ELiRF-UPV at BioNNE-R Task: Transfer Learning on Transformer-based Models for Biomedical Relation Classification

The implementation of NLP solutions in the biomedical domain is being applied to an elevated number of fields, thereby increasing the complexity of the problems posed by new, but widely applicable tasks. The BioNNE-R Shared Task is an example of it. This task involves extracting biomedical entity relations, with the added difficulty of the presence of both nested named entities and relations. The task also presents the challenge of addressing this task in English and Russian. In this paper, we present the ELiRF-UPV participation in this shared task, in the context of the BioASQ Lab at CLEF 2026. We show the different models implemented to approach this challenge as a Relation Classification task. Our systems rely on Transformer models based on BERT followed by a post-processing phase based on the dataset characteristics and the experimental results for the validation set. We also present a model trained through a Transfer Learning approach, obtaining competitive results in all tracks, achieving the best results in the English one. During the Test Phase, our best systems achieved a value of 0.5060 in F1-macro for the English Track, 0.4717 for the Russian Track, and 0.4540 for the Bilingual Track.