NormDef-FR: an annotated corpus of normative definitions for the automatic processing of French legal codes

<title>Abstract</title> <p>Normative definitions play a central role in the interpretation and application of the law: they establish the legal meaning of the terms, categories and regimes used in legislative and regulatory texts. However, these definitions are scattered across more than 161,000 articles currently in force, spread throughout the entire body of French legal codes, with no structured resource available to automatically identify or index them. To our knowledge, no annotated reference corpus dedicated to this task is available for French law. This article introduces \textit{NormDef-FR}, an annotated open-source corpus for the automated study of normative definitions in French legal codes. Version 1.0.0 contains 693 normative definitions of \textit{gold} standard, extracted from 363 articles covering 50 distinct legal codes. Each instance links a defined term, a definition, legal metadata, a definition type and, where the text permits, information on the scope of application. The corpus was constructed using a hybrid pipeline combining detection via linguistic patterns, automatic extraction, manual annotation and systematic deduplication. Inter-annotator agreement achieves a Kappa coefficient~\citep{cohen_coefficient_1960} of \textit{0.98} for the binary scheme and \textit{0.97} for the three-class scheme, confirming the robustness of the annotation scheme. The corpus is accompanied by a manually verified article-level benchmark, comprising 300~definitional articles and 300~verified true negatives, as well as 449~annotated candidates for automatic validation. The dataset is distributed with comprehensive documentation (\texttt{DATA\_CARD}, \texttt{MODEL\_CARD}, bilingual annotation guides ) and reproducible scripts. In order to characterise the difficulty of the tasks associated with the resource, we provide illustrative baselines at three levels: rule-based term-definition extraction, the detection of definitional articles, and the automatic validation of candidates. These experiments show that French normative definitions exhibit a high degree of formal regularity that can be exploited by explicit patterns, that the detection of articles is well resolved against verified negatives, and that the fine-grained validation of candidates constitutes the main bottleneck, with performance stabilising around \textit{0.71 macro-F1} in cross-validation. NormDef-FR thus constitutes a reusable reference resource for French-language legal NLP, designed to support the development of systems for knowledge extraction, legal information retrieval and retrieval-augmented generation.</p>