United States

We investigate how tokenization granularity affects the representation of medical terminology in language models. Prior work links tokenization granularity to downstream performance under contextualized settings for specifically pretrained and fine-tuned models. We instead ask whether this relationship already emerges at the level of isolated term representations across existing pretrained models. We introduce an intrinsic definition retrieval task using UMLS term-definition pairs, with comparison to WordNet. We show that despite substantially heavier fragmentation of medical terminology, the models remain relatively robust in maintaining semantic alignment between medical terms and their definitions. At the same time, tokenization granularity still correlates with retrieval performance, indicating that effects previously observed in downstream biomedical tasks are already reflected at the level of isolated term representations. Encoder models benefit primarily from whole-token preservation, while for decoder LLMs, tokenization effects emerge mainly at deeper retrieval ranks.

ACL 2026

Tokenization Granularity and Medical Term Representations in Language Models

medical term representations

tokenization granularity

language models

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.

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.

workshop paper

Clinical dialogue-to-note generation is challenging because clinically salient evidence is noisy, distributed across turns, and often revised later in the encounter. Direct transcript-only prompting and coarse intermediate scaffolds can therefore suffer from omissions, section leakage, unsupported fill-in, and brittle final-state tracking.
 
 We propose Clinical Atomic Propositions (CAPs), a dialogue-aware intermediate representation for faithful clinical note generation. CAPs extract source-grounded clinical assertions while preserving modifiers such as verification status, temporality, speaker/source, and action type. We also study an optional event consolidation layer that groups CAPs into problem-oriented care bundles before note rendering.
 
 We evaluate five methods on a 197-case ACI-Bench cohort: a transcript-only baseline, prompt-based reimplementations of Cluster2Sent and MEDSUM-ENT, CAP, and CAP+Event. The main task uses a sectioned-note template, with SOAP-template rendering and transcript-free rendering reported as ablations. We use MEDSUM-ENT-style GPT-R/P/F1 metrics and a proposition-grounded semCAP-R/P/F1 audit to measure concept-level and source-grounded faithfulness, complemented by case-level win/tie/loss analysis and clinician deep review.
 
 Results show that CAP improves preservation of transcript-grounded clinical propositions while remaining competitive on concept-level GPT metrics. CAP+Event is not uniformly better than CAP, but qualitative and boundary analyses show when problem-oriented consolidation can improve organization and when compression can introduce omissions. We release code, prompts, intermediate representations, generated notes, and evaluation artifacts at a public repository.

CAP: A Source-Grounded Proposition Scaffold for Faithful Clinical Dialogue-to-Note Generation

Clinical reasoning over electronic health records (EHRs) involves heterogeneous operations, including text interpretation, numerical computation, temporal filtering, and guideline-based aggregation. However, many existing LLM-based approaches still cast these heterogeneous operations as a single end-to-end generation process, obscuring their different reliability requirements and making intermediate failures difficult to inspect. We therefore propose a framework based on operation-mechanism alignment that represents clinical reasoning as a directed acyclic graph of typed operations, where each node is assigned to the execution mechanism best suited to its reliability requirements. The framework also preserves structured evidence provenance for intermediate results. Across six clinician-annotated binary decision tasks, the framework outperforms direct prompting, single-step retrieval-augmented prompting, and chain-of-thought baselines, supporting operation-mechanism alignment as a practical design principle for reliable clinical reasoning over EHRs.

Operation-Mechanism Alignment for Reliable Clinical Reasoning over Electronic Health Records

Despite Spanish being one of the most widely spoken languages in the world, biomedical NLP resources and systematic evaluations remain limited relative to English. We address this gap by constructing and releasing two Spanish biomedical corpora: (1) **MeSHClass-ES**, a 29,063 abstract bilingual corpus translated from PubMed with Opus-MT, and (2) **AnatEM-ES**, the AnatEM anatomical entity corpus translated with a chunk-level LLM-based pipeline that jointly preserves BIO annotations across 13,849 entity mentions. Both corpora achieve a mean COMET score of 0.73 despite using different translation systems. We benchmark nine encoder models spanning general-domain Spanish, domain-specific, and multilingual architectures for both tasks. RigoBERTa-2.0 leads both tasks (micro-F1 classification 0.69, tied with SciBETO-large; NER F1 0.66). Both domain pretraining and model capacity drive performance, with the gap slightly more pronounced for NER (4-point spread) than classification (3-point spread). XLM-RoBERTa-large emerges as a competitive multilingual baseline. A parallel evaluation of four open-weight decoders (7–9B) reveals a task-dependent encoder-decoder gap: QLoRA-adapted Gemma-2-9B reaches 88% of the best encoder on classification (micro-F1 .61 vs .69), but for NER every decoder configuration we tested stays at or below 40% of the best encoder F1. We release both corpora on the HuggingFace Hub1, translation pipelines, and evaluation code on GitHub.

MeSHClass-ES and AnatEM-ES: Open Resources for Spanish Biomedical NLP

When Evidence Conflicts: Uncertainty and Order Effects in Retrieval-Augmented Biomedical Question Answering

Clinical sources and patient-authored reviews often describe antidepressant side effects in different ways, but these differences are rarely measured directly. We present ClinPeer-AE, a linked dataset for comparing side-effect evidence from PubMed, ClinicalTrials.gov, WebMD, and Drugs.com while preserving source identity. Across five widely prescribed antidepressants, we find low overlap between clinical and peer sources, large differences in relative emphasis, and evidence that many peer-only effects also appear in U.S. Food and Drug Administration Adverse Event Reporting System (FAERS) reports. These findings suggest that patient reviews provide useful context about recurring medication experiences and offer a complementary view of how side effects are described outside formal clinical settings.

