NeuReg is a neuro-symbolic QA generation framework that transforms complex regulatory documents into intelligent and explainable question–answering systems. It seamlessly integrates ontology-guided knowledge graphs (KGs) and regulatory text chunks to generate high-quality, semantically grounded QA pairs. By combining structured symbolic knowledge from domain ontologies with the contextual richness of unstructured policy text, NeuReg enables accurate, diverse, and interpretable QA generation using large language models (LLMs).
- 🔄 Pipeline Overview
- 🧠 Model Architecture
- 🚀 Motivation
- ✨ Key Contributions
- 🎯 QA Generation Types
- 📂 Repository Structure
- ⚙️ Installation
▶️ Getting Started- 📖 Citation
- 📄 License
NeuReg consists of two main stages: Knowledge Extraction and Question Answer Generation.
- Regulatory text is split into coherent text chunks.
- A domain-specific Educational Funding Regulation Ontology (EFRO) is used to guide schema extraction and triple generation using GPT-4 turbo.
- Each output triple is structured as (subject, predicate, object) with post-processing.
Each chunk is mapped to its corresponding KG (based on chunk_id). The QA generation follows four steps:
- Question Type Selection – Factual, Relational, Comparative, Inferential
- Prompt Augmentation – Zero-shot, One-shot, Few-shot prompting strategies
- QA Filtering – Answer length, semantic similarity (cosine < 0.85), retry up to 3 times
- Validation – Human annotation and automatic scoring from LLMs
Figure 1: NeuReg: Neuro-symbolic framework for regulatory QA generation using ZS, OS, and FS prompting with ontology-guided KG extraction.
Access to education funding is governed by complex and evolving regulations. These policies are often communicated through lengthy documents that are difficult for students and institutional staff to interpret. NeuReg addresses this challenge by transforming unstructured regulatory guidance into structured and explainable QA datasets, bridging the gap between dense policy language and actionable decision support.
We present NeuReg, a neuro-symbolic question–answer generation framework that integrates the generative power of large language models (LLMs) with structured knowledge from ontology-guided knowledge graphs and their aligned regulatory text segments. This hybrid approach enables the generation of high-quality, semantically grounded QA pairs tailored to complex regulatory domains.
We construct a domain-specific QA dataset for regulatory compliance in education funding, encompassing four distinct question types: Factual (FactQ), Relational (RelQ), Comparative (CompQ), and Inferential (InferQ). These QA pairs are generated using multi-strategy prompting and rigorously validated through comparative assessment by expert human annotators and state-of-the-art (SOTA) LLM judges. To the best of our knowledge, this is the first QA dataset of its kind within this domain.
We conduct controlled ablation studies to quantify the individual contributions of structured KG triples and unstructured text chunks to QA generation quality, demonstrating the indispensable role of symbolic knowledge. Additionally, we evaluate the practical utility of the generated datasets by fine-tuning multiple LLMs (T5, FLAN-T5), analyzing the effects of prompting strategies (ZS, OS, FS) and model scale on QA performance.
