3.1. Datasets Explanation

Digital copies of several legal documents pertinent to the education sector, such as acts, rules, circulars, and institutional policies, were included in the educational legal instruments dataset ³⁰. The materials were obtained from open-access repositories, university legal archives, and official government portals to guarantee their legitimacy and coverage. Every document underwent preprocessing, which included tokenization for model input, conversion to plain text, and removal of superfluous formatting. Because of the dataset’s varied vocabulary, wording, and structural variances, EduLegal-DL was able to acquire strong drafting patterns and context-aware compliance mapping across several educational legal areas.

Educational policy compliance dataset ³¹ consisted of annotated policy documents highlighting compliance and non-compliance segments. It included university policies, administrative guidelines, and accreditation requirements, annotated by legal experts to indicate whether specific clauses aligned with regulatory frameworks. The dataset captured multi-label compliance indicators such as academic governance, student rights, staff obligations, and institutional accountability. Preprocessing involved normalization, clause segmentation, and tagging for training the compliance-testing module. This dataset was essential for validating the generative drafts produced by EduLegal-DL, ensuring they adhered to both legal and institutional regulatory standards. The summary of these datasets is given in Table 2.

3.2. Data Collection and Preprocessing

The first step of the EduLegal-DL framework is data collection, where diverse educational legal corpora are gathered to ensure comprehensive model training. These corpora include laws, policies, treaties, and institutional regulations obtained from reliable sources such as government portals, university archives, and open repositories. Collecting data from multiple jurisdictions enhances the diversity of legal language, thereby improving the model’s ability to generalize across different contexts of education law.

Once collected, the raw text undergoes preprocessing, a crucial step to convert unstructured legal content into a standardized form suitable for deep learning models. The preprocessing pipeline consists of the following phases, as in Table 3.

The preprocessing equation parameters determine the structural components of the legal text and their mappings. In tokenization, \({T}=\{{t}_1,t_2,\dots,{t}_{{n}}\}\) is a tokenized stream of a document such that each token \(t_i\) is from the vocabulary set \(V\) and \(n\) represents the number of tokens. In clause segmentation, a clause is defined as \({C}_{{j}}=\{{t}_{{a}},{t}_{{a}+1},\dots,{t}_{{b}}\}\), where \({C}_{{j}}\) is the \(j\)-th clause of the document, and is a continuous token subset from position \(a\) to \(b\). The document has \(m\) clauses. At the embedding step, every clause \(C_{{j}}\) is mapped to a dense vector representation by an embedding function \({f}(\cdot)\), for example, BERT. The output vector \({E}({C}_{{j}})\in\mathbb{R}^{{d}}\) is the semantic and contextual representation of the clause, where \(d\) is the embedding dimension that decides the size of the numerical representation space. These parameters collectively represent the transition from unparsed legal text to parsed, machine-readable inputs for deep learning.

Table 4 illustrates the preprocessing pipeline applied to a sample legal sentence. Each stage transforms raw text into progressively structured forms-cleaned text, tokens, normalized roots, clause-level segmentation, and embeddings. This systematic process ensures that unstructured legal documents become machine-readable, preserving both semantic meaning and legal context for deep learning models.

3.3. Transformer-Based Legal Draft Generation

The EduLegal-DL system uses a transformer model to generate initial drafts of educational legal documents automatically. Transformers are very effective in legal text modeling as they utilize self-attention mechanisms that can represent long-range dependencies of complex clauses. Given a preprocessed clause sequence \({X}=\{{x}_1,{x}_2,\dots,{x}_{{n}}\}\), each token \({x}_{{i}}\) is converted into an embedding vector, \({e}_{{i}}={f}({x}_{{i}})\in\mathbb{R}^{{d}}\) where \({f}(\cdot)\) is the embedding function (e.g., BERT), and \(d\) is the embedding dimension. The self-attention mechanism computes contextual representations by comparing tokens with one another. For queries \(Q\), keys \(K\), and values \(V\), attention is defined as in Eq. (1).

