United States

Medical query rewriting transforms verbose
consumer health questions into concise clin-
ical queries, a critical step in health infor-
mation retrieval. Large language models
(LLMs) perform well on this task by stan-
dard metrics, yet high ROUGE or BERTScore
does not guarantee preservation of clinical
content. To address this issue, we intro-
duce MedFaith-F1, a category-level faithful-
ness metric over four clinically salient cate-
gories: clinical problems, medications, proce-
dures, and follow-up intent. We further propose
a hybrid Evidence and Knowledge-Grounded
Retrieval-Augmented Generation (EKG-RAG),
an evidence and knowledge-grounded frame-
work combining hybrid retrieval over PubMed
and MedlinePlus resources with UMLS (Uni-
fied Medical Language System)-aligned ontol-
ogy grounding. Evaluating LLMs LLaMA-
3 and Qwen2.5 across zero-shot, few-shot,
and QLoRA settings on MeQSum and med-
ical question-pair (MQP) datasets revealed
that base models exhibit category-level faith-
fulness failure rates (CHR) exceeding 40%,
invisible to standard metrics, while EKG-
RAG with QLoRA reduces CHR to 26.75%,
achieving MedFaith-F1 of 0.73. Our find-
ings call for faithfulness-aware evaluation
in clinical query rewriting, and MedFaith-
F1 provides a reproducible step in that di-
rection.

ACL 2026

Beyond Lexical Similarity: Evaluating Faithfulness in LLM-Based Medical Question Reformulation

medical question reformulation

knowledge graph

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.

workshop paper

This paper presents a hybrid solution for a binary classification of medical PubMed articles created for identifying reports that associate clinical metadata with SARS-CoV-2 genomic sequences. The system is designed to catch the subtle distinction between reports of sequence-associated patient metadata and sentences where such metadata is either unrelated, irrelevant, or linked to previous studies. The biggest challenge is the fact that the medical dataset is highly imbalanced, consisting of only 13.3 % of medical reports labeled positive. Our system proposes a hybrid solution combining four approaches that include dual-evidence tagging, negation-aware suppression, semantic frame extraction, and adversarial training. All these approaches were tested on multiple models: BiomedBERT-baseabstract, BioLinkBERT-large, PubMedBERTbase-fulltext, followed by a best-subset ensemble search to obtain an F1 score of 0.786, setting a new benchmark and positioning the solution in 1st place in the competition.

MetaMiners at SMM4H-HeaRD 2026: A Semantic-StructuralKnowledge-Enriched Ensemble for SARS-CoV-2 Metadata Identification

This paper describes the BIT.UA sys-
tem for the MultiClinNER shared task at
#SMM4H–HeaRD 2026, targeting multi-
lingual clinical named entity recognition
across seven languages for three entity types
(Disease, Procedure, Symptom). We extend
the Multi-Head CRF architecture, originally
developed for multi-class NER on Spanish
clinical text, to the multilingual setting. To
enable joint multi-entity training despite per-
entity text variations in the dataset, we develop
an adaptive text consolidation pipeline that
preserves over 94% of annotations. Our central
finding is that a single xlm-roberta-large model,
trained jointly on all seven languages and
three entity types, achieves competition rank 2
for five of seven languages, outperforming
dedicated monolingual models by up to +6.94
F1 points, while requiring only a single set
of weights. Ensembling multiple seeds of
this model achieves rank 1 for those five
languages, and combining it with monolingual
models yields rank 1 for the remaining
two.

BIT.UA at #SMM4H–HeaRD 2026: Towards Multi-Class Multilingual
Clinical Entity Recognition with Multi-Head CRF Ensembles

Word-level cross-lingual alignment is central to
annotation projection, translation auditing, and
cross-lingual faithfulness estimation, yet exist-
ing neural aligners are rarely adapted to special-
ized domains. In this paper, we introduce Clin-
icalAligner26AM, a large-context multilingual
aligner model for biomedical and clinical text
initialized from ClinicalEncoder26AM (Remy,
2026b). Our training recipe is inspired by
AWESoME Align (Dou and Neubig, 2021). We
build our soft alignment target by sharpening
with Sinkhorn–Knop optimal transport a cost
matrix established for parallel clinical texts and
conversations through the fusion of sentence-
level, phrase-level, and token-level signals. We
distill this sharpened alignment matrix directly
into our student aligner, by encouraging its
naive cosine-based token similarity scores to
match this target. At inference time, we project
source-span scores through the learned token
alignment matrix and decode the longest valid
high-scoring span in the target text, optionally
supported by MultiClinNER predictions sum-
marized in Appendix B. We evaluate CA26AM
on the MultiClinCorpus shared task, which
projects Spanish clinical entity annotations into
six target languages (Gallego-Donoso et al.,
2026; Lopez-Garcia et al., 2026). Our two sub-
mitted systems ranked respectively first and
second across all languages and entity types,
with character-weighted F1 scores above 0.95
in nearly all settings.

