Automated extraction of Italian tax-court judgments with citation-hallucination control

The authors present an automated pipeline that decomposes approximately 330,000 Italian tax-court judgments into individual legal issues and extracts structured XML representations grounded in the IRAC framework. The system uses DeepSeek V3 for cost-efficient processing and couples extraction with a hallucination-detection filter that compares model-generated references against those identified by the Linkoln parser.

• Targets a corpus of 330,000 first- and second-instance Italian tax-court decisions.
• Uses DeepSeek V3 to process documents at sustainable cost.
• Validates the pipeline on 50 expert-annotated judgments, measuring inter-annotator and LLM-vs-expert agreement.
• Normalizes citations to standard identifiers (URN-NIR, ECLI, CELEX) using Linkoln.

This structured extraction provides a concrete starting point for downstream applications such as issue-level retrieval, citation-network analysis, and the construction of large-scale legal reasoning datasets.