From Trials to Reviews: A Cross-Source Dataset for Antidepressant Side Effects

Retrieval-augmented generation (RAG) holds promise for clinical question answering over electronic health records (EHRs), but existing systems treat retrieval as an opaque subroutine, limiting auditability and reliability in patient care workflows. 
 
 We introduce a deterministic multi-stage retrieval pipeline for longitudinal EHR question answering that decomposes retrieval into four distinct, ablated stages where each stage is instrumented with diagnostic metrics, making the flow of clinical evidence measurable and auditable at every step. Evaluated on a broad LLM-annotated cohort and an expert-annotated cardiovascular benchmark developed alongside clinicians from real ICU records, the full pipeline achieves 22-23% relative recall gain over a strong dense retrieval baseline across both cohorts, with consistent improvements in downstream answer quality. 
 
 The pipeline's deterministic and transparent design addresses a critical gap in clinical NLP: retrieval systems that clinicians and researchers can not only rely on, but inspect, audit, and build upon for real-world deployment.

A Deterministic Multi-Stage Retrieval Pipeline for Longitudinal EHR Question Answering

Transformer-based models such as PLM-CA achieve strong performance for automatic ICD coding, but their attention weights do not provide faithful explanations of their predictions. 
 This is a major limitation for electronic medical records, where users often need concise and trustworthy evidence for each assigned code. 
 To address this issue, we jointly train a sentence extractor and an ICD code classifier such that predictions are based only on the extracted sentences. 
 As a result, the extracted sentences serve as faithful rationales for each predicted code and substantially reduce the effort required to inspect long medical records. 
 Experiments on MIMIC-III show that our method approaches the performance of a transformer baseline that processes the full record while using only a small fraction of the document.

Interpretable ICD Code Classification with Faithful Sentence Extraction

Highly technical medical terms are difficult for patients to understand during fast-paced hospital consultations, leading them to rely on Large Language Models (LLMs) for simplified explanations. However, LLMs can produce inaccurate or false information. Since expert evaluation is costly and time-consuming, LLM-as-a-Judge (LaaJ) approach is increasingly adopted to assess the quality of LLM-generated text. In this paper, we investigate the reliability and robustness of LaaJ for specialized medical knowledge by evaluating six LLMs for their judgment capabilities on three dimensions: correctness, readability, and completeness. We utilized three judgment setups: Vanilla, Epistemic, and Bias to probe robustness, and assess them against human expert annotations to measure alignment. To address the lack of specialized medical benchmarks, we introduce BrainCancerDB, an English dataset of 219 brain cancer terms with 23,652 annotations. Our findings indicate that while LLM-Judges and humans display similar trends in ranking simplified explanations, LLM-Judges tend to be more lenient on correctness, which may have serious implications in medical setting. Additionally, we observe that hallucinations in LaaJ setups can be mitigated by epistemic markers.

Evaluating LLM-as-a-Judge for Medical Term Simplification

Large Language Models such as GPT-4o and GPT-5 achieve strong zero-shot performance on biomedical claim verification, but cost and opacity limit scalable use. We fine-tune three small LLMs; Phi-3-mini (3.8B), Qwen2.5-3B, and Mistral-7B; via QLoRA on SciFact and HealthVer, providing the first study of QLoRA models against GPT-4o and fine-tuned BioLinkBERT encoders. Mistral-7B QLoRA achieves higher F1 than both GPT-4o and GPT-5 (up to 12% gain) at 44.5x lower cost using just 1,008 training examples, representing a compelling cost-quality trade-off. We conduct extensive in-domain and cross-domain evaluation: models trained on SciFact tested on HealthVer and vice versa, at matched sizes to isolate dataset structure from data quantity. We identify a previously unreported structural artifact in SciFact that inflates in-domain scores, and show through bidirectional out-of-domain evaluation that training on structurally sound data enables robust cross-domain transfer. We plan to release all code and adapter checkpoints.

Small LLMs for Biomedical Claim Verification: Cost-Effective Fine-Tuning, Structural Dataset Shortcuts, and Cross-Domain Generalization

Reliable biomedical and clinical retrieval requires more than strong ranking performance: it requires a practical way to find systematic model failures and curate the training evidence needed to correct them.
 Late-interaction models such as ColBERT provide a first solution thanks to the interpretable token-level interaction scores they expose between document and query tokens.
 Yet this interpretability is shallow: it explains a particular document--query pairwise score, but does not reveal whether the model has learned a clinical concept in a stable, reusable, and context-sensitive way across diverse expressions.
 As a result, these scores provide limited support for diagnosing misunderstandings, identifying irreasonably distant biomedical concepts, or deciding what additional data or feedback is needed to address this.
 In this short position paper, we propose Diagnosable ColBERT, a framework that aligns ColBERT token embeddings to a reference latent space grounded in clinical knowledge and expert-provided conceptual similarity constraints.
 This alignment turns document encodings into inspectable evidence of what the model appears to understand, enabling more direct error diagnosis and more principled data curation without relying on large batteries of diagnostic queries.

Content not yet available

Downloads

Next from ACL 2026

CAP: A Source-Grounded Proposition Scaffold for Faithful Clinical Dialogue-to-Note Generation

Similar lecture

Temporal Tokenization Strategies for Event Sequence Modeling with Large Language Models

Content not yet available

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

CAP: A Source-Grounded Proposition Scaffold for Faithful Clinical Dialogue-to-Note Generation

Similar lecture

Temporal Tokenization Strategies for Event Sequence Modeling with Large Language Models

Downloads