Algorithm 1 establishes the transformer-based drafting process. The input sequence is first mapped to a dense vector space and passed through the transformer encoder to generate contextual representations (Steps 1–2). A new blank draft is initialized (Step 3), and the generation is initiated. For every time step \(t\), the model generates the probability distribution of the next token by taking a softmax of the hidden state (Step 4). The most probable token is obtained and added to the draft. This goes on until the draft attains a maximum length \(T\) or an end-of-sequence token is produced. The last draft \(D\) is then returned (Step 5). This cycle of generation ensures that every newly created word is conditioned on the previous sequence, allowing the model to maintain legal consistency, semantic coherence, and compliance alignment throughout the drafting process.

3.4. Compliance Testing Module (BERT-Based)

The moment the transformer model generates initial education legal drafts, these are verified by a compliance test module developed based on BERT-based classifiers. The primary role of this module is to verify whether the generated drafts comply with primary compliance dimensions like equity, inclusivity, accessibility, data privacy, and UNESCO compliance. Bidirectional Encoder Representations from Transformers (BERT) is most appropriate to perform this function since it is a bidirectional text encoding, taking both the left and right context of words in a clause. It is therefore able to identify fine-grained legal dependences as well as compliance cues. Given a clause \({C}_{{j}}=\{{t}_{{a}},{t}_{{a}+1},\dots,{t}_{{b}}\}\), it is first converted into embeddings and passed through the BERT encoder to obtain a contextualized representation as in Eq. (3).

The draft document (\(D\)) consists of \(m\) clauses, and the clauses \({C}_{{j}}\) is the \(j\)-th clause. \({P}({y} \mid {C}_{{j}})\) indicates the likelihood of a clause \({C}_{{j}}\). Complies with classifier calculations. To calculate the sum, use the formula: \(\sum_{{j}=1}^{{m}} {P}({y}\mid {C}_{{j}})\) calculates compliance probabilities for all clauses using the factor \({1}/{{m}}\). Calculates the document’s average compliance by normalizing the score by clause count. The \({S}({D})\) measure of document-level compliance between 0 and 1 indicates the degree to which the created legal draft meets regulatory criteria. This research verifies compliance clause by clause and comprehensively at the document level. It ensures clause-level evaluation, allowing the system not only to determine if a document is compliant but also to pinpoint specific clauses that may require revision.

Clause-Level Feedback

Clause-level outputs are important because they offer actionable feedback. An example is where a clause on students’ accessibility is not UNESCO compliant; the system indicates that it is non-compliant and flags it for revision. This detailed analysis revises drafts incrementally and ensures the final legal document is complete and in sync with the world. With the inclusion of this compliance verification system using BERT, EduLegal-DL ensures that automatically generated versions are not only contextually sound but also ethically and legally compliant, minimizing the possibility of lack of oversight in policy-making.

Algorithm 2 starts with feeding all the clauses in the draft into the BERT encoder to generate contextual representations (Step 2). For each clause, the classifier outputs a probability distribution over the categories of compliance using softmax, and the most likely label becomes the compliance status of a clause. Each outcome is saved as clause-level feedback, marking compliant and non-compliant sections. The total compliance score for the draft is then calculated as the average probability for all the clauses (Step 3). The system finally returns the document-level compliance score along with explanatory clause-level labels (Step 4), which could be helpful for policymakers to pinpoint and identify offending passages. Table 5 shows the clause-level compliance evaluation of generated legal drafts using the BERT-based module.

3.5. Optimization Through Hyperparameter Tuning

The deep learning model operations, such as transformers and BERT-based classifiers, heavily depend on selecting the appropriate hyperparameters. These training parameters, which are not learned, determine how the learning process will happen. Hyperparameter tuning becomes a necessity in EduLegal-DL to ensure that the system finds a balance between drafting accuracy, compliance accuracy, and computational cost. Some of the most important hyperparameters are the learning rate (\(\eta\)), batch size (\(B\)), number of attention mechanism heads (\(h\)), and dropout probability (\(p\)). All have immediate effects on how the model is learning from data. For instance, the update rule for model parameters while training is given by Eq. (6).

