has made significant progress in rationale-based learning, explanation evaluation, and bias analysis in hate speech detection, these aspects are typically studied in isolation. Existing approaches either focus on improving predictive performance, generating explanations, or analyzing model behavior across specific dimensions such as identity terms. However, there remains a lack of unified frameworks that (i) integrate explanation supervision directly into model training, (ii) evaluate explanation quality using complementary plausibility and faithfulness metrics, and (iii) examine moderation behavior under socially and linguistically variable conditions in a deployment-relevant setting. This fragmentation limits our understanding of how explanation-guided learning affects moderation system behavior beyond standard benchmarks.

Comparison to Relevant Literature

To situate our work within the existing literature, Table 1 compares several representative studies across key aspects of explainable hate speech moderation. Prior work has made important contributions in individual areas, including rationale-based learning [19,20], explanation evaluation [21], and analysis of identity-linked model behavior [7,11]. However, these dimensions are typically explored independently.

In contrast, this work adopts an integrated perspective by combining explanation-guided training, multi-metric explanation evaluation, and deployment-oriented behavioral analysis within a single experimental framework. The goal is not to introduce a new modeling paradigm, but to provide a structured approach for understanding how explanation supervision influences moderation system behavior across different evaluation dimensions.

Unlike prior work that focuses on individual components of explanation or bias analysis, this study emphasizes the interaction between explanation supervision, explanation quality, and moderation behavior under deployment-relevant conditions.

3. Methodology

This section describes the datasets, rationale-supervised moderation pipeline, explanation analysis framework, and evaluation procedures used in this study. Our objective is to examine how integrating human rationales into model training affects both explanation quality and moderation behavior across socially relevant and linguistically variable contexts. This design allows us to analyze explanation behavior, predictive performance, and moderation reliability within a single controlled experimental setting. Accordingly, the methodological design focuses on controlled analysis of explanation behavior and moderation outcomes under consistent modeling conditions.

3.1. Datasets

We conduct experiments on two datasets that represent complementary moderation settings: HateXplain and BullySent. These datasets were selected not to maximize dataset diversity but to represent two complementary moderation regimes: a high-resource rationale-supervised setting and a lower-resource linguistically variable setting. This contrast allows us to examine how explanation supervision behaves under different annotation and linguistic conditions.

HateXplain

HateXplain [29] is a benchmark dataset designed specifically for explainable hate speech detection. The dataset contains social media posts annotated with hate speech labels as well as human-provided rationales highlighting the tokens that justify each decision. Each example is annotated by multiple annotators who independently identify both the class label and the corresponding rationale spans. The final dataset contains three classes: hate speech, offensive language, and normal.

Importantly, HateXplain also includes identity annotations indicating which demographic groups are referenced within each post. These annotations enable analysis of moderation behavior across identity-linked content, allowing evaluation of whether models exhibit differential error patterns when processing language referencing particular demographic groups.

Because it provides both class labels and explanation annotations, HateXplain represents a high-resource setting for explanation-aware moderation models.

BullySent

The BullySent[42] dataset contains annotated instances of abusive and bullying language collected from social media platforms. Unlike HateXplain, this dataset does not include human rationale annotations. Instead, it provides binary labels indicating whether a post contains bullying or abusive language.

This dataset therefore represents a lower-resource setting in which explanation supervision is unavailable. Evaluating models across both datasets allows us to examine whether rationale-aligned training learned from explanation-rich data generalizes to moderation contexts where explicit rationales are not provided.

Dataset statistics are summarized in Table 2.

3.2. Rationale-Supervised Moderation Training

To investigate the effect of rationale supervision, we train transformer-based hate speech detection models using RoBERTa [43] and XLM-R [44]. Training follows a multi-objective formulation that combines standard classification loss with a rationale alignment objective. While our experiments use a single backbone architecture per dataset, this design isolates the effect of explanation-guided supervision from architectural variation, allowing clearer comparison of attribution mechanisms. Prior work suggests that attribution-based explanation behavior is often consistent across transformer backbones, indicating that the observed trends are unlikely to be strongly architecture-specific. Figure 1 shows an overview of the pipeline that integrates rationale-supervised feedback into the training pipeline for both datasets.

