Source-linked AI summary
Overview of BioASQ 2026: The fourteenth BioASQ Challenge on Large-Scale Biomedical Semantic Indexing and Question Answering
Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara, Martin Krallinger, Miguel Rodríguez-Ortega, Eduard Rodriguez-López, Natalia Loukachevitch, Igor Rozhkov, Elena Tutubalina, Dimitris Dimitriadis, Vasiliki Patsiou, Grigorios Tsoumakas, George Giannakoulas, Alexandra Bekiaridou, Athanasios Samaras, Giorgio Maria Di Nunzio, Nicola Ferro, Stefano Marchesin, Marco Martinelli, Gianmaria Silvello, Georgios Paliouras
TL;DR
Biomedical language processing spans tasks such as multilingual clinical summarization, Greek clinical coding, and relation extraction between nested entities. This paper surveys BioASQ 2026’s six shared tasks and participating systems; 87 teams took part, and several systems achieved competitive performance, with some improving over baselines or prior state-of-the-art results.
Problem
Biomedical information-processing needs span multiple tasks, including multilingual clinical summarization, underrepresented Greek clinical text, and relation extraction involving nested entities.
Source: §2.3 · p. 4, §2.4 · pp. 4–5, §2.5 · p. 6Method
The paper overviews BioASQ 2026’s six shared tasks, datasets, participating systems, and their performance.
Source: §5 · p. 20Results
87 distinct teams participated, and several systems achieved competitive performance, with some improving over baselines or previous state-of-the-art results.
Source: §5 · p. 20Takeaways & Limitations
BioASQ continues to support research in biomedical semantic indexing, question answering, information extraction, and language processing through established and newer tasks.
Source: §5 · p. 20Best macro-F1 scores stayed near 0.50–0.53, indicating that relation extraction over nested entities remains an open challenge.
Source: §5 · p. 20
Abstract
from arXiv · showhide
This paper presents an overview of the fourteenth edition of the BioASQ challenge, organized in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2026. BioASQ is an international challenge series that supports progress in biomedical language processing tasks ranging from semantic indexing and information extraction to question answering and summarization. In 2026, BioASQ included six shared tasks: a) Task 14b on biomedical semantic question answering. b) Task Synergy14 on question answering for developing biomedical top- ics. c) Task MultiClinSum-2 on multilingual clinical summarization. d) Task BioNNE-R on extracting relations between nested named entities in Russian and English. e) Task ELCardioCC on clinical coding in cardiology. f) Task GutBrainIE on gut-brain interplay information extrac- tion. Across these six tasks, 87 distinct teams participated, submitting more than 1000 runs overall. As in previous editions, several submissions reached competitive performance, reflecting the continued progress of state-of-the-art methods across biomedical language processing tasks.
1 Introduction
BioASQ advances biomedical semantic indexing and question answering through annual shared tasks and benchmark datasets reflecting biomedical experts’ information needs. This paper gives a compact overview of the 2026 challenge’s tasks, datasets, participants, and performance.
1 Introduction
- BioASQ uses annual shared tasks and benchmark datasets to support biomedical language-processing research and exchanges among teams worldwide.Its tasks address expert information needs and include established tracks and newer tasks for emerging needs.Source: §1 · p. 2
- The paper summarizes the fourteenth challenge’s shared tasks and datasets, participation, and system performance.Performance is reported using state-of-the-art evaluation measures or manual assessment.Source: §1 · p. 2
2 Overview of the tasks
BioASQ 2026 comprised six shared tasks spanning biomedical question answering, clinical summarization, nested-entity relation extraction, cardiology coding, and gut-brain information extraction. The tasks used distinct datasets and evaluation setups, including multilingual tracks and developing-topic questions.
