A workflow-oriented and risk-aware system for Turkish legal named entity recognition: integrating transformer-based models with legal knowledge graphs

Abstract

The transparent and efficient conduct of judicial processes depends on transforming large, unstructured legal texts into computationally structured and machine-processable information. This study presents a workflow-oriented and risk-aware system architecture based on the complex morphological structure and high error cost of Turkish legal texts. The tasks of legal named entity recognition, anonymization, citation extraction, and institutional analysis are positioned as fundamental components for legal workflows. Accordingly, the Turkish Legal Named Entity Recognition (TLNER) Dataset, consisting of decisions from the Council of State and the Court of Cassation, the highest judicial bodies in Turkey, is presented. Transformer-based language models with different pre-training strategies are analyzed for legal workflows using this dataset. To overcome the inadequacy of standard performance metrics in measuring legal risks, the Workflow-Aware Risk Score (WARS) formulation is introduced. Thus, the models are considered for different error types, and the cost of these error types in legal workflows is quantified. The experimental results demonstrated that while the BERTurk model generally established a consistent and robust baseline, the ConvBERTurk model exhibited highly competitive performance, particularly securing the highest accuracy and lowest risk in citation extraction workflows. Furthermore, a legal knowledge graph layer is incorporated into the system to support the transformation of named entity recognition outputs into structured information. This allows the establishment of queryable relational links between courts, legislations, and cases. The end-to-end architecture not only anonymizes raw text but also transforms judicial decisions into an institutional knowledge entity, providing a scalable decision support mechanism for legal professionals.