Classification Loss

Let x denote an input post and $y\in \{0,1\}$ the ground truth label for the target class. The model produces a prediction probability $\widehat{y}=f\left(x\right)$. For binary classification tasks we use binary cross-entropy (BCE) loss:

$${\mathcal{L}}_{cls}=-\left[ylog\left(\widehat{y}\right)+(1-y)log(1-\widehat{y})\right]$$

(1)

For multi-class settings the loss is computed using standard cross-entropy over class probabilities.

Explanation Alignment Loss

For examples in HateXplain, annotators provide rationale masks identifying tokens that support the moderation decision. Let $r_i$ denote the human rationale mask for token i where:

$$r_i=\left\{\begin{array}{cc}1\hfill & \mathrm{if}\mathrm{token}i\mathrm{is}\mathrm{part}\mathrm{of}\mathrm{the}\mathrm{rationale}\hfill \\ 0\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

Given an attribution score $s_i$ produced by an explanation method for token i, we define an explanation alignment loss that encourages agreement between model attributions and human rationales:

$${\mathcal{L}}_{exp}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{(s_i-r_i)}^2$$

(2)

To ensure comparability across inputs, attribution scores are rescaled to a common range using min-max normalization within each instance prior to computing alignment loss. This loss penalizes discrepancies between tokens identified by the model as important and those highlighted by human annotators. We adopt a squared alignment objective that treats attribution scores and human rationale masks as continuous alignment signals over the input space. This formulation can be interpreted as minimizing the Euclidean distance between model-attributed importance and human-identified evidence, encouraging the model to allocate importance mass to tokens that align with human reasoning.

Compared to distributional objectives such as KL-divergence, this formulation avoids normalization constraints on attribution scores and provides a stable optimization signal across varying attribution scales. Importantly, the objective does not enforce exact correspondence between model attributions and human rationales, but instead provides a soft alignment signal that biases the model toward human-relevant features while preserving predictive flexibility. This design is consistent with our goal of studying the behavioral impact of explanation supervision rather than introducing a new alignment objective.

Furthermore, human-provided rationales may themselves be incomplete or subjective; therefore, the alignment objective is intended as a guiding signal rather than a ground-truth constraint, allowing the model to deviate where necessary for predictive performance.

Combined Training Objective

The final training objective combines classification and explanation alignment losses:

$$\mathcal{L}={\mathcal{L}}_{cls}+\lambda {\mathcal{L}}_{exp}$$

(3)

The hyperparameter $\lambda$ controls the strength of explanation supervision and is selected via validation-based tuning to balance predictive performance and explanation alignment. Lower values prioritize classification accuracy, while higher values enforce stronger alignment with human rationales, potentially constraining model flexibility. This trade-off is explicitly examined in our experimental analysis. When $\lambda =0$, the model reduces to a standard classification model without explanation supervision.

3.3. Explanation Extraction and Selection Strategies

In our experiments, we employ multiple attribution methods, including Integrated Gradients (IG), attention-based attribution, and perturbation-based methods such as LIME, to generate token-level importance scores. These methods are selected to represent diverse explanation families with different underlying assumptions. Rather than treating them as competing algorithms, we use them as analytical tools to examine how explanation-guided supervision influences moderation behavior. Each method produces token-level attribution scores $S=\{s_1,s_2,...,s_n\}$ indicating the relative importance of each token for the model prediction. We do not automatically assume that attention always constitutes a faithful explanation, but rather use it as a tractable proxy for alignment with human rationales.

Top-K Strategy

The Top-K strategy selects the K tokens with the highest attribution scores:

$$E_{TopK}=\{i\mid s_i\in \mathrm{TopK}(S,K)\}$$

(4)

This approach ensures that explanations have fixed length across examples.