2 Overview of the tasks
- BioASQ 2026 included six tasks covering biomedical QA, developing-topic QA, multilingual clinical summarization, nested-entity relations, cardiology coding, and gut-brain information extraction.The overview emphasizes differences from previous editions.Source: §2 · p. 2
2.1 Task 14b
- Task 14b combined retrieval of PubMed documents and snippets with answer generation across three phases, using 280 new test questions and 5,729 training questions.Phase A retrieves evidence, Phase A+ produces exact and ideal answers, and Phase B generates answers from expert-selected material.Source: §2.1 · p. 3, Table 1 · p. 3
2.2 Task Synergy 14
- Synergy14 ran as a four-round dialogue in which systems retrieved evidence, experts assessed it, and answerable questions received exact and ideal answers.The edition considered 66 distinct questions, including 63 newly added developing-topic questions and three carried over from earlier editions.Source: §2.2 · pp. 3–4
2.3 Task MultiClinSum-2
- MultiClinSum-2 expanded clinical case-report summarization to 15 languages using full clinical cases paired with reference summaries.The language tracks include English, French, Spanish, Portuguese, Italian, Russian, Catalan, Norwegian, Danish, Romanian, German, Greek, Dutch, Czech, and Swedish.Source: §2.3 · p. 4
2.4 Task BioNNE-R
- BioNNE-R targets relations between nested biomedical entities, adding English and Russian annotations and monolingual and bilingual evaluation tracks.The annotation includes eight entity types and 14 relation types; nested spans can create relations connecting internal, uppermost, or external entities.Source: §2.4 · pp. 4–5
2.5 Task ELCardioCC
- ELCardioCC asks systems to assign cardiology-related ICD-10 codes to Greek hospital discharge letters at the document level.Its dataset contains 2,500 training and 500 test letters, and system performance is evaluated with F1.Source: §2.5 · pp. 6–7
2.6 Task GutBrainIE
- GutBrainIE supports extraction and normalization of gut-brain-axis knowledge and expanded its dataset to more than 6,000 documents with concept-level annotations.Its four subtasks are NER, mention-level relation extraction, entity recognition and disambiguation, and concept-level relation extraction; development and test sets contain expert annotations only.Source: §2.6 · p. 7, Table 5 · p. 8
3 Overview of participation
Across BioASQ 2026’s six tracks, 87 distinct teams participated, with most focusing on one task. Systems ranged from retrieval and generation pipelines to task-specific extraction, coding, and summarization approaches.
3 Overview of participation
- 87 distinct teams participated across six tasks, with participation ranging from 45 teams in 14b to six in ELCardioCC.Four teams entered both 14b and Synergy14; most teams focused on one task.Source: §3 · p. 8
3.1 Task 14b
- 45 teams submitted 821 runs from 151 systems in Task 14b, using varied retrieval and generation methods including dense or hybrid retrieval and multi-stage RAG.Approaches also included reranking, query expansion, ontology-guided methods, LLMs, prompting, ensembles, and agentic workflows.Source: §3.1 · p. 8
3.2 Task Synergy 14
- Seven teams submitted 79 Synergy14 runs from 18 systems, using hybrid retrieval, evidence-grounded prompting, and RAG approaches.Four teams also participated in Task 14b.Source: §3.2 · p. 9
3.3 Task MultiClinSum-2
- Nine teams submitted 75 MultiClinSum-2 runs; English received 20 runs, Spanish 14, Portuguese nine, and Norwegian none.Baseline systems nevertheless provided coverage for Norwegian.Source: §3.3 · p. 9, Table 6 · p. 9
3.4 Task BioNNE-R
- Seven teams submitted BioNNE-R predictions, mostly using typed-marker pair classification with transformer encoders and postprocessing to control precision.Teams also explored synthetic training examples, LoRA fine-tuning of an 8B-parameter LLM, and a document-level relational graph model.Source: §3.4 · pp. 9–10
3.5 Task ELCardioCC
- Six ELCardioCC teams submitted 19 systems, often combining entity recognition, linking, and document-level ICD-10 classification.Approaches used Greek-specific or multilingual transformers, structured clinical knowledge, and ensembles or optimization strategies to address class imbalance and long-tail labels.Source: §3.5 · pp. 10–11
3.6 Task GutBrainIE
- GutBrainIE drew 18 teams, with 17 in NER, 14 in NERD, 11 in M-RE, and 10 in C-RE.Participants commonly used biomedical transformers and ensembles; NERD extended NER pipelines with entity linking, while relation extraction used entity-pair classification and some prompt- or RAG-based systems.Source: §3.6 · p. 11
4 Results
The six task sections report task-specific evaluations and varied system outcomes across biomedical question answering, summarization, relation extraction, coding, and information extraction. Results include competitive systems alongside task-specific constraints, such as retrieval difficulty and performance drops across more demanding evaluation settings.