Parallia at #SMM4H-HeaRD 2026:
ClinicalAligner26AM: A Cross-Lingual Aligner for Dataset Translation;
Evidences from the MultiClinCorpus Shared Task

We describe two system submissions to Task 4
of the SMM4H-HeaRD 2026 Shared Task
on automated SOAP note generation from
doctor–patient dialogues. Our first submis-
sion is a standalone entity-conditioned gen-
eration model: Mistral-7B-Instruct-v0.1 fine-
tuned with QLoRA on 8,529 MedSynth train-
ing dialogues, where both training and infer-
ence prompts include clinical entities extracted
and grouped by SOAP section. Our second
submission is a four-stage modular pipeline
that additionally incorporates a hybrid retrieval
stage and a rule-based verification stage. The
key finding of this work is that incorporating
structured clinical domain knowledge, in the
form of NER entities grouped by SOAP sec-
tion, directly into the generation prompt pro-
duces consistent and reliable improvements
over dialogue-only generation. Our four-stage
pipeline submission achieved an average score
of 0.54 on the official test set, ranking first on
the shared task leaderboard.

NU_DeepHealthNLP at #SMM4H-HeaRD 2026: Entity-Conditioned
Generation and a Four-Stage Pipeline for Automated SOAP Note
Generation

TNM cancer staging is a critical process for
characterizing tumor burden and guiding clini-
cal decisions. Nevertheless, its automated ex-
traction remains challenging due to the unstruc-
tured and heterogeneous nature of free-text
pathology reports. This paper describes the par-
ticipation of the URJC-Team in Task 6 of the
Social Media Mining for Health/Health Real-
World Data (#SMM4H-HeaRD) 2026 Shared
Tasks. It focuses on predicting TNM stag-
ing from pathology reports. The proposed
workflow combines hand-crafted regular ex-
pressions with a Large Language Model (LLM).
First, explicit TNM mentions are extracted us-
ing rule-based patterns. Then, any stage not
recovered by these rules is inferred by an LLM.
Overall, the proposal provides competitive re-
sults across all official shared-task phases.

URJC-Team at #SMM4H-HeaRD 2026: TNM Stage Extraction with a
Regex-LLM Workflow

Six tasks from the SMM4H–HeaRD 2026
workshop were addressed with task-specific
large-language-model (LLM) pipelines relying
on prompt engineering, strict structured (JSON)
responses, and deterministic rule sets. The
pipelines utilize no task-specific fine-tuning
and can be adapted across diverse clinical and
social media data. This study demonstrates that
general-purpose LLMs (gpt-5.4-mini and gpt-
5.4) can accurately extract and classify crucial
health information when constrained by strict
output schemas. Notably, our hybrid approach
achieved the best overall performance among
all participating systems for Task 2 (Insomnia
Detection).

MedMind AI at #SMM4H-HeaRD 2026: Data Extraction and Generation Using Prompt Engineering and Structured Outputs (Tasks 1–6)

Detecting insomnia from clinical narratives re-
quires both accurate classification and clini-
cally grounded reasoning with interpretable evi-
dence. We present our systems for the SMM4H-
HeaRD 2026 shared task, which leverages
MIMIC-III notes annotated with rule-based in-
somnia criteria and supporting evidence spans.
We explore two complementary approaches:
parameter-efficient fine-tuning of lightweight
models using QLoRA and LoRA, and few-shot
prompting of large language models for joint
reasoning and evidence extraction. Our best
system achieves an F1-score of 0.7333 on bi-
nary classification and a micro-F1 of 0.6535 on
multi-label rule prediction, with up to 0.5192
partial-match F1 for evidence extraction. Re-
sults show that lightweight fine-tuned models
can outperform larger models in classification,
while larger models demonstrate stronger rea-
soning but struggle with precise span localiza-
tion, highlighting a key gap in clinically inter-
pretable NLP systems.