Threshold Strategy

Alternatively, tokens whose attribution score exceeds a threshold $\tau$ are selected:

$$E_{\tau }=\{i\mid s_i\ge \tau \}$$

(5)

This method allows explanation lengths to vary depending on attribution strength.

3.4. Explanation Quality Evaluation

Explanation quality is evaluated along two complementary dimensions: plausibility and faithfulness. Importantly, we do not assume that attribution methods provide fully faithful explanations of the model’s internal reasoning process. Instead, attribution scores are treated as analytical proxies that provide partial and complementary views of model behavior. Accordingly, we evaluate explanations along both plausibility and perturbation-based faithfulness dimensions to capture different aspects of explanation quality.

Plausibility

Plausibility measures agreement between model explanations and human rationales. Let E denote the set of tokens selected by the explanation method and R denote the set of human rationale tokens.

Token-level F1 is computed as:

$$F1={\displaystyle \frac{2|E\cap R|}{\left|E\right|+\left|R\right|}}$$

(6)

Intersection-over-Union is defined as:

$$IoU={\displaystyle \frac{|E\cap R|}{|E\cup R|}}$$

(7)

Higher values indicate stronger agreement between model explanations and human annotations.

Faithfulness

Faithfulness measures whether explanation tokens genuinely influence model predictions. We estimate faithfulness using prediction-drop analysis [38]. Given an explanation token set E, the prediction drop is defined as:

$$\Delta =f_y\left(x\right)-f_y(x\setminus E)$$

(8)

where $f_y\left(x\right)$ denotes the predicted probability of the target class. Larger values of $\Delta$ indicate that removing explanation tokens significantly alters the model prediction.

3.5. Identity and Linguistic Moderation Analysis

To examine moderation behavior across socially relevant and linguistically variable contexts, we conduct two complementary analyses using the HateXplain and BullySent datasets. HateXplain enables evaluation across identity-linked content through demographic annotations, while BullySent allows examination of linguistic stability under varying degrees of code-switching.

Identity-Linked Analysis (HateXplain)

HateXplain provides annotations indicating whether a post references specific demographic groups. We first partition the dataset into two subsets based on identity presence: posts that reference at least one identity group and posts without identity references. Let $I\left(x\right)$ denote an indicator function that equals 1 if input x contains an identity reference and 0 otherwise.

We compute classification performance separately for identity-present and identity-absent subsets. In addition, we perform a per-identity analysis by evaluating performance for each identity group g present in the dataset.

For each group g, we compute false positive rate (FPR), false negative rate (FNR), and F1 score. False positive rate is defined as:

$$FP{R}_g={\displaystyle \frac{F{P}_g}{F{P}_g+T{N}_g}}$$

(9)

and false negative rate as:

$$FN{R}_g={\displaystyle \frac{F{N}_g}{F{N}_g+T{P}_g}}$$

(10)

where $T{P}_g$, $T{N}_g$, $F{P}_g$, and $F{N}_g$ denote the number of true positives, true negatives, false positives, and false negatives respectively for examples referencing identity group g.

We also compute the F1 score for each subset as:

$$F1={\displaystyle \frac{2TP}{2TP+FP+FN}}$$

(11)

Comparing these metrics across identity-present and identity-specific subsets allows us to identify disparities in moderation outcomes when models process language referencing different demographic groups.

Linguistic Stability Analysis (BullySent)

To analyze linguistic stability, we evaluate moderation performance under varying levels of code-switching using the BullySent dataset. Code-switching refers to the presence of multiple languages within a single input text.

Let $C\left(x\right)$ denote the proportion of tokens in input x belonging to a secondary language. The proportion $C\left(x\right)$ is computed using token-level language identification, where each token is assigned a language label by a pre-trained language detection model, and the fraction of tokens belonging to a secondary language is then calculated. Based on this value, we categorize examples into code-switching bins representing different linguistic conditions:

• Low code-switching: $C\left(x\right)<0.1$
• Medium code-switching: $0.1\le C\left(x\right)<0.3$
• High code-switching: $C\left(x\right)\ge 0.3$

For each bin c, we compute classification performance using the same evaluation metrics defined above:

$$F{1}_c=F1\left(\{x\mid C\left(x\right)\in c\}\right)$$

(12)

as well as corresponding false positive and false negative rates.