| Type | Description |
|---|---|
| FactQ | Extract concrete details (e.g., definitions, thresholds, dates) for grounded information retrieval. |
| RelQ | QExamine entity interactions within regulatory structures, reflecting KG-based links (e.g., between providers and funding authorities). |
| CompQ | Contrast policies, programmes, or entities to highlight distinctions or trade-offs. |
| InferQ | Require synthesis or multi-hop reasoning across text and KG to derive implicit conclusions |
NeuReg/
├── README.md # Overview of the project, contributions, pipeline, and structure
├── LICENSE # Project license (MIT)
├── requirements.txt # Python dependencies for reproducing the results
├── data/ # Preprocessing and knowledge graph construction
│ ├── README.md # Overview of chunk & triple-level statistics
│ ├── chunks/ # Extracting regulatory text chunks
│ │ ├── chunks.csv # chunk dataset
│ │ └── chunks.ipynb # Chunk extraction notebook
│ ├── ontology/ # Ontology schema and KG triples
│ │ ├── ontology_schema.json # Extracted ontology schema in JSON
│ │ ├── Ontology_Guided_Triples.csv # Ontology-guided KG triples
│ │ ├── Ontology_Guided_Triples_statistics.json # Stats on generated triples
│ │ ├── EFRO_Schema_Extraction.ipynb # Extract ontology schema from guidance
│ │ └── KG_Extraction.ipynb # Generate KG using ontology + chunks
├── qa_generation/ # QA dataset generation using prompting
│ ├── README.md
│ ├── Zero-shot.ipynb # Zero-shot QA generation
│ ├── One-shot.ipynb # One-shot QA generation
│ ├── Few-shot.ipynb # Few-shot QA generation
│ ├── Zero-Shot_qa_dataset.json # Output QA dataset (zero-shot)
│ ├── One-Shot_qa_dataset.json # Output QA dataset (one-shot)
│ ├── Few-Shot_qa_dataset.json # Output QA dataset (few-shot)
│ ├── Zero_Shot_QA_analysis_report.json # Analysis report (zero-shot)
│ ├── One_Shot_QA_analysis_report.json # Analysis report (one-shot)
│ └── Few_Shot_QA_analysis_report.json # Analysis report (few-shot)
├── evaluation/ # Complete evaluation framework
│ ├── README.md # Central summary of all evaluation types and modules
│ ├── Ontology-Guided_KG_Evaluation/ # Evaluation of KG triples
│ │ ├── README.md
│ │ ├── Evaluation.ipynb # Validates triple structure and semantics
│ │ ├── evaluation_results.csv # Per-triple validation outcomes
│ │ └── evaluation_report.json # Aggregate KG validation statistics
│ ├── LLM-as-a-Judge/ # LLM-based QA evaluation (5 models)
│ │ ├── README.md # Overview of LLM evaluation setup and metric definitions
│ │ ├── DeepSeek-R1-Distill-Llama-70B/ # Evaluation results from DeepSeek-R1
│ │ │ ├── DeepSeek-R1-Distill-Llama-70B.ipynb # ipynb file
│ │ │ ├── DeepSeek_zeroshot_evaluation_results.csv # Zero-Shot QA results
│ │ │ ├── DeepSeek_oneshot_evaluation_results.csv # One-Shot QA results
│ │ │ └── DeepSeek_fewshot_evaluation_results.csv # Few-Shot QA results
│ │ ├── Gemma-2 Instruct 27B/ # ipynb file and evaluation results
│ │ ├── LLaMA 3.3 70B/ # ipynb file and evaluation results
│ │ ├── mixtral-8x22b-instruct-v0.1/ # ipynb file and evaluation results
│ │ └── Qwen3-32B/ # ipynb file and evaluation results
Each model folder includes: one `.ipynb` notebook + 3 CSVs for Zero-/One-/Few-Shot QA evaluation results
│ ├── llms results analysis/ # Cross-model aggregation and statistics
│ │ ├── README.md
│ │ ├── LLM results analysis.ipynb # Compare results across LLM judges
│ │ └── comprehensive_analysis_report.json # Metrics summary (means, deviations, majority voting agreement)
│ ├── Human Judgements/ # Human evaluation and sampling
│ │ ├── README.md
│ │ ├── Evaluation_Template.md # Annotation form and scoring rubric
│ │ ├── stratified sampling method.ipynb # Script for stratified QA sampling
│ │ ├── QA_Human_Eval_Stratified_5percent.csv # Final sampled QA set for annotation
│ │ ├── QA_Sampling_Summary_Statistics.csv # Summary of sampled distribution
│ │ ├── QA_Stratified_Sampling_Visualization.png # Sample distribution plots
│ │ ├── human results analysis.ipynb # Human score processing and statistics
│ │ └── human_evaluation_analysis_report.json # Metrics summary (means, deviations, majority voting agreement)
│ ├── LLM vs Human/ # Correlation between LLM and human scores
│ │ ├── README.md
│ │ ├── LLM vs Human.ipynb # Notebook to compare LLM vs human scores
│ │ └── human_llm_comparison_results.csv # EM,f1
├── analysis/ # Statistical analysis & insights
│ ├── README.md
│ ├── Statistical_Analysis.ipynb
│ ├── Readability_Analysis.csv # FKGL, Flesch, etc.
