Published October 4, 2026 | Version v1

GetGut: A Multi-Stage Pipeline for Gut-Brain Axis Information Extraction

  • 1. Department of Computer Science, Aalborg University, Aalborg, Denmark

Description

This paper presents the GetGut@AAU pipeline, a multi-stage, waterfall information extraction architecture designed for the GutBrainIE Task within the BioASQ Lab at CLEF 2026. Our solution addresses the extraction of structured knowledge across four subtasks: Named Entity Recognition (NER), Named Entity Recognition and Disambiguation (NERD), Mention-Level Relation Extraction (M-RE), and Concept-Level Relation Extraction (C-RE). To address the disparities in annotation density and quality across the training datasets, we introduce a Weighted Funnel Fine-Tuning methodology for our NER and M-RE models. We progressively train our pipeline by starting with large-scale, weakly supervised data and gradually fine-tuning on smaller, expert-annotated datasets. This allows learning broad biomedical terminology before refining precise entity boundaries. Furthermore, we address data quality on a per-task basis. We mitigate positive-class bias in M-RE through a controlled negative sampling mechanism, and we boost NERD accuracy by excluding distantly supervised noise from our reference dictionary. Our proposed system achieved third place in both the M-RE and C-RE subtasks with Micro F1 scores of 0.4054 and 0.2020, respectively, and achieved 6th and 5th place in the NER and NERD subtasks with Micro F1 scores of 0.8014 and 0.5517. All code regarding the implementation can be found at: https://github.com/SophusJ/GetGut-AAU.

Files

paper17.pdf

Files (2.2 MB)

Name Size
md5:292cc466c8608cab4aeec683d96ad278
2.2 MB Preview Download

Additional details

Funding

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