These analyses are designed to approximate deployment-relevant scenarios in which moderation systems must operate across diverse user populations and heterogeneous linguistic inputs, providing insights into how explanation-guided training may influence real-world moderation reliability and decision-support workflows.

Statistical Considerations

To ensure robustness of reported results, experiments are conducted across multiple random seeds, and performance metrics are reported as mean values with standard deviation. Statistical significance between baseline and explanation-guided models is assessed using paired t-tests at a significance level of $p<0.05$. These procedures are intended to ensure that observed differences reflect consistent behavioral trends rather than random variation.

4. Results and Discussion

This section reports results across the three main dimensions considered in this study: classification performance, explanation quality, and runtime efficiency. We compare attribution-guided regularization methods under both a high-resource setting (HateXplain; HX) and a lower-resource, linguistically noisier setting (BullySent; BS). Table 3 summarizes the full set of results, including Accuracy and Macro-F1 for classification, IoU and Token-F1 for explanation plausibility, Prediction Drop for partial faithfulness, and training/inference throughput for computational efficiency. Together, these results allow us to examine how explanation-guided supervision affects predictive quality, explanation grounding, and deployment-relevant trade-offs across different data regimes. The findings are therefore interpreted primarily in terms of moderation behavior, explanation quality, and operational suitability rather than purely predictive gains. The broader goal is to provide empirical insight into how explanation-aligned learning influences moderation systems in socially sensitive and deployment-relevant settings.

4.1. Classification Performance and Regularization Sensitivity

We first examine whether explanation-guided regularization improves predictive performance and how this effect depends on the strength of rationale supervision. As described in Section 3, the explanation alignment term is weighted by the regularization coefficient $\lambda$. We tune $\lambda$ separately for each dataset by selecting the value that maximizes validation Macro-F1. Figure 2 shows the resulting validation curves for HateXplain and BullySent.

A clear difference emerges between the two resource regimes. In HateXplain, performance improves steadily as rationale supervision becomes stronger, reaching its peak at $\lambda =5$. This suggests that the richer supervision available in HateXplain, including multi-annotator rationales and more stable explanation signals, allows the model to benefit from relatively strong regularization. In this setting, explanation alignment appears to reinforce rather than disrupt the classification objective, indicating that human rationales provide a useful supervisory signal during training.

In contrast, BullySent exhibits a much lower optimal value, with validation performance peaking at $\lambda =0.05$ and declining as the regularization weight increases further. This behavior is consistent with the lower-resource nature of the dataset. Compared with HateXplain, BullySent is smaller, more linguistically variable, and supported by weaker explanation supervision. In such conditions, heavy regularization can over-constrain the model toward sparse or unstable explanation cues, harming generalization. As a result, BullySent benefits from a lighter form of rationale influence, where explanation alignment acts as an auxiliary signal rather than a dominant supervisory force. The $\lambda$ sweep therefore functions as an implicit ablation of explanation influence, illustrating how varying levels of rationale supervision affect predictive performance across different data regimes.

Using these tuned values, Table 3 shows that explanation-guided training generally improves or preserves classification performance across both datasets relative to the non-regularized baseline. On HateXplain, all attribution-guided variants match or exceed the baseline Macro-F1 of 0.82, with LIME and Attention achieving the highest F1 of 0.84 and improved accuracy of 0.80. These gains are meaningful because they indicate that rationale-guided training can improve predictive performance in a high-resource setting while also supporting explanation-grounded moderation decisions.

