@article{banasaz-nouri-etal-2026-novel,
    title = "A Novel Synthetic Dataset for Few-Shot Legal Relation Extraction in {G}erman",
    author = "Banasaz Nouri, Shiva  and
      Leitner, Elena  and
      Moreno-Schneider, Julian  and
      Rehm, Georg",
    editor = "Piperidis, Stelios  and
      Bel, N{\'u}ria  and
      van den Heuvel, Henk  and
      Ide, Nancy  and
      Krek, Simon  and
      Toral, Antonio",
    journal = "International Conference on Language Resources and Evaluation",
    volume = "main",
    month = may,
    year = "2026",
    address = "Palma de Mallorca, Spain",
    publisher = "ELRA Language Resource Association",
    url = "https://preview.aclanthology.org/ingest-lrec/2026.lrec-main.830/",
    pages = "10579--10591",
    abstract = "The legal domain is particularly challenging for natural language processing due to the personal and confidential information it contains. Despite the significant advances of large language models (LLMs), applying them to relation extraction (RE) in legal texts remains challenging, not only because of the task{'}s linguistic and semantic complexity, but also due to privacy, compliance, and infrastructure constraints under regulations such as the EU AI Act. To address these challenges, we propose a novel synthetic dataset for German legal relation extraction, created using LLMs through a controlled, privacy-preserving, template-based pipeline. The dataset allows for reproducible and legally compliant experimentation. We benchmark it using two few-shot learning paradigms, a description-enhanced Model-Agnostic Meta-Learning (MAML) framework and Prototypical Networks with supervised contrastive loss and curriculum-aware prototype enrichment. Our results demonstrate that combining few-shot learning with structured semantic knowledge achieves robust and interpretable results, with the curriculum-aware Proto-Contrastive model reaching an F1-score of 99.83{\%}."
}