@article{gabryszak-etal-2026-facts,
    title = "From Facts to Hypotheses: Joint Detection of Biomedical Relations and Epistemic Commitment Using {LLM}s",
    author = {Gabryszak, Aleksandra  and
      Truong, Phuc Tran  and
      Binder, Arne  and
      Milosevic, Nikola  and
      Keese, Felix-Sebastian  and
      Rheinl{\"a}nder, Astrid  and
      Thomas, Philippe},
    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.602/",
    pages = "7592--7605",
    abstract = "Determining the factual status of biomedical statements, whether affirmed, negated, or uncertain, is essential for accurate understanding. To support research in this area, we introduce BioRelFact, a publicly available, expert-annotated dataset of 1,767 English biomedical sentences labeled with nine relation types and five levels of epistemic commitment. Using this dataset, we evaluate eight large language models (LLMs) from the GPT, Qwen, and Gemma families for joint relation extraction and epistemic classification. Among the evaluated models, GPT-OSS-20B performs best in both tasks (F1 77.3 for relation, 65.3 for commitment), followed by GPT-4o (75.9 and 60.2), while Qwen3-8B (Thinking) shows strong performance despite its smaller size (74.6 and 57.2). Domain adaptation has mixed effects: relative to their general-purpose counterparts, MedGemma-27B improves (+3.6 F1 for relation, +4.4 for factuality), whereas Qwen2.5-Aloe-Beta-7B declines ({--}4.3 and {--}3.5, respectively). Moreover, definition-based few-shot prompts consistently yield the best results for most models, and an explorative analysis of prediction errors suggests which specific linguistic features may drive model confusions."
}