GutHub: Tackling the GutBrainIE Task Through Entity Recognition, Disambiguation, and Relation Extraction
Authors/Creators
- 1. Department of Computer Science, Aalborg University, Aalborg, Denmark
Description
In this paper, we introduce GutHub, a comprehensive extraction pipeline developed for the CLEF 2026 GutBrainIE task, addressing Named Entity Recognition (NER), Named Entity Disambiguation (NERD), and Relation Extraction at both the mention (M-RE) and concept (C-RE) levels. The primary obstacle in this domain is navigating extreme class imbalances across a multi-tiered training corpus that ranges from massive, noisy auto-generated text to sparse, expert-curated data. To overcome this, GutHub employs a stacked transformer ensemble combining domain-specific and zero-shot models to mitigate label bias. For disambiguation, SapBERT dense retrieval paired with a Mini-LM reranker trained on the MS-MARCO dataset outperformed several domain-specific alternatives. For mention-level relation extraction we combined class-specific confidence thresholding, deterministic heuristic pruning, and negative sampling. We then mapped the mention-level relations to unique URIs to output the concept-level relations. On the official test set, the micro-F1 scores achieved for each subtask were 0.7893 (NER), 0.3497 (NERD), 0.4029 (M-RE) and 0.1295 (C-RE).