Real-Time Natural Language Processing for Regulatory Change Management: Extracting Obligations from SEC, FCA, and MAS Filings

Abstract

The rapidly changing nature of global financial regulation has created an unprecedented need for automated systems that can track, understand and operationalise regulatory change in real time. In this paper, we present a novel hybrid natural language processing (NLP) framework leveraging domain-adapted transformer-based language models and graph-based knowledge representation to extract, classify and monitor regulatory obligations from filings issued by three major regulatory bodies, the U.S. Securities and Exchange Commission (SEC), the UK Financial Conduct Authority (FCA) and the Monetary Authority of Singapore (MAS). The suggested system consists of a fine-tuned FinBERT encoder, a bi-directional conditional random field (BiCRF) decoding layer and a dynamic knowledge graph to track obligation-linkage. The framework is evaluated on a curated corpus of 10,625 regulatory documents containing 2,853,180 tokens and 779 manually annotated obligations. It achieves a macro-averaged F1-score of 0.917, an AUC-ROC of 0.951 and a mean processing latency of 124 milliseconds per document, significantly outperforming baseline transformer architectures. Ablation investigations verify the contribution of each subsystem component, and cross-jurisdictional transfer tests show strong generality across regulatory environments. The findings have meaningful implications for RegTech deployment, proving that real-time, interpretable and scalable NLP may consistently help compliance teams in financial institutions navigate multi-jurisdictional regulatory settings.