United States

This paper presents an overview of the Perspective-aware Answer Summarization (PerAnsSumm) Shared Task on summarizing healthcare answers in Community Question Answering forums hosted at the CL4Health Workshop at NAACL 2025. In this shared task, we approach healthcare answer summarization with two subtasks: (a) perspective span identification and classification and (b) perspective-based answer summarization (summaries focused on one of the perspective classes). We
define a benchmarking setup for comprehensive evaluation of predicted spans and generated summaries. We encouraged participants to explore novel solutions to the proposed problem and received high interest in the task with 23 participating teams and 155 submissions. This paper describes the task objectives, the dataset, the evaluation metrics and our findings. We share the results of the novel approaches adopted by task participants, especially emphasizing the applicability of Large Language Models in this perspective-based answer summarization task.

NAACL 2025

Overview of the PerAnsSumm 2025 Shared Task on Perspective-aware Healthcare Answer Summarization

llms

shared task

healthcare

span

summarization

### Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Welcome to the 2025 meeting of the Nations of the Americas Chapter of the Association for Computational Linguistics! I am proud to help organize the first NAACL conference to carry the new name of our organization, one that emphasizes inclusion for all of the Americas. I am also pleased to welcome you to Albuquerque, New Mexico, a state whose unique blend of cultural influences will make for an excellent backdrop for NAACL 2025, especially with this year’s special theme on NLP in a Multicultural World. 
**[Continue reading...](https://drive.google.com/file/d/1jX-qGhqVSZZCIrAnJaz798pflu5Irrdn/view?usp=sharing)**

*- Colin Cherry, Google, NAACL 2025 General Chair* 

[![](https://assets.underline.io/markdown_image/1/image/b087f8a4dc5816d6e1a6514e59c59ac3.png)](https://drive.google.com/file/d/1T96GzPqObXrMln2BMByCSXTSizjTg69P/view?usp=sharing)

You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

To access NAACL 2025 event page you are required to register. Please follow [**this link**](https://2025.naacl.org/registration/registration/) to register. Access will depend on your registration type.

Please register!

Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

workshop paper

Healthcare community question-answering (CQA) forums provide multi-perspective insights into patient experiences and medical advice. Summarizations of these threads must account for these perspectives, rather than relying on a single "best" answer. This paper presents the participation of the WisPerMed team in the PerAnsSumm shared task 2025, which consists of two sub-tasks: (A) span identification and classification, and (B) perspectivebased summarization. For Task A, encoder models, decoder-based LLMs, and reasoningfocused models are evaluated under finetuning, instruction-tuning, and prompt-based paradigms. The experimental evaluations employing automatic metrics demonstrate that DeepSeek-R1 attains a high proportional recall (0.738) and F1-Score (0.676) in zero-shot settings, though strict boundary alignment remains challenging (F1-Score: 0.196). For Task B, filtering answers by labeling them with perspectives prior to summarization with Mistral-7B-v0.3 enhances summarization. This approach ensures that the model is trained exclusively on relevant data, while discarding non-essential information, leading to enhanced relevance (ROUGE-1: 0.452) and balanced factuality (SummaC: 0.296). The analysis uncovers two key limitations: data imbalance and hallucinations of decoder-based LLMs, with underrepresented perspectives exhibiting suboptimal performance. The WisPerMed team's approach secured the highest overall ranking in the shared task.

WisPerMed @ PerAnsSumm 2025: Strong Reasoning Through Structured Prompting and Careful Answer Selection Enhances Perspective Extraction and Summarization of Healthcare Forum Threads

Automated summarization of healthcare community question-answering forums is challenging due to diverse perspectives presented across multiple user responses to each question. The PerAnsSumm Shared Task was therefore proposed to tackle this challenge by identifying perspectives from different answers and then generating a comprehensive answer to the question. In this study, we address the PerAnsSumm Shared Task using two complementary paradigms: (i) a training-based approach through QLoRA fine-tuning of LLaMA-3.3-70B-Instruct, and (ii) agentic approaches including zero- and few-shot prompting with frontier LLMs (LLaMA-3.3-70B-Instruct and GPT-4o) and a Mixture-of-Agents (MoA) framework that leverages a diverse set of LLMs by combining outputs from