Source-linked AI summary
TWIX: a Two-Stage Approach for End-To-End Named Entity Recognition and Relation Extraction
Marco Martinelli, Laura Menotti
TL;DR
The expanding gut-brain-axis literature creates a need for information extraction systems, especially because end-to-end subtasks propagate upstream errors. TWIX addresses this with three interconnected two-stage modules and ranks first across all four GutBrainIE test subtasks while substantially outperforming baselines.
Problem
The rapidly growing gut-brain-axis literature and upstream-to-downstream error propagation create challenges for end-to-end information extraction.
Method
TWIX is a precision-oriented pipeline with EFCL NER, retrieve-and-rerank NEL, and two-stage RE training across three interconnected modules.
Results
TWIX substantially outperforms baselines and ranks first among participant submissions across all four GutBrainIE subtasks.
Takeaways & Limitations
The results support the effectiveness of TWIX’s two-stage pipeline for improving end-to-end information extraction performance in the GutBrainIE setting.
Takeaways & Limitations
Final NERD performance combines upstream mention extraction with concept assignment, so it cannot establish that TWIX’s NEL module is superior from the final score alone.
Abstract
from arXiv · showhide
The exponential growth of scientific publications calls for automatic Information Extraction (IE) systems to support knowledge discovery. In this context, the GutBrainIE benchmark evaluates Named Entity Recognition (NER), Named Entity Recognition and Disambiguation (NERD), and Relation Extraction (RE) systems in the gut-brain axis domain. We propose Two-stage Workflow for Information eXtraction (TWIX), an end-to-end IE pipeline featuring three interconnected modules, each leveraging a two-stage framework to solve all four GutBrainIE subtasks. Evaluation on the development and test sets shows that our method substantially outperforms the baseline by a wide margin, while also ranking first among all participant submissions across all subtasks. These results indicate that the proposed two-stage pipeline effectively improves both precision and recall in practical settings.
1. Introduction
The rapidly expanding gut-brain-axis literature creates a need for automatic information extraction, while end-to-end task dependencies make error propagation a central challenge. TWIX addresses this with three interconnected two-stage modules and achieves leading benchmark performance.
- Gut-brain-axis PubMed publications more than doubled between 2020 and 2025, increasing from around 300 to more than 700 articles per year.
- GutBrainIE benchmarks entity mentions, concept-level links, and semantic relations across four subtasks of increasing complexity.
- In the end-to-end setting, downstream tasks depend on predictions from upstream NER and NERD components.
- TWIX uses three interconnected two-stage modules: EFCL for NER, retrieve-and-rerank for NEL, and staged noisy-data pretraining followed by high-quality fine-tuning for RE.
- TWIX substantially outperforms baselines on development data and ranks first across all four subtasks on the test set.
2. Related Work
Prior work commonly uses pipeline or joint architectures for end-to-end NER and relation extraction, with transformers and retrieve-and-rerank designs prominent in specific subtasks. The related work motivates extending standard formulations while recognizing pipeline competitiveness.
- Pipeline architectures separate entity identification from relation prediction, whereas joint and multi-task models represent entities and relations simultaneously.
- Pipeline methods may suffer from propagation errors, yet simple and well-engineered pipelines remain highly competitive.
- NER commonly uses a pretrained transformer with a token-level classification head, the dominant strategy in the 2025 GutBrainIE challenge.
- Similarity-based NEL systems often retrieve a ranked top-k candidate set before contextual reranking.
- Transformer-based contextual models markedly improved document-level relation extraction and are now standard in DocRE.
3. Data and Task Description
GutBrainIE uses PubMed titles and abstracts to evaluate structured extraction of biomedical entities, links, and relations in the gut-brain domain. Its four subtasks span entity extraction and normalization, and mention- or concept-level relation extraction.
- GutBrainIE contains PubMed titles and abstracts focused on gut-brain interactions and their implications for neurological and mental diseases.
- The corpus covers 13 entity types, including general biomedical categories and gut-brain-axis-specific categories such as microbiome and dietary supplement.
- Entity mentions are linked to six standardized biomedical vocabularies and a custom GutBrainIE ontology for unmapped mentions.
- The relation schema contains 17 predicates and 55 distinct subject-label, predicate, and object-label triples.
- The Gold, Silver, and Bronze collections represent expert-curated, supervised student, and fully automatic annotation quality levels, respectively.
- Entity extraction and normalization: NER identifies and classifies entity mentions into one of 13 categories, while NERD additionally links each predicted mention to a concept URI.
- Relation extraction: M-RE predicts relations between document-level entity mentions, whereas C-RE predicts relations between linked biomedical concepts and their categories.
- The official leaderboard uses micro-averaged F1 as its reference metric because entity and relation labels are strongly imbalanced.
4. Methodology
TWIX is an end-to-end pipeline with three interconnected modules for NER, NEL, and RE, each using a two-stage strategy. Its design separates candidate generation from validation or scoring to improve precision and limit error propagation.
- TWIX comprises interconnected NER, NEL, and RE modules for processing GutBrainIE documents.The NER module operates over document titles and abstracts; downstream modules consume its predictions.
- NER module: NER uses Extract First Classify Later: Term Extraction identifies candidate spans, then Term Classification validates them and assigns entity labels.Term Extraction uses generic BIO term tags, while Term Classification can reject candidates as not entities.
- NER module: The two-stage NER design filters false positives during extraction and classification while allowing high recall for potentially relevant biomedical terms.This architecture is explicitly intended to improve precision before predictions reach downstream modules.
