Abstract
Legal and policy papers often articulate eligibility criteria in unstructured natural language, rendering automated reasoning over these materials a formidable challenge. This study presents a transformer-based methodology for extracting structured eligibility restrictions from legal and regulatory texts utilizing pre-trained language models, particularly BERT. The approach emphasizes the identification of entities, the extraction of condition-action pairings, and the formalization of entitlement logic in a machine-readable manner. Each sentence is examined to see if it conveys an eligibility criterion, and if it does, the pertinent constraints are extracted and organized into structured representations. The pipeline encompasses clause segmentation, legal condition annotation, and constraint normalization. Experiments are performed using a meticulously assembled dataset of eligibility-related stipulations from social benefit legislation and administrative procedures. The assessment against expert-annotated ground truth indicates that the model attains excellent precision in recognizing eligibility-related sentences and extracting essential logical criteria. The suggested system facilitates downstream applications like automated legal assistants, eligibility verification for public services, and digital compliance engines. This endeavor aims to enhance the interpretability and applicability of unstructured legal text via natural language processing and machine learning methodologies.