4.1 Task 14b
- Task 14b Phase A document retrieval averaged 0.142±0.013 and snippet F1 0.095±0.008, below BioASQ 13's 0.323±0.08 and 0.244±0.021.In Phase A+, best factoid and list scores were lower overall than Phase B, while yes/no scores remained close; final results awaited completion of expert assessment.Source: §4.1 · pp. 11–13, Table 8 · p. 12, Table 9 · p. 12
4.2 Task Synergy 14
- In Synergy 14, four rounds supported answering 61 of 63 new questions (97%), and systems provided at least one ideal answer for 48 questions (73%).Round 1 had no answer-ready questions or answers; the task supported collecting relevant material and answers on developing biomedical topics.Source: Table 10 · p. 13, §4.2 · p. 13
4.3 Task MultiClinSum-2
- Team ixa-sum ranked first across most MultiClinSum-2 language sub-tracks, while consistency was the weakest LLM-judged clinical dimension.The evaluation combined automatic summarization metrics with LLM judging across faithfulness, completeness, fluency, and consistency; teams used both fine-tuning and inference-only strategies.Source: §4.3 · pp. 13–14, Table 11 · p. 15
4.4 Task BioNNE-R
- BioNNE-R macro-F1 scores clustered at 0.44–0.53; LODAC-NII led Russian (0.5283) and bilingual (0.5015), while ELiRF-UPV led English (0.5060).Mean macro-F1 among teams entering both Russian and English subtasks was 0.4858 and 0.4866, respectively; graph and typed-marker approaches performed comparably.Source: Table 12 · p. 16, §4.4 · p. 16
4.5 Task ELCardioCC
- In ELCardioCC, stanimeros achieved the top F1 (0.8667), with Georgios 1 close behind at 0.8638.The baseline BERT system had precision 0.9036 but lower recall; the results describe ensemble and correction-heavy methods as dominating the leaderboard.Source: §4.5 · pp. 16–17, Table 13 · p. 17
4.6 Task GutBrainIE
- GutBrainIE top micro-F1 was 0.87 for NER, 0.69 for NERD, 0.61 for M-RE, and 0.35 for C-RE, with all teams beating the baseline.The authors report a drop of over 40% from M-RE to C-RE, attributing it to errors accumulated across entity recognition and linking.Source: §4.6 · p. 17
5 Conclusions
BioASQ 14 brought together 87 teams across six tracks, spanning 15 languages, multiple document types, and specialized biomedical domains. Participating systems achieved competitive results, with some improving over baselines or previous state-of-the-art results, while task findings also identified areas for further work.
5 Conclusions
- Task 14b systems scored strongly on yes/no answer generation, even in Phase A+, while list and factoid answers were less consistent and benefited from expert-selected material.Phase A retrieval varied across batches, and performance dropped compared with task 13b when task 14b included more questions from new experts.Source: §5 · pp. 18–19
- Synergy14 results suggest current systems can support biomedical scientists seeking specialized information on developing problems, while further advances remain needed.Source: §5 · p. 19
- MultiClinSum-2 drew nine teams using diverse methods, and its enhanced evaluation combined lexical, semantic, and LLM-as-judge measures with notable results across all subtracks and assessment dimensions.Source: §5 · p. 19
- All BioNNE-R submissions beat the multilingual baseline, but best macro-F1 scores stayed near 0.50–0.53, leaving nested-entity relation extraction an open challenge.Source: §5 · pp. 19–20
- ELCardioCC results favored hybrid, ensemble, and knowledge-enhanced approaches, especially Greek-specific transformer backbones and imbalance optimization, over simpler BERT baselines generally.Source: §5 · p. 20
- GutBrainIE's strongest systems used supervised deep learning with pretrained biomedical models and ensembles; prompt-based and generative approaches generally scored lower.Source: §5 · p. 20
- BioASQ 14 drew 87 teams across six tracks, spanning 15 languages, multiple document types, and specialized domains including gut-brain interaction and cardiology.Several systems achieved competitive performance, and some improved over baselines or previous state-of-the-art results.Source: §5 · p. 20
- Future BioASQ directions include expanding QA datasets through community processes, involving more biomedical experts, and broadening task resources across document types, languages, and subdomains.Source: §5 · p. 20