autoReg: An Integrated NLP Pipeline for Automated Regulatory Impact Tracking of RBI Circulars

Abstract

India’s banking sector operates under one of the world’s most complex regulatory frameworks, with the Reserve Bank of India (RBI) issuing over 200 circulars annually spanning KYC/AML, digital lending, cybersecurity, and risk management. Manual interpretation of these dense, cross-referential documents remains a significant bottleneck for compliance teams. This paper presents autoReg, the first integrated NLP pipeline to automate the complete regulatory compliance lifecycle for RBI circulars across four stages: (1) automated retrieval and structured parsing; (2) abstractive summarization using fine-tuned Pegasus; (3) semantic change detection via Sentence-BERT clause-level embeddings; and (4) department-level compliance mapping through a hybrid Retrieval-Augmented Generation (RAG) pipeline. Evaluated on a curated corpus of 75 RBI circulars across five regulatory domains, autoReg achieves ROUGE-L 0.41 and BERTScore F1 0.86 for summarization (32% improvement over TextRank baseline), weighted F1 0.78 for clause-level change detection, and a RAGAS faithfulness score of 0.83. The modular architecture supports extension to SEBI and IRDAI. To the best of our knowledge, autoReg is the first system to unify all four stages of the RBI circular compliance lifecycle in a single, production-ready framework.

Introduction

This paper presents autoReg, an end-to-end NLP-based Regulatory Technology (RegTech) system designed to automate the compliance lifecycle of Reserve Bank of India (RBI) circulars for Indian banks. The system addresses the growing complexity of banking regulations by automating four key tasks: regulatory document ingestion, abstractive summarization, semantic change detection, and compliance question answering through a hybrid Retrieval-Augmented Generation (RAG) framework.

The proposed architecture includes four layers: data ingestion from RBI circulars, NLP processing using a fine-tuned Pegasus model and Sentence-BERT, a hybrid RAG pipeline combining dense retrieval and BM25 for accurate compliance mapping, and a user interface for browsing circulars, comparing regulatory changes, and querying compliance information. The system was evaluated on a curated dataset of 75 RBI circulars covering KYC/AML, digital lending, cybersecurity, risk management, and payment systems.

Experimental results demonstrate that the fine-tuned Pegasus model achieved a ROUGE-L score of 0.41 and BERTScore F1 of 0.86, representing a 32% improvement over the TextRank baseline. The Sentence-BERT change detection module achieved a weighted F1-score of 0.78 for clause-level amendment identification, while the hybrid RAG pipeline obtained a faithfulness score of 0.83, indicating reliable, source-grounded compliance responses. Domain-wise evaluation showed consistent performance across all regulatory categories.

The study concludes that autoReg provides an effective, domain-specific solution for automating RBI regulatory compliance by integrating summarization, change detection, and compliance mapping into a single pipeline. Although the current dataset is relatively small and some risk of AI hallucination remains, the system demonstrates strong potential for improving compliance efficiency, transparency, and accuracy in Indian banking. Future work includes expanding the dataset, supporting additional regulators such as SEBI and IRDAI, enabling multilingual processing, and integrating the system with banks' core operational platforms.

Conclusion

We presented autoReg, the first integrated NLP pipeline for the complete regulatory compliance lifecycle of RBI circulars. By unifying fine-tuned Pegasus summarization, Sentence-BERT semantic change detection, and hybrid RAG-based compliance mapping, autoReg substantially reduces manual regulatory interpretation effort in Indian banking. The system achieves ROUGE-L 0.41 and BERTScore F1 0.86 for summarization (32% improvement), weighted F1 0.78 for change detection, and faithfulness 0.83 for compliance querying on a curated 75-circular dataset. These results validate transformer-based NLP for Indian regulatory text and position autoReg as a foundation for a broader Indian RegTech platform.

References

[1] Reserve Bank of India, “Master Directions and Circulars,” https://www.rbi.org.in, 2024. [2] A. Arora et al., “RegNLP: NLP for Financial Regulation Compliance,” Proc. ACM ICAIF, pp. 1–8, 2021. [3] A. Okonkwo et al., “Automating Financial Compliance with AI,” IJSRA, vol. 11, no. 1, pp. 2646–2659, 2024. [4] Congruence Market Insights, “RegTech & Compliance Automation Market 2025–2032,” 2024. [5] J. Zhang et al., “PEGASUS: Pre-training with Extracted Gap-Sentences for Abstractive Summarization,” Proc. ICML, pp. 11328–11339, 2020. [6] N. Reimers and I. Gurevych, “Sentence-BERT: Sentence Embeddings Using Siamese BERT-Networks,” Proc. EMNLP, pp. 3982–3992, 2019. [7] P. Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” Proc. NeurIPS, vol. 33, pp. 9459–9474, 2020. [8] R. Mihalcea and P. Tarau, “TextRank: Bringing Order into Text,” Proc. EMNLP, pp. 404–411, 2004. [9] I. Chalkidis et al., “LEGAL-BERT: The Muppets Straight Out of Law School,” Findings of ACL, pp. 2898–2904, 2020. [10] S. Rusiya and A. Jamatia, “Implementation of Legal Documents Text Summarization,” Proc. MISP 2022, Springer, pp. 329–341, 2023. [11] I. Chalkidis et al., “NLP for the Legal Domain: A Survey,” arXiv:2410.21306, 2024. [12] Y. Gao et al., “RAG for Large Language Models: A Survey,” arXiv:2312.10997, 2023. [13] A. Pipitone and G. Alami, “LegalBench-RAG,” arXiv:2408.10343, 2024. [14] N. Wiratunga et al., “CBR-RAG for Legal QA,” Proc. ICCBR, Springer, pp. 445–460, 2024. [15] S. Es et al., “RAGAS: Automated Evaluation of RAG,” Proc. EACL, pp. 150–163, 2024. [16] M. Dahl et al., “Large Legal Fictions: Profiling Legal Hallucinations in LLMs,” J. Legal Analysis, 2024. [17] M. Bommarito and D. Katz, “LexNLP: NLP for Legal and Regulatory Texts,” arXiv:1806.03688, 2018. [18] M. Cherubini et al., “Enhancing Legal Document Building with RAG,” Computer Law & Security Review, 2025. [19] S. Robertson and H. Zaragoza, “The Probabilistic Relevance Framework: BM25 and Beyond,” Found. Trends IR, vol. 3, no. 4, pp. 333–389, 2009. [20] C.-Y. Lin, “ROUGE: A Package for Automatic Evaluation of Summaries,” Proc. ACL Workshop, pp. 74–81, 2004. [21] T. Zhang et al., “BERTScore: Evaluating Text Generation with BERT,” Proc. ICLR, 2020. [22] H. Kuhn, “The Hungarian Method for the Assignment Problem,” Naval Research Logistics Quarterly, vol. 2, pp. 83–97, 1955. [23] S. Paul et al., “InLegalBERT: Pre-Trained Language Model for the Indian Legal Domain,” Hugging Face, 2023. [24] J. Devlin et al., “BERT: Pre-Training of Deep Bidirectional Transformers,” Proc. NAACL-HLT, pp. 4171–4186, 2019. [25] A. Vaswani et al., “Attention is All You Need,” Proc. NeurIPS, vol. 30, pp. 5998–6008, 2017. [26] J. Johnson et al., “Billion-Scale Similarity Search with GPUs,” IEEE Trans. Big Data, vol. 7, no. 3, pp. 535–547, 2021.

Copyright

Copyright © 2026 Kshitij Markandeya, Atharva Chitambare, Soham Kumbhar, Vaishnavi Koshatwar. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.