United States

Large language models achieve strong performance on biomedical question answering and summarization benchmarks, yet traditional evaluation metrics often fail to detect clinically significant factual errors. We introduce a unified evaluation framework that combines reference-based measures with evidence-grounded factuality verification to assess biomedical text generation. Evaluating four open-source models across three benchmarks (BioASQ, PubMedQA, MedLFQA), we find that 13.4–24.7% of generated claims are contradicted and 23–41% are unsupported, despite high lexical overlap scores. Our proposed Fact-Aligned Score (FAS) correlates strongly with claim-level verifiability (rho=0.68), substantially outperforming ROUGE-L (rho=0.41). We release an open-source toolkit with model outputs and analysis scripts to support reproducible factuality evaluation and safer deployment of biomedical LLMs.

ACL 2026

Towards Unified Factuality Evaluation for Biomedical QA and Summarization: Aligning Metrics with Clinical Use-Cases

fas metric

evaluation metrics

text generation

**Welcome to ACL-IJCNLP 2021!**

The great event is jointly organized by the Association for Computational Linguistics (ACL) and Asian Federation of Natural Language Processing (AFNLP). 

As in previous years, the program of the conference includes a poster session, tutorials, workshops and demonstrations in addition to the main conference.


We were able to keep the registration fees similar to those charged for the virtual ACL 2020. The one fee allows attendance at the main conference and any/all tutorials and workshops. These fees would be $125 Regular Early and $175 Regular Late; $50 Student Early and $75 Student Late. Early registration closes at midnight July 11, 2021 (Eastern Daylight Time).

**Reminder:** It is ACL’s policy that at least one author of each accepted paper (including ACL Finding papers) must register for the conference.

**Reminder2:** Underline site will open closer to the event. If you already registered you will receive access detail

Registration is now open

ACL-IJCNLP 2021

The great event is jointly organized by the Association for Computational Linguistics (ACL) and Asian Federation of Natural Language Processing (AFNLP).

workshop paper

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

Seizure freedom is a key clinical outcome for people with epilepsy (PWE) yet it is primarily recorded in free-text notes and letters in the United Kingdom, making it difficult to aggregate and track at scale. This paper introduces a generative LLM-based pipeline boosted by synthetic data to identify a PWE's seizure freedom status in clinicians' records. We fine-tuned seven different LLMs with between 4-14 billion parameters using LoRA to compare models trained on synthetic records against those trained on expert annotated records. The best performing configuration, based on Qwen-2.5-14B, was trained entirely on synthetic records and used chain-of-thought (CoT) reasoning (both generated by GPT-5). This achieved an F1 score of 0.90±0.02 on double-annotated test data and outperformed the equivalent model trained on authentic clinician records, which achieved 0.87±0.04. The synthetically trained models also have the benefit of outputting their CoT reasoning process for greater decision-making transparency and can also make use of the unused supervised training data for significantly increased test examples. This work has implications for monitoring a key treatment outcome for PWE automatically and at scale.

Using Synthetic Records to Improve Automated Identification of Seizure Freedom in Clinical Text about People with Epilepsy

This paper studies how to improve biomedical named entity recognition (NER) using large language models (LLMs), especially for low-resource languages like Bangla and Basque. The main goal is to understand how different prompt styles and output formats affect model performance.
 
 The study finds that the way we design prompts is very important. Among all methods, question-style prompting works best across all languages. It helps the model understand the biomedical task more clearly and improves accuracy. In fact, improvements are much greater in Bangla and Basque compared to high-resource languages like English and Spanish.
 
 Another key finding is about the output format. Traditional BIO tagging (labeling each word) performs poorly with LLMs because it is strict and sensitive to small errors. Instead, span-based extraction (directly extracting text phrases) works much better and gives higher F1 scores. This is because LLMs naturally generate text spans rather than token-level labels.
 
 The paper also analyzes errors. Common problems include hallucination, missing entities, and boundary mistakes. Translation-based prompts can reduce hallucination, while question-style prompts reduce empty outputs in biomedical NER.
 
 Overall, the study shows that choosing the right prompt and output format is very important, especially for low-resource high-vocabulary languages. It provides useful guidance for building better multilingual medical information extraction systems.