On BullySent, the improvements are similarly consistent but reflect a different regime. The non-regularized baseline achieves 0.83 Macro-F1, while LIME and Integrated Gradients increase this to 0.84 and both Grad×Input and Attention reach 0.85. Attention also achieves the highest accuracy at 0.86. Although the absolute gains are modest, they remain notable given the noisier and lower-resource nature of the dataset. In particular, the fact that multiple regularized variants outperform the baseline suggests that explanation-aware supervision can remain beneficial even when rationale quality is weaker, provided that the regularization strength is carefully controlled.

Taken together, these results show that explanation-guided supervision can improve classification quality across both high-resource and low-resource moderation settings, but the appropriate degree of supervision depends strongly on annotation quality and dataset characteristics. HateXplain benefits from stronger alignment with human rationales, whereas BullySent benefits from a lighter-touch regularization regime. From a deployment perspective, this suggests that rationale supervision should not be treated as a fixed training ingredient, but rather as a dataset-dependent signal whose influence must be calibrated to the reliability of available explanations.

4.2. Explanation Quality and Attribution Behaviour

We next analyze explanation quality across attribution methods using both plausibility and faithfulness-oriented metrics. Table 3 summarizes the primary quantitative results, while Figure 4 and provide deeper insight into how explanations change under different extraction strategies and across attribution methods.

Explanation quality is evaluated using three complementary metrics: IoU and Token-F1, which measure plausibility relative to human rationales, and Prediction Drop, which serves as a perturbation-based proxy for partial faithfulness by quantifying how model confidence changes when highlighted tokens are removed. IoU captures the overlap between predicted and human rationale spans, while Token-F1 balances precision and recall at the token level. Prediction Drop does not provide a definitive causal account of explanation validity, but it offers a useful indication of whether highlighted tokens materially affect model confidence.

Across both datasets, the attribution methods exhibit clear behavioral differences. On HateXplain (HX), which contains richer multi-annotator rationales, internal attribution mechanisms such as Attention consistently achieve the highest plausibility scores. As shown in Table 3, Attention reaches the highest Token-F1 (0.60) and IoU (0.38) under threshold extraction, suggesting that its highlighted tokens align most closely with human rationale annotations. Integrated Gradients (IG) achieves the highest Prediction Drop, with values reaching 0.74, indicating stronger dependence of model confidence on the tokens it highlights. Grad×Input (GXI) provides a balanced trade-off, achieving moderately strong plausibility and faithfulness-oriented performance while remaining computationally efficient.

The lower-resource BullySent (BS) dataset exhibits a different pattern. Explanation metrics are generally lower due to annotation sparsity, code-switching, and weaker rationale supervision. Under these conditions, perturbation-based explanations such as LIME become more competitive for plausibility, achieving the highest Token-F1 (0.45) and IoU (0.27) under Top-K extraction. This suggests that perturbation-based explanations may sometimes better tolerate noisy or weak rationale supervision. Nevertheless, IG again produces the largest Prediction Drop, while GXI offers a practical compromise between explanation quality and computational cost.

4.2.1. Agreement Between Attribution Methods

Beyond metric values, it is important to understand how different XAI methods interpret the same examples. Figure 3 shows the Kendall’s $\tau$ agreement heatmap measuring rank correlation between token-level attribution scores.

The heatmap reveals clear clustering among gradient-based approaches. Attention and Grad×Input exhibit the highest agreement ($\tau \approx 0.45$), indicating that both methods tend to highlight similar tokens despite using different attribution formulations. Integrated Gradients shows moderate agreement with both methods ($\tau \approx 0.25$–0.30), suggesting that these gradient-based approaches capture related internal importance signals.

In contrast, LIME shows near-zero or slightly negative correlation with the gradient-based methods. This divergence arises from LIME’s perturbation-based mechanism, which evaluates token importance by sampling modified inputs rather than inspecting model gradients directly. As a result, LIME explanations often emphasize different linguistic cues, providing complementary interpretability signals rather than redundant ones. From a deployment perspective, this suggests that explanation method choice is not interchangeable and should instead be matched to the intended moderation workflow.

