生物医学領域における知識ベースの引用理由推論

Citation is central to scholarly communication, enabling researchers to navigate rapidly expanding literature and identify relevant prior work. Yet the reasoning behind why a particular paper is cited is often implicit or opaque. Although academic search engines and literature tools rank candidate papers for a query, the motivations underlying these rankings are rarely transparent, making it difficult for scholars to interpret and act on retrieved results-especially in biomedical research where domain knowledge is essential.

We propose an encoder-decoder framework that leverages curated biomedical knowledge to generate explanations of citation motivation in a structured bio-triplet format. We evaluate the approach against recent families of pretrained language models for text generation, including BERT-style (and variants) and GPT-style (and variants) models. In cancer-focused experiments using PubMed Central, we annotate over 10,000 citation relations with bio-triplets grounded in curated knowledge from multiple biomedical databases. Trained on these annotations, our model outperforms strong sequence-generation baselines, improving precision, recall, and F1 for citation-motivation generation.

Code and data are available at Zenodo (archival DOI: 10.281/zenodo.14893445) and GitHub: https://github.com/zhongxiangboy/Knowledge-based-Citation-Reasoning-for-Biomedical-Domain.

No supplementary information is available for this manuscript.