Analyzing Prompt Design Choices in Biomedical Information Extraction for Low-Resource Languages

Extracting structured knowledge from unstructured text is a fundamental challenge in machine learning, particularly for concepts organized within complex hierarchical ontologies. In genomics, identifying phenotypes from clinical narratives is crucial for diagnostic precision, yet current methods struggle with contextual interpretation and subtle clinical descriptions.
 We present a hierarchy-aware workflow for ontology-based phenotype linking that combines semantic and hierarchical signals. Our approach integrates Large Language Models for span detection with retrieval and a hybrid reranking strategy using both Euclidean (semantic) and hyperbolic (hierarchical) embeddings trained on the Human Phenotype Ontology.
 We show that while hyperbolic embeddings alone do not outperform standard semantic retrieval, they provide complementary structural signals that improve performance over strong baselines when combined with Euclidean representations. In particular, the hybrid approach outperforms existing state-of-the-art methods and yields more hierarchically coherent predictions, especially in settings involving implicit phenotype mentions.
 Experiments on a public benchmark (ID-68) and a newly released clinical dataset (CHU-50), publicly released with code and data, highlight both performance gains and improved alignment with ontology structure. We further introduce a hierarchy-aware evaluation framework that reflects clinical relevance beyond exact-match metrics.

Hierarchy-Aware Hyperbolic and Semantic Reranking for Ontology-Based Phenotype Linking

The MedCPT model has demonstrated that strong biomedical retrievers can be trained using proprietary PubMed search logs. In this work, we study whether freely available citation sentences are sufficient to train similarly effective models. We construct a large-scale training dataset of ~ 62 million citation sentence-abstract pairs extracted from PubMed Central. We train a lightweight BERT-based retriever-reranker model called CiteRec on this dataset and evaluate it across three benchmark settings: (a) the biomedical subset of BEIR for information retrieval, (b) SciRepEval for generalizable scientific document embeddings, and (c) CitancePlus, a new set of ~ 90 thousand citation sentence-abstract pairs for PubMed-scale citation recommendation.
 
 We show that CiteRec performs competitively with MedCPT on the biomedical BEIR subset and outperforms it on SciRepEval. On CitancePlus, CiteRec achieves strong performance for citation recommendation over the full PubMed corpus, outperforming both MedCPT and a substantially larger Qwen3-Embedding-8B retriever.

Training Biomedical Retrievers From Large-Scale Citation Contexts

Standard clinical Natural Language Processing (NLP) benchmarks often yield inflated metrics by forcing deterministic classification on ambiguous instances, thereby obscuring the clinical risks of overconfident predictions. To bridge this gap, we propose a risk-aware hybrid selective classification framework, evaluated on early Human Immunodeficiency Virus suspicion identification in Spanish clinical notes. Our dual-verification approach explicitly decouples aleatoric uncertainty through Mondrian conformal prediction and epistemic uncertainty using a Multi-Centroid Mahalanobis Distance veto. Empirical evaluations reveal that standard uncertainty metrics and baseline classifiers are structurally insufficient for safe medical triage, suffering severe coverage collapse when forced to operate under strict reliability constraints. In contrast, by demanding that clinical narratives pass both probabilistic and geometric safeguards, the proposed framework successfully isolates a highly trustworthy operational domain.The obtained results show that explicit, decoupled uncertainty quantification is essential for translating biomedical NLP into responsible clinical practice.

Reliable Automated Triage in Spanish Clinical Notes: A Hybrid Framework for Risk-Aware HIV Suspicion Identification

Retrieval strategy selection is a critical but understudied design decision in biomedical RAG systems. Existing evaluations rely on lexical metrics that miss answer grounding, or require proprietary infrastructure that limits reproducibility. We present BioRAG, a head-to-head ablation of seven retrieval strategies on BioASQ-13b, evaluated using four RAGAs metrics with a locally deployed judge at zero monetary cost. Hybrid BM25 plus dense retrieval with Reciprocal Rank Fusion achieves faithfulness of 0.534 and context recall of 0.507, improvements of 50% and 85% over naive dense retrieval, confirmed across three random seed re-samples. HyDE improves faithfulness by 14% but reduces context precision by 52%, a tradeoff not previously documented on BioASQ. No single strategy dominates all four metrics, indicating that strategy selection must be application-driven. Sensitivity analysis across k in {3,5,10} confirms ranking stability. A domain mismatch diagnostic confirms 2% corpus coverage failure. The full pipeline runs on consumer hardware without paid APIs, directly addressing BioNLP 2026's emphasis on reproducibility and evaluation frameworks for health-related applications.