A3S@C-DAC at #SMM4H-HeaRD 2026: Reasoning Meets Evidence:
LLMs for Interpretable Insomnia Detection with Evidence Extraction in
Clinical Notes

This paper describes the Gladiators system for
Task 1 of the SMM4H 2026 shared task on bi-
nary classification of adverse drug event (ADE)
mentions in multilingual social media posts.
Our system fine-tunes three XLM-RoBERTa-
large models with different random seeds us-
ing focal loss (α=0.75, γ=2.0) and 3×pos-
itive oversampling, then averages their pre-
dicted probabilities and applies per-language
threshold optimization. On the development
set, our ensemble achieves a pooled binary F1
of 0.7505. On the official test set—which in-
troduced surprise Farsi comprising 35.5% of
samples—our system achieves F1 = 0.6039,
above the competition mean (0.5465) and me-
dian (0.5798). We evaluated eleven approaches,
document key negative results including catas-
trophic translation augmentation failures, and
analyze precision–recall trade-offs under se-
vere class imbalance. Post-evaluation, a six-
model cross-regime ensemble improved dev F1
to 0.7585.

Gladiators at #SMM4H–HeaRD 2026: Multi-Seed XLM-RoBERTa
Ensemble with Focal Loss
and Per-Language Threshold Optimization for Multilingual Adverse Drug
Event Detection

This paper describes the participation of the
LSI_UNED team in the firt sub-task of Mul-
tiClinAI at the #SMM4H-HeaRD 2026 Work-
shop, which focuses on multilingual clinical
named entity recognition in seven languages.
The task requires identifying mentions of dis-
eases, procedures, and symptoms in clinical
case reports. We propose a set of systems
based on the W2NER architecture, with a sepa-
rate model trained for each language and entity
type. For Spanish, we use a RoBERTa-based
model with data augmentation from additional
NER resources, while English and Italian sys-
tems are based on different biomedical BERT
variants. Results show consistent performance
across languages, with the best overall results
obtained for Spanish. Data augmentation im-
proves recall and F1, while English and Italian
models achieve competitive but slightly lower
scores. Symptom recognition remains the most
challenging entity type across all languages.

LSI_UNED at #SMM4H–HeaRD 2026: Grid-Based Biomedical Named
Entity Recognition Across Languages and Entity Types

We describe the LLATMU systems submitted
to the #SMM4H–HeaRD 2026 shared tasks,
covering two related clinical text structuring
problems: dialogue-to-SOAP note generation
(Task 4) and TNM staging classification from
pathology reports (Task 6). Although the two
tasks differ in modeling paradigm—text gen-
eration versus supervised classification—both
require transforming unstructured clinical nar-
ratives into structured representations.
For Task 4, we instruction-tuned LLMs with
parameter-efficient adaptation and submitted
a QLoRA-based Ministral-3B system, achiev-
ing an official blind test average score of 0.53
and outperforming the task-wide mean and me-
dian. For Task 6, we formulate TNM predic-
tion as a three-head classification problem us-
ing BioClinical-ModernBERT-large with long-
context encoding, class-weighted loss, and
normalized partial-label training. The model
achieves a validation average macro-F1 of
0.9196 and continues to outperform the offi-
cial baseline on the more challenging tie-break
test set.
Across both tasks, our results suggest that ro-
bust data handling, stable fine-tuning, and task-
appropriate supervision are important for prac-
tical clinical NLP under constrained and imper-
fect shared-task settings.

Content not yet available

Downloads

Next from ACL 2026

MetaMiners at SMM4H-HeaRD 2026: A Semantic-StructuralKnowledge-Enriched Ensemble for SARS-CoV-2 Metadata Identification

Similar lecture

MED-COPILOT: A Medical Assistant Powered by GraphRAG and Similar Patient Case Retrieval

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

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

MetaMiners at SMM4H-HeaRD 2026: A Semantic-StructuralKnowledge-Enriched Ensemble for SARS-CoV-2 Metadata Identification

Similar lecture

MED-COPILOT: A Medical Assistant Powered by GraphRAG and Similar Patient Case Retrieval

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads