Published October 4, 2026 | Version v1

Do You Need a Frontier LLM, or Can Domain-Specific Encoders Compete in Gut–Brain Information Extraction?

  • 1. Ontotext, Sofia, Bulgaria
  • 2. FMI, Sofia University, Sofia, Bulgaria
  • 3. Heidelberg University, Heidelberg, Germany
  • 4. NOVA University of Lisbon | CLUNL, Lisboa, Portugal
  • 5. IICT, Bulgarian Academy of Sciences, Sofia, Bulgaria

Description

This paper presents a comprehensive suite of Information Extraction systems developed for the GutBrainIE Task at the BioASQ Lab of CLEF 2026, aimed at elucidating the gut-brain interplay from biomedical literature. For the Named Entity Recognition (NER) subtask, we evaluate domain-specific BERT models, GLiNER architectures, and ensemble methods, incorporating distant supervision and data augmentation. Our GLiNER-based approach demonstrated superior robustness on unseen test data, securing second place. For Named Entity Recognition and Disambiguation (NERD), which we address as an entity linking problem, we propose a two-stage hybrid system—which also ranked second. It combines deterministic frequency-based gazetteers for in-vocabulary terms and tackles out-of-vocabulary disambiguation with PubMedBERT and a fine-tuned SapBERT embedding models. Finally, for Mention-level (M-RE) and Concept-level (C-RE) Relation Extraction, we contrast an encoder-based ATLOP architecture, supervised fine-tuning of an open-weights LLM (Qwen3.5-4B), and few-shot prompting of proprietary GPT-5 models (GPT-5.4 and GPT-5.5). Our Concept-level Relation Extraction system also ranked second in the respective subtask.

Files

paper13.pdf

Files (1.6 MB)

Name Size
md5:5ea597167568629e42a3420510236e77
1.6 MB Preview Download

Additional details

Funding

European Commission
HEREDITARY - HetERogeneous sEmantic Data integratIon for the guT-bRain interplaY 101137074