- NEL module: NEL retrieves a ranked top-k set of candidate concepts and reranks them contextually to select the final concept identifier.The retriever scores mention–concept pairs, while the reranker performs more accurate disambiguation over the reduced candidate set.
- NEL module: NEL retrieval maps mentions and concept descriptions into a shared vector space and ranks candidates by similarity.The implementation experiments with dual encoders and text embedding models, using biomedical transformer backbones.
- RE module: The RE module uses two-stage ATLOP training: noisy large-scale pretraining precedes fine-tuning on a high-quality manually annotated dataset.The model uses RoBERTa-large as its ATLOP transformer backbone.
5. Experiments and Results
Development-set experiments compare standard biomedical-transformer NER, majority-voting ensembles, the EFCL NER architecture, NERD configurations, and relation-extraction training settings. Across these analyses, the two-stage EFCL design achieves the strongest NER results, while NEL configurations remain closely clustered and RE precision benefits from gold data and longer training.
- Standard NER: The best standard transformer reaches a micro-averaged F1 of 0.8120, while the lowest scores 0.7977, leaving only 0.0143 F1 points between them.Model choice alone has limited impact under standard token classification.
- NER ensembles: The best majority-voting ensemble reaches 0.8462 F1, improving by 0.0342 over the best single-model baseline through higher precision.Precision rises from 0.8009 to 0.8835, while recall remains comparable; ensembles of four or five models perform best.
- EFCL NER: The best EFCL TE+TC configuration reaches 0.8925 micro-averaged F1, improving by 0.0805 over the best single model and 0.0463 over the best ensemble.Its strongest configurations use BiomedBERT-large-uncased-abstract as term extractor with BiomedBERT-large-uncased-abstract or BiomedBERT-base-uncased-abstract-fulltext as classifier.
- NEL: NEL configurations differ by only 0.0119 F1 points, with the best BiomedBERT retriever-reranker configuration reaching 0.8036 micro-averaged F1.The small spread motivated fixing text-embedding retrieval and cross-encoder reranking while varying transformer backbones.
- Relation Extraction: For mention-level and concept-level RE, training-data choices affect precision but not recall, while gold annotations and longer training improve precision.Using gold annotations as the manual dataset and silver and bronze annotations as distant data outperforms the alternative split; the same trend appears for concept-level RE.
6. Submission Results
On the official test set, TWIX’s two-stage configurations outperform the baseline across NER and NERD, with gains concentrated in precision and end-to-end performance. The results also show that NERD remains dependent on upstream mention extraction, while retriever–reranker backbone choices have limited impact.
- NER: 0.8244 micro-averaged F1 makes the best majority-voting ensemble 0.0248 points better than the baseline.Precision rises from 0.7782 to 0.8639, while recall generally remains below the baseline.
- NER: 0.8740 micro-averaged F1 makes the best EFCL NER run 0.0744 points better than the baseline.The run reaches precision 0.9548 and recall 0.8058; its F1 gain is mainly precision-driven.
- NERD: 0.6890 micro-averaged F1 makes the best NERD run 0.2492 points better than the official baseline.All submitted runs exceed the baseline, with scores ranging from 0.6797 to 0.6890.
- NERD: Changing the retriever and reranker backbones between BiomedBERT and BioLinkBERT has limited impact once the retrieve-and-rerank architecture is fixed.The 24-run performance range is narrower than one F1 point, and the strongest configurations tend to use BiomedBERT for both components.
- Interpretation: NERD scores combine mention extraction and concept assignment because the NEL component assigns one concept to every received mention.NEL cannot recover mentions missed by NER or remove incorrectly extracted spans, so the final score alone cannot establish NEL superiority over other systems.
- Overall results: TWIX ranks first among participating systems for NERD and supports the development-set design decisions on the official test set.The reported pipeline combines high-quality EFCL mentions with stable retriever–reranker linking.
7. Conclusion
TWIX is a precision-oriented pipeline with interconnected NER, NEL, and RE modules, each using a two-stage strategy. Test results report improvements over the baseline for NER, NERD, and both relation-extraction subtasks.
- Contribution: TWIX addresses all four GutBrainIE subtasks through three interconnected modules for NER, NEL, and RE.The pipeline is explicitly designed as a precision-oriented IE system.
- Architecture: EFCL extracts candidate mentions before assigning entity types, while NEL retrieves top-k concepts before contextual reranking.The RE module uses ATLOP with pre-training on noisy labels and fine-tuning on high-quality annotations.
- Results: 87.40% versus 79.96% precision marks the reported NER improvement over the baseline.The conclusion attributes the result to the effectiveness of the EFCL architecture.
- Results: 61.32% versus 38.86% F1 marks the mention-level RE improvement, while concept-level RE rises from 13.45% to 34.75%.The conclusion reports significant improvements for both RE settings.
Declaration on Generative AI
The authors disclose using ChatGPT, Grammarly, and Comet for writing-style and formatting assistance, followed by review and editing.
- ChatGPT, Grammarly, and Comet were used for writing-style and formatting assistance.The authors state that they reviewed and edited the content afterward and take responsibility for the publication.
A. EFCL NER Results on the Development Set
The appendix reports development-set NER performance for the EFCL architecture across all tested configurations, using micro-averaged precision, recall, and F1.
- Experimental setup: 49 EFCL configurations were tested on the development set.The appendix table reports these configurations for the proposed architecture.
- Results table: Each EFCL configuration combines a term extractor and term classifier, with results sorted by decreasing F1.The table reports precision, recall, and F1, highlighting the best and underlining the second-best result for each metric.