4.2.2. Effect of Explanation Extraction Strategy

Figure 4 provides a detailed view of how explanation metrics change as Top-K and threshold parameters vary. Several consistent trends emerge. First, threshold-based extraction generally produces stronger plausibility and faithfulness in high-resource settings. On HateXplain, IoU and Token-F1 curves show clear peaks for Attention and IG when moderate thresholds are applied, indicating that allowing explanation length to adapt to attribution magnitude improves alignment with human rationales.

Second, Top-K extraction produces more stable curves in the lower-resource BullySent dataset. Fixed-length selection prevents thresholding from producing overly sparse explanations when attribution scores are noisy. This effect is particularly visible for LIME, whose Top-K curves outperform threshold extraction in several plausibility metrics.

As shown in the Prediction Drop plots, IG consistently produces the highest drop values across both datasets, suggesting that gradient path integration tends to highlight tokens on which model confidence is comparatively more dependent. GXI and Attention follow similar but slightly weaker trends, while LIME produces comparatively lower drops, reflecting the indirect relationship between perturbation-based explanations and model gradients.

Figure 4. Variation of explanation metrics across extraction strategies. Top rows correspond to HateXplain (HX), while bottom rows correspond to BullySent (BS). Metrics include IoU and Token-F1 (plausibility) and Prediction Drop (faithfulness). Threshold-based extraction generally improves plausibility in high-resource settings, while Top-K selection produces more stable behavior under low-resource and noisy rationale conditions.

Figure 4. Variation of explanation metrics across extraction strategies. Top rows correspond to HateXplain (HX), while bottom rows correspond to BullySent (BS). Metrics include IoU and Token-F1 (plausibility) and Prediction Drop (faithfulness). Threshold-based extraction generally improves plausibility in high-resource settings, while Top-K selection produces more stable behavior under low-resource and noisy rationale conditions.

4.2.3. Qualitative Attribution Patterns and Multi-Objective Trade-offs

Figure 3b illustrates qualitative attribution examples on two code-switched hate speech instances. Gradient-based methods (Attention, GXI, and IG) consistently highlight semantically meaningful tokens associated with abusive intent, often capturing subword fragments and compound slurs that arise from multilingual or morphologically complex language.

LIME explanations differ in two ways. First, they operate at a coarser word-level granularity rather than subword tokenization. Second, perturbation sampling occasionally fails to isolate components of compound expressions, causing parts of abusive phrases to remain unhighlighted. However, this word-level granularity can also make LIME explanations easier for human moderators to interpret, particularly when explanations are surfaced in moderation dashboards or audit settings.

To synthesize trade-offs between explanation quality and computational efficiency, we perform a multi-objective Pareto analysis. A configuration is considered Pareto-optimal if no other method performs strictly better across plausibility (Token-F1), Prediction Drop, and runtime efficiency simultaneously.

Table 4 summarizes these Pareto-optimal configurations. On HateXplain, Attention and GXI offer the strongest balance between explanation quality and efficiency, making them suitable candidates for large-scale moderation pipelines. Integrated Gradients provides the strongest faithfulness-oriented signal but incurs higher computational cost, making it better suited for auditing and diagnostic workflows.

On BullySent, GXI and Attention remain efficient choices for deployment, while LIME becomes competitive for plausibility under noisy rationale conditions. These results demonstrate that the most suitable attribution method depends on deployment requirements: gradient-based approaches are better aligned with real-time moderation settings, while perturbation-based explanations may be more useful for offline analysis or moderator-facing inspection.

4.3. Identity and Linguistic Stability Analysis

Beyond overall classification and explanation performance, it is important to examine how moderation models behave across socially sensitive contexts and linguistically diverse inputs. We therefore analyze two additional dimensions of model behavior: (1) identity-linked moderation outcomes using the HateXplain dataset and (2) linguistic stability under varying degrees of code-switching using BullySent.

