This paper introduces ReguMap, a neural architecture designed for automated compliance rule mapping, using legal NLP, semantic textual similarity, and information retrieval. It includes a multi-stage pipeline for regulatory text segmentation, semantic embedding generation, cross-document alignment, and mapping prediction. The evaluation uses legal NLP benchmarks and a novel compliance mapping dataset.
Key findings
Proposes ReguMap, a multi-stage neural architecture for automated compliance rule mapping.
Introduces domain-adapted embedding approach using contrastive learning on regulatory text pairs.
Presents a comprehensive evaluation framework including baseline comparisons and human evaluation.
Anticipates significant improvements in mapping accuracy with interpretable justifications for compliance decisions.
Limitations & open questions
Challenges in semantic heterogeneity, complex relationships, and contextual dependencies remain.
Need for further mitigation strategies for practical deployment.