Here, \(\theta\) denotes the trainable model parameters, and \({L}_{{B}}\) is the average loss over a mini-batch of size \(B\). Each pair \(({x}_{{i}},{y}_{{i}})\) signifies input sample \({x}_{{i}}\) (example: educational legal document clause) and target label \({y}_{{i}}\) (compliant, non-compliant). The term \({L}({x}_{{i}},{y}_{{i}})\) a loss function (e.g., cross-entropy) that measures the difference between the model’s prediction and the actual label for sample \(i\). The operator \(\nabla_{\theta}\) indicates loss gradient relative to model parameters. This approach offers stable and efficient parameter updates during training by averaging gradients across all \(B\) samples in the mini-batch. This research helps EduLegal-DL optimize its transformer and compliance modules for legal writing and verification. A small batch size introduces stochastic noise that can improve generalization, while a large batch size provides more stable gradients but requires higher computational resources.

Algorithm 3 specifies how EduLegal-DL selects its hyperparameters. It begins by specifying a search space (\(\Lambda\)) consisting of learning rate, batch size, number of attention heads, and dropout rate (Step 1). The model then searches through each candidate setting (Step 4), training the model and calculating the validation loss \(Lval\)(\(\lambda\)). When the new arrangement beats the current best result obtained, the system updates the stored best arrangement \(\lambda^{*}\). Once the system tests all candidates, pseudocode returns the most effective subset of hyperparameters, which maintains compliance accuracy and effectiveness (Step 5).

3.6. RLHF

While transformer models produce drafts and BERT-based classifiers ensure compliance, AI systems by themselves might not have the capacity to appreciate the intricacies of legal interpretation. To prevent this, EduLegal-DL uses RLHF, where drafts produced are reviewed by legal experts and evaluative feedback is provided. With this, the system not only produces syntactically correct and compliant text but also is morally upright, equitable, and legally viable. In RLHF, the environment is abstracted as an agent and an environment (legal drafting problem). The agent creates a draft policy \(D\), and human evaluators provide a reward score \(R(D)\) according to legal correctness, comprehensiveness, and ethical merit. The objective is to maximize the expected reward, as shown in Eq. (8).

Here, \(\pi_{\theta}\) a policy model with parameters \(\theta\) that generates draft papers \(D\). The reward score \(R(D)\) a draft based on expert comments and reflects legal correctness, compliance, and inclusivity. It is expected \({E}_{({D}\sim\pi_{\theta})}[\cdot]\) indicates the total goal. \({J}(\pi_{\theta})\) is the policy’s average reward across revisions. This research aims to maximize \({J}(\pi_{\theta})\), which ensures EduLegal-DL develops legal drafts that meet compliance standards and expert preferences. Policy gradient methods are applied to update the model, as shown in Eq. (9).

RLHF is an iterative improvement process that reinforces the drafting process in EduLegal-DL. The transformer model first produces a draft, which is then filtered by legal experts for justice, compliance, and ethical appropriateness. Experts label numerical or categorical reward scores (e.g., compliant \(=+1\), non-compliant \(=0\)), and these are employed in training the model parameters using reinforcement learning. This feedback mechanism allows the system to learn human drafting habits. In cycles of repeated training, the model improves its capacity to draft legally compliant and ethically mindful instruments. Thus, RLHF improves EduLegal-DL in three ways: it improves legal terminology by following professional guidelines, adds ethical alignment by introducing principles of fairness and openness, and improves compliance strength by following stricter adherence to UNESCO standards and local regulations.

3.7. Iterative Refinement and Finalization

The iterative refinement and completion is the final step of the EduLegal-DL pipeline, where the output continuously gets refined by successive iteration cycles of generation, compliance checking, and optimization. During each iteration cycle, the transformer-based generation module produces a draft, which is checked by the BERT-based compliance module. Compliance classification criticism and reinforcement hints derived from human annotators are used for hyperparameter tuning and RLHF for system parameter fine-tuning.

The outcome of this refinement process is a final one that meets three key requirements:

1. ・

   Legal Precision: The paper meets international standards and national laws.

2. ・

   Contextual Consistency: The composition reflects logical consistency and semantic coherence clause by clause.

