Ontology-Guided Legal NER and Relation Extraction for Uzbek Statutes with a Juridical Knowledge Graph
Abstract
This article presents an ontology-oriented algorithm for knowledge extraction from Uzbek regulatory texts, generating a temporal knowledge graph suitable for data-dependent legal queries. The approach combines a preprocessing module in the form of alphabetical normalization, high-precision rules and gazetteers for structural templates (act identifiers, articles, dates, event triggers), a transformer-based NER with a CRF and embedded rule features, and relation classification using a span-marker transformer under ontological constraints. A separate temporal semantics module interprets entry/expiration formulas and updates norm validity intervals, with the consistency of the results verified using OWL/SHACL. Furthermore, a corpus of approximately 18,000 sentences with two-layer annotation (BIOES entities and inter-entity relations) was generated for training neural network models, demonstrating that the algorithm is robust to variability in writing and genres. Furthermore, the implemented approach ensures high structural accuracy at the graph level and significantly reduces inconsistent relationships compared to unsupervised approaches. Potential applications include legal searches, compliance checks, and restoring the current version of a regulation as of a given date.