BioRAG: A Systematic Ablation Study of Retrieval Strategies for Biomedical Question Answering

Drug repurposing methods rely heavily on knowledge graph (KG) embeddings, but building and curating these graphs takes considerable effort. We present two findings on the Hetionet drug-disease benchmark and an epilepsy ranking task. First, PubMedBERT text embeddings, fed through the same downstream classifiers and identical 10-fold splits as four re-trained KG baselines (TransE, ComplEx, DistMult, RotatE), reach AUROC $0.910$, above all four (best: RotatE, $0.854$); a Random Forest on the same vectors scores $0.880$. The comparison is asymmetric in one important way: PubMedBERT was pretrained on the literature Hetionet was curated from, so the result is best read as ``text-with-literature-supervision vs.graph-only,'' and a head-to-head with text-augmented KG methods (KG-BERT, TxGNN) is left as follow-up. Second, across all seven combinations of text, molecular (ECFP4), and gene expression (LINCS L1000) features, cross-attention fusion of weaker modalities into text consistently degrades performance, despite a gated mechanism intended to suppress unhelpful modalities; the residual path forces the strong modality to absorb noise. The model also ranks proconvulsants (amoxapine, flumazenil) near the top, because text embeddings encode strength of association with a disease but not its direction.

Beyond Knowledge Graphs: PubMedBERT Embeddings as a Competitive Standalone Modality for Drug Re-purposing

Clinical NLP systems are increasingly used for triage support, prediction, and decision assistance in EHR-based settings, where demographic fairness is a critical concern. A common evaluation approach is counterfactual demographic perturbation: modifying attributes such as age or sex while holding clinical evidence fixed and measuring output changes. 
 However, we show that such counterfactual audits can be misleading when interpreted in isolation. Across three clinical LLMs, we find that non-demographic control perturbations (e.g., paraphrases) often induce output variability comparable to or greater than demographic edits. 
 This can contribute to overestimation or misinterpretation of demographic bias.
 To address this, we propose a baseline-aware audit framework that explicitly compares demographic perturbations against control baselines. 
 Our analysis reveals that (i) label-level stability can mask substantial variation in generated rationales and recommendations, and (ii) age-based perturbations generally induce larger effects than sex-based ones in borderline cases. 
 Crucially, we identify a high intrinsic instability ("noise floor"; 0.46--0.71 Jaccard instability) in clinical LLM generations, while additional matched-metric analyses show that demographic perturbations are often comparable to non-demographic baseline variability.
 These findings highlight a key limitation of existing fairness evaluations: without establishing appropriate baselines, apparent demographic sensitivity may be over- or mis-attributed to bias rather than broader generative instability. We argue that baseline-aware counterfactual audits, which explicitly compare demographic effects against intrinsic model noise, provide a more reliable lens for evaluating clinical NLP systems in high-stakes settings.

When Demographic Sensitivity Isn't What It Seems: Baseline-Aware Counterfactual Audits for Clinical NLP

MedCAT is an open-source framework for clinical named entity recognition and linking (NER+L) widely used in research and healthcare settings. We present MedCAT v2, a re-engineered version designed to improve modularity, extensibility, and maintainability while preserving the core functionality and performance of previous releases. The new architecture introduces a registry-based component system and a flexible pipeline that enables easy substitution of components, integration of alternative methods, and future expansion, including support for pre-trained components across the full NER+L and contextualisation workflow. This enables systematic exploration of clinical NER+L design trade-offs by evaluating different components in the pipeline. Evaluation across multiple public datasets shows equivalent or improved performance compared to earlier versions, with reduced integration overhead and improved runtime flexibility. The framework also supports optional extensions such as meta-annotation, relation extraction, providing a unified and reproducible environment for clinical NLP in real-world settings.

MedCAT v2: a modular, extensible architecture for clinical named entity recognition and linking under real-world privacy and compute constraints

The development of large language models (LLMs) has led to increased focus on their adaptation to specialized domains and languages, yet the effectiveness of domain adaptation strategies remains unclear. We present a study of medical domain adaptation using French medical question answering (QA) as a case study. We compare continual pretraining (CPT), supervised fine-tuning (SFT), and their combination across three model families, multiple sizes, and three initialization types, explicitly disentangling adaptation effects from base model choice.
 We evaluate both multiple-choice (MCQA) and open-ended QA (OEQA) under greedy and constrained decoding using automatic metrics and LLM-as-a-Judge evaluation. For MCQA, CPT+SFT most often achieves the best scores, but gains over SFT are small and frequently not statistically significant, making SFT a strong and cost-effective default. For OEQA, CPT consistently improves overlap-based metrics, while SFT often degrades generation quality; instruction tuning and CPT+SFT are preferred by LLM-based evaluation. Cross-lingual experiments further show effective transfer from French adaptation to English benchmarks. Overall, we provide practical guidelines for selecting adaptation strategies under computational constraints.

Downloads

Next from ACL 2026

Using Synthetic Records to Improve Automated Identification of Seizure Freedom in Clinical Text about People with Epilepsy

Similar lecture

GO FIGURE: A Meta Evaluation of Factuality in Summarization