3. ・

   International Flexible Adaptable Compliance: The instrument is flexible to fairness, transparency, and ethics in a manner that lends itself to flexibility geographically.

Through incremental refining and tuning, EduLegal-DL not only guarantees that the end product is computationally optimal but also legally feasible and policyable for implementation by policymakers and schools.

The improved EduLegal-DL system is rendered deployable in actual policymaking environments in schools and government ministries to enable automatic generation of legal documents and compliance verification. The system is 92% accurate in determining compliance, thus ensuring high reliability to generate legally correct and ethically correct documents. Automation using the system decreases preparation time with considerable improvement over existing manual processes, therefore speeding up the policymaking process. In terms of scalability, EduLegal-DL employs transformer models and BERT inference parallelization in processing large sets of legal documents in a variety of domains. Its architecture is strong in performance as it can scale from institution-scale policy to education legislation sizes at national- or international-scope scales. This scalability is then complemented by optimization techniques like hyperparameter learning and human feedback-conditioned reinforcement learning, which provide the model with the capacity to generalize across regimes of regulation, languages, and jurisdictions. EduLegal-DL is a viable, efficient, and scalable platform for replication in other education governance settings, allowing policymakers to produce legally adequate, context-scaled, and world-converged instruments in bulk.

4.1. Accuracy of Legal Compliance Detection

The accuracy of legal compliance detection was evaluated to determine how effectively the proposed EduLegal-DL framework identifies whether drafted legal documents align with existing educational regulations and policies. Accuracy was calculated as the proportion of correctly classified legal clauses—compliant versus non-compliant—out of the total tested clauses. Higher accuracy values reflect the model’s robustness in understanding legal language and applying regulatory criteria. This metric is crucial because even small misclassifications in compliance-sensitive contexts can lead to significant legal and administrative issues.

Figure 2 illustrates the superiority of the novel EduLegal-DL framework over baselines. EduLegal-DL consistently outperforms all types of compliance—Contract, Policy Compliance, Guideline Enforcement, and Regulatory Framework—due to its well-tuned deep learning architecture, which incorporates domain-specific compliance signals and contextual embeddings. On the other hand, AIS-BERT-PSO excels in feature optimization but lacks contextual generalizability. MuseRec-LSTM learns mention-to-mention relations but struggles with long-distance legal text semantics. FinBERT-Audit, although domain-specific for financial auditing, is inflexible in application to generic educational compliance contexts. The performance metric is quantified using classification accuracy \(\textrm{Acc}=(\textrm{TP}+\textrm{TN})/(\textrm{TP}+\textrm{TN}+\textrm{FP}+\textrm{FN})\times 100\), where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively. These results highlight EduLegal-DL’s enhanced ability to minimize false classifications and maintain robust compliance detection across heterogeneous legal categories.

4.2. F1-Score for Legal Classification

The F1-score measure is a more balanced measure of performance than accuracy because it considers both precision and recall, which is very important in legal compliance detection where false negatives can have destructive implications. The proposed EduLegal-DL had the highest F1-score in all the types of compliance, which shows its high capacity to identify relevant legal clauses and avoid misclassifications correctly. For comparison, AIS-BERT-PSO improved in accuracy through feature optimization but was poor in recall and concluded with a mediocre F1-score. MuseRec-LSTM performed well in terms of sequence dependency but struggled with sparse legal vocabulary, resulting in reduced recall. FinBERT-Audit, though domain optimized for finance, failed to generalize over education compliance texts; therefore, it kept its overall F1-score low.

Table 6 presents the F1-score comparison of EduLegal-DL with baseline models across compliance categories. The F1-score, defined as F1 \(=2\times (\textrm{Precision}+\textrm{Recall})/(\textrm{Precision}\times\textrm{Recall})\), provides a balanced evaluation by jointly considering precision (\(\textrm{TP}/(\textrm{TP}+\textrm{FP})\)) and recall (\(\textrm{TP}/(\textrm{TP}+\textrm{FN})\)), where TP, FP, and FN denote true positives, false positives, and false negatives, respectively. EduLegal-DL consistently achieved superior scores (average 0.92), indicating its ability to minimize both Type I (false positives) and Type II (false negatives) errors in legal compliance classification. This performance advantage highlights its robustness in handling imbalanced datasets and complex semantic dependencies, unlike AIS-BERT-PSO, which suffered from optimization convergence issues, MuseRec-LSTM, which lacked deep contextual awareness, and FinBERT-Audit, which was domain-specific to financial corpora and less effective in educational compliance.