These analyses provide insight into how explanation-guided training affects moderation behavior and reliability across both high-resource and low-resource environments. HateXplain represents a high-resource dataset with rich multi-annotator rationales and explicit identity annotations, allowing fine-grained evaluation of moderation outcomes across identity groups. In contrast, BullySent is a smaller, lower-resource dataset characterized by substantial code-switching and noisier annotations, making it suitable for evaluating classification stability on linguistic variation.

Since Attention-based attribution achieved the strongest overall balance of plausibility, predictive performance, and efficiency in earlier experiments, we compare the Attention model against a baseline model without rationale supervision when analyzing identity-linked and code-switching effects.

4.3.1. Identity-Linked Moderation Behaviour (HateXplain)

We first analyze moderation outcomes when identity terms are present in text. Table 5 compares classification performance for samples containing identity references versus those without identity references.

When identity references are present, the attention-based model achieves higher Macro-F1 and accuracy than the baseline model. In particular, the baseline model exhibits a higher false positive rate (FPR), suggesting that models without rationale supervision may be more prone to over-predict harmful content when identity-related language appears. By contrast, the attention-based model reduces this tendency while maintaining broadly similar false negative rates, indicating improved reliability for identity-linked content.

We further examine moderation behaviour across grouped identity categories in Table 6.

The per-group analysis in Table 6 reveals meaningful variation across identity categories. For some groups, particularly content referencing Muslim identities, the attention-based model improves Macro-F1 relative to the baseline while also reducing false positive rates. This pattern is consistent with the interpretation that explanation-guided training can encourage the model to attend more to contextual cues than to identity terms alone.

However, disparities across identity groups remain visible. Some categories continue to exhibit elevated false positive rates, and the improvements are not uniform across all groups. These findings should therefore be interpreted as diagnostic evidence of behavioral asymmetry rather than as a fairness guarantee. Nevertheless, the overall pattern suggests that rationale-aligned training can improve moderation stability for at least some forms of identity-linked content.

4.3.2. Linguistic Stability Under Code-Switching (BullySent)

We next analyze classification stability to linguistic variation using the BullySent dataset, which contains substantial code-switching between languages. Table 7 reports performance across bins representing different levels of code-switching in the input text.

Across most code-switching levels, the attention-based model achieves equal or higher Macro-F1 than the baseline model. The improvements are particularly visible in the moderate code-switching ranges (10–50%), where explanation-guided supervision appears to help the model maintain more stable predictions despite linguistic variation. At the highest code-switching bucket, however, the baseline remains competitive or stronger on some metrics, suggesting that the benefits of rationale alignment are not uniform across all linguistic conditions and may become less predictable in very small or highly variable slices.

Finally, Table 8 compares high and low code-switching conditions using bootstrap confidence intervals. Although performance patterns differ across linguistic conditions, the attention-based model generally maintains comparable or slightly improved Macro-F1 relative to the baseline. This suggests that explanation-guided training can contribute to stability under linguistic variability, but the effect should be interpreted as a diagnostic trend rather than a universal stability guarantee.

Across both analyses, a consistent pattern emerges: explanation-guided supervision generally improves or preserves moderation stability, particularly in contexts where identity-linked language or linguistic variation might otherwise introduce prediction instability. These analyses are intended as behavioral diagnostics rather than definitive fairness measurements.

5. Limitations

While the proposed framework provides useful insight into explanation-guided moderation systems, several limitations should be acknowledged. First, our experiments evaluate explanation-guided regularization using a single backbone model per dataset. This choice was deliberate: the primary objective of this study is to compare attribution methods and their interaction with human rationale supervision rather than to benchmark classification architectures. Using a consistent model architecture reduces confounding effects and helps isolate the behavioral impact of explanation-guided supervision. Nevertheless, broader evaluation across additional transformer backbones would strengthen the generality of the findings.

