Graphwise @ CLEF-2025 GutBrainIE: Towards Automated Discovery of Gut-Brain Interactions - Deep Learning for NER and Relation Extraction from PubMed Abstracts

Aleksis Datseris, Mario Kuzmanov, Ivelina Nikolova-Koleva, Dimitar Taskov, Svetla Boytcheva

Published: 01 Sept 2025, Last Modified: 28 Apr 2026CEUR Workshop ProceedingsEveryoneRevisionsCC BY-SA 4.0
Abstract: This paper presents a set of approaches to tackle the named entity recognition and relation extraction from scientific literature, specifically targeting the gut-brain axis related terms and relationships between them. The proposed methods participated in the GutBrainIE Task at CLEF 2025 BioASQ Lab. The solutions rely on fine-tuned BERT-based models (BioBERT , BiomedNLP ELECTRA , BioBERT PubMed) and GLiNER for the named entity recognition task, ATLOP and REBEL fine-tuning for the relation extraction task. Hybrid models and ensemble of models are also demonstrated for end-to-end tasks. Notably, one of our proposed solutions ranked 2nd on the most difficult task of the challenge - Ternary Mention-based Relation Extraction, achieving micro-F1 37.29%. Our best system for Named Entity Recognition over the test set achieved micro-F1 80.1%. On the Binary Tag-based Relation Extraction subtask, our best solution achieved micro-F1 65.38% on the test set and on the Ternary Tag-based Relation Extraction subtask, our best result was micro-F1 63.72%. All of the proposed approaches demonstrated good performance, consistently outperforming baseline results across all subtasks of the GutBrainIE Task at CLEF 2025 BioASQ Lab.