4.3. Generative Quality of Legal Drafting

The generative quality of legal drafting was tested to determine how effective EduLegal-DL is in producing consistent, contextually coherent, and legally valid text compared to baselines available. Evaluation was carried out using BLEU (Bilingual Evaluation Understudy), ROUGE-L (Recall-Oriented Understudy for Gisting Evaluation), and BERTScore metrics. BLEU is applied to quantify n-gram overlap between a draft and a reference draft, ROUGE-L is applied to quantify the longest common subsequence for structural coherence, and BERTScore applies semantic embeddings for contextual coherence. EduLegal-DL performed much better on all three, with higher levels of maintenance of legal vocabulary, logical consistency, and clause-level dependencies.

Figure 3 illustrates the generative quality progression of EduLegal-DL, AIS-BERT-PSO, MuseRec-LSTM, and FinBERT-Audit across training epochs, where the curves exhibit an upward trend with irregular oscillations. The generative quality score at epoch \(e\) for model \(m\) is represented as \({Q}_{{m}}({e})={B}_{{m}}+\alpha_{{m}}\cdot (1-{e}^{-{e}/{\tau{m}}})+\delta({e})\), where \({B}_{{m}}\) denotes the base performance, \(\alpha_{{m}}\) is the maximum achievable gain, and \(\tau_{{m}}\) controls the rate of convergence. The fluctuation term \(\delta({e})\) captures stochastic training variations, producing the zig-zag patterns observed. EduLegal-DL demonstrates superior convergence behavior, stabilizing at higher scores compared to baseline methods, thereby indicating its stronger capability for high-quality legal drafting.

4.4. Adaptability Across Domains

The adaptability across domains evaluation reveals how legal AI models generalize across various legal settings like corporate, criminal, civil, and compliance areas. Given that legal language varies in structure, vocabulary, and meaning across the domains, adaptability guarantees that the model is not highly specialized to a single dataset. EduLegal-DL exhibits better adaptability with its deep contextual embeddings and multi-domain training and attains comparable high F1-scores in all the domains.

The average F1-score, which is defined as \({F}1_{{avg}}=(1/{D})\sum_{i=1}^{{D}} {F}1_{{i}}\) is used to evaluate the adaptability across domains, where \(D\) indicates the number of domains and \({F}1_{{i}}\) reflects the score in domain \(i\). According to Fig. 4, EduLegal-DL maintains consistent performance across the corporate, criminal, civil, and compliance domains while achieving the maximum adaptability with \({F}1_{{avg}}\approx 0.90\). While MuseRec-LSTM has limited transferability (\({F}1_{{avg}}=0.83\)), AIS-BERT-PSO records modest adaptability (\({F}1_{{avg}}=0.77\)) but shows sensitivity in compliance. FinBERT-Audit performs worse in criminal (\({F}1_{{avg}}=0.81\)) but has domain-specific strengths in corporate and compliance. EduLegal-DL guarantees balanced generalization across many legal situations, as this analysis demonstrates.

4.5. Human Feedback Integration

Human feedback incorporation into legal AI systems is a vital process to maintain ongoing model improvement, ethical correctness, and more trustworthy users. EduLegal-DL process employs RLHF for improving the output of document drafting and classification. In this arrangement, human professionals verify the system-created legal drafts or classification outputs, giving preference scores that determine the model improvements.

Table 7 sets out the importance of adding human feedback to the improvement of AI-powered legal drafting. Legacy models such as SVM and BERT-Legal are pretty accurate and descriptive, but less sensitive to sophisticated legal thinking. GPT-Legal has draft quality and alignment improvements, but remains behind in ethics resilience. EduLegal-DL works best among baseline models independently; when provided with human feedback, the performance of the system is optimal—94.6% accuracy, 0.83 BLEU score, and 0.91 ethical alignment. The overall reward function, \({R}(\theta)=\sum_{{i}=1}^{N} (\alpha\cdot{Acc}_{{i}}+\beta\cdot{Qual}_{{i}}+\gamma\cdot{Align}_{{i}})\), \({R}(\theta)\) is the overall reward function that evaluates the performance of the legal AI model with parameters \(\theta\). \({N}\) is total number of evaluation instances (e.g., number of documents, cases, or tasks). \({Acc}_{{i}}\) is accuracy score for the \(i\)-th task, measuring correctness in legal classification or reasoning. \({Qual}_{{i}}\) is drafting quality score for the \(i\)-th task, often measured using BLEU, ROUGE-L, or linguistic fluency metrics. \({Align}_{{i}}\) is ethical/legal alignment score for the \(i\)-th task, reflecting fairness, bias reduction, and compliance with legal norms. \(\alpha\), \(\beta\), \(\gamma\) are weighting parameters that balance the contribution of accuracy, quality, and ethical alignment to the overall reward. Typically, \(\alpha+\beta+\gamma=1\) to normalize the importance.

4.6. Computational Efficiency and Scalability

The EduLegal-DL framework highlights computational scalability and efficiency, and hence it is prepared for application in real-world large-scale legal document analysis. By combining the transformer-based representations with optimization techniques, the framework prevents redundant computation and achieves faster convergence. The adaptive attention mechanisms lead to low memory overhead but retain contextual richness such that the system can handle longer legal text without exponential growth in complexity. In addition, scalability is facilitated through parallelizing training and inference on multiple GPUs in a manner that maintains performance regardless of dataset size and domain diversity. The resulting speed-accuracy-resource-balanced availability of the presented solution renders it a practical one for large-scale legal AI systems.

Table 8 shows the computation cost and scalability effectiveness of the EduLegal-DL method for different dataset sizes. One epoch training time (\({T}_{{\mathit{train}}}\)) and inference delay (\({T}_{{\mathit{inf}}}\)) show that EduLegal-DL is overhead light compared to baseline models employing shared optimized parameters and light-weight attention mechanisms. GPU memory consumption (\({M}_{{\mathit{gpu}}}\)) is kept moderate to support viable deployment even in constrained environments. Scalability (\({S}_{{\mathit{data}}}\)) is confirmed as model performance maintenance in the process of switching from small-sized to large-sized datasets. EduLegal-DL tends to have better trade-offs on average, achieving low computational cost while keeping scalability for real-world legal document processing.

In the proposed RLHF mechanism, drafts generated by the EduLegal-DL framework are evaluated by domain experts with legal expertise in educational regulations. For example, when the model generates a draft institutional policy clause related to student assessment procedures, legal experts review the text to verify compliance with applicable education laws, accreditation standards, and statutory language requirements. If the draft contains ambiguous terminology (e.g., undefined evaluation criteria) or omits mandatory regulatory provisions, the experts annotate the draft, provide corrective suggestions, and assign structured feedback scores reflecting legal accuracy, clarity, and compliance. Drafts meeting regulatory standards receive positive feedback, whereas non-compliant or partially compliant drafts are penalized. This expert feedback is then converted into reward signals that guide subsequent policy updates, enabling the model to progressively generate legally consistent and compliant educational legal instruments.

5.1. Limitation

The primary shortcoming of EduLegal-DL is its dependency on domain-specific training data, which may restrict generalizability to underrepresented legal domains. Second, high-compute-demand large-scale training is a resource problem. Third, although human feedback enhances alignment, it introduces subjectivity, and caution should be exercised in matching expert feedback and optimization by the automated system.

5.2. Future Work

Future research will focus on scaling up EduLegal-DL to encompass cross-jurisdictional and multilingual legal systems, incorporating multimodal information such as statutes, case laws, and contracts, and utilizing human-in-the-loop optimization for scaling. In addition, adaptive reinforcement policies and explainable AI mechanisms will be investigated to enhance transparency, accountability, and trustworthiness in computer-aided legal drafting.