Second, our evaluation relies on two datasets, HateXplain and BullySent, selected because they provide complementary moderation conditions and, in the case of HateXplain, token-level rationale annotations that are essential for studying explanation alignment. However, the limited availability of rationale-annotated moderation datasets remains a broader challenge for the field. Constructing large-scale explanation-rich moderation datasets is costly and time-consuming, and the conclusions reported here should therefore be interpreted within the scope of the available benchmarks rather than as exhaustive evidence across all moderation settings.

Third, our explanation evaluation remains limited to plausibility and perturbation-based faithfulness proxies. Although these metrics provide useful complementary signals, they do not constitute definitive causal validation of explanations. We also do not conduct counterfactual explanation analysis or direct human-subject evaluation of explanation usability for moderators, both of which would strengthen claims regarding practical interpretability in moderation workflows.

Fourth, the identity-linked and code-switching analyses presented in this work are diagnostic in nature. They are intended to surface behavioral asymmetries and stability trends, not to establish compliance with formal group fairness criteria or to provide a fairness intervention. The scope of these analyses is further constrained by the identity categories, language mixtures, and sample sizes present in the available datasets.

Finally, although the paper discusses deployment-oriented implications for moderation systems, the proposed workflow is evaluated analytically rather than through live deployment or user studies with moderators. As a result, the operational value of different explanation methods for decision support, contestability, or moderation efficiency remains an important direction for future applied evaluation.

Despite these limitations, the results provide consistent evidence that integrating human rationales into model training can improve explanation alignment, preserve or improve predictive performance, and support more structured behavioral analysis of moderation systems. The findings therefore highlight the value of expanding rationale-annotated datasets and developing more systematic evaluation frameworks for explanation-aware moderation.

6. Conclusions and Future Work

This work presents Human-Aligned Rationale Learning (HARL) as an explanation-aware framework for hate speech moderation that integrates attribution-based alignment signals directly into model training. Rather than proposing a new rationale-learning paradigm, the framework is designed to examine how explanation-guided supervision influences predictive performance, explanation quality, and moderation behavior under deployment-relevant conditions.

Our results highlight several main findings. First, explanation-guided training generally improves or preserves predictive performance while also improving explanation plausibility across datasets with different resource characteristics. Second, attribution methods exhibit distinct trade-offs: Attention provides the strongest alignment with human rationales, while gradient-based approaches such as Integrated Gradients produce the strongest perturbation-based faithfulness signals at higher computational cost. Third, the effectiveness of explanation supervision depends on dataset characteristics. High-resource datasets such as HateXplain benefit from stronger rationale regularization, whereas lower-resource settings such as BullySent benefit from more conservative regularization because of noisier annotations and greater linguistic variability.

Beyond aggregate predictive performance, the results also show that explanation-guided training can support more informative behavioral analysis of moderation systems. Identity-linked evaluation reveals non-uniform moderation outcomes across demographic references, while code-switching experiments show that rationale-aligned training can preserve or improve stability across linguistically variable inputs in many settings. These findings should be interpreted as diagnostic evidence about moderation behavior rather than as guarantees of fairness or universal stability. Nevertheless, they suggest that aligning model-highlighted evidence with human rationales can improve interpretability while also supporting more reliable inspection of moderation decisions.

Several directions for future research follow from this work. A key priority is the development of additional rationale-annotated datasets across languages, cultures, and moderation domains, as such resources are essential for advancing explanation-aware and human-in-the-loop moderation research. Future work may also explore interactive annotation pipelines in which moderators iteratively refine explanations, allowing models to adapt to evolving linguistic patterns and community norms. Extending the framework to multilingual and multimodal moderation settings is another important avenue, particularly as real-world systems increasingly operate across diverse platforms and media types. In addition, future studies should examine counterfactual and causal explanation evaluation, explanation usability for moderators, and uncertainty-aware explanation mechanisms for decision support.

Overall, this work argues for viewing rationale-aligned learning not primarily as a new algorithmic contribution, but as a practical framework for jointly studying explanation grounding, predictive behavior, and deployment-oriented trade-offs in automated moderation systems.