│ ├── Vocabulary_Diversity_Analysis.csv
│ ├── Length_Distribution_Analysis.csv
│ ├── LLMs_based_results_analysis.ipynb
│ └── LLMs_Analysis_report.csv
Ablation Studies/
│
├── Ablation Study 1/
│ ├── chunks_only_qa_dataset.ipynb
│ ├── Ablation_1_chunks_only_analysis_report.json
│ └── Ablation_1_chunks_only_qa_dataset.json
│
├── Ablation Study 2/
│ ├── KG_only_qa_dataset.ipynb
│ ├── Ablation_2_kg_only_analysis_report.json
│ └── Ablation_2_kg_only_qa_dataset.json
│
├── Evaluation/
│ ├── chunks_only_Evaluation/
│ │ ├── DeepSeek-R1-Distill-Llama-70B/
│ │ ├── Gemma-2 Instruct (27B)/
│ │ ├── Llama 3.3 70B/
│ │ ├── mixtral-8x22b-instruct-v0.1/
│ │ └── Qwen3-32B/
│ │
│ └── KG_only_Evaluation/
│ ├── DeepSeek-R1-Distill-Llama-70B/
│ ├── Gemma-2 Instruct (27B)/
│ ├── Llama 3.3 70B/
│ ├── mixtral-8x22b-instruct-v0.1/
│ └── Qwen3-32B/
│
└── Results Analysis/
├── Chunks Only/
│ └── Chunks Only Evaluation Analysis.ipynb
└── KG Only/
└── KG Only Evaluation Analysis.ipynb
├── fine_tuning/ # Fine-tuning experiments on QA datasets
│ ├── README.md
│ ├── t5_small/
│ ├── t5_base/
│ ├── t5_large/
│ ├── flan_t5_small/
│ ├── flan_t5_base/
│ └── flan_t5_large/ # results
git clone https://github.com/RGU-Computing/NeuReg.git
cd NeuReg
pip install -r requirements.txtTo reproduce the NeuReg QA generation pipeline:
cd data/chunks/
jupyter notebook chunks.ipynbcd data/ontology/
jupyter notebook EFRO_Schema_Extraction.ipynb # Extract EFRO ontology
jupyter notebook KG_Extraction.ipynb # Generate KG triplesChoose your prompting strategy:
cd qa_generation/
# Choose one of the following:
jupyter notebook Zero-shot.ipynb # Zero-shot prompting
jupyter notebook One-shot.ipynb # One-shot prompting
jupyter notebook Few-shot.ipynb # Few-shot promptingcd evaluation/llm_judges/[ModelName]/
# Example:
cd evaluation/llm_judges/DeepSeek-R1-Distill-Llama-70B/
jupyter notebook DeepSeek-R1-Distill-Llama-70B.ipynbcd evaluation/llm_vs_human/
jupyter notebook llm_vs_human_Analysis_results_analysis.ipynbcd fine_tuning/t5_small/ # or flan_t5_base/, flan_t5_large/, etc.
# Choose based on your dataset:
jupyter notebook t5_small_zero.ipynb # Zero-shot dataset
jupyter notebook t5_small_one.ipynb # One-shot dataset
jupyter notebook t5_small_few.ipynb # Few-shot datasetArshad, U., Corsar, D., & Nkisi-Orji, I. (2025). NeuReg: Neuro-Symbolic QA Generation from Regulatory Compliance. In Proceedings of the 13th Knowledge Capture Conference (K-CAP ’25), 232–235. https://doi.org/10.1145/3731443.3771375
GitHub Repository: https://github.com/RGU-Computing/NeuReg
This project is licensed under the MIT License. © 2025 School of Computing, Engineering and Technology, Robert Gordon University, UK. For full license details, see the LICENSE file.