@inproceedings{Kerdraon-Mimouni:CORIA-TALN-2026:2026,
    author = "Kerdraon, Gabriel and Mimouni, Nada",
    title = "Recherche d{\textquoteright}information juridique par r\'eseaux de neurones de graphe sur un corpus de droit",
    booktitle = "Actes de CORIA-TALN 2026. Actes de la 21e Conf\'erence en Recherche d'Information et Applications.  Volume 1 : articles scientifiques originaux",
    month = "6",
    year = "2026",
    address = "Nantes, France",
    publisher = "Association pour le Traitement Automatique des Langues",
    pages = "136-150",
    note = "",
    abstract = "Legal practitioners often struggle to identify the applicable legal provisions due both to the sheer volume of legal materials and the dispersion of normative sources. We developed a prototype based on a Retrieval-Augmented Generation architecture; however, its use in practical settings proves computationally expensive. We therefore propose an alternative approach based on modeling the corpus as a graph. Graph neural networks propagate contextual information, enrich the dense representations of legal articles, and improve the retrieval of relevant texts upstream of the generation step. The results show that explicitly exploiting structural information makes it possible to simplify the architecture while maintaining, and in some cases improving, performance when certain modules of the initial pipeline are retained.",
    keywords = "vinformation retrieval, legal domain, regulatory question answering, dense retrieval, graph neural networks, French regulatory corpus",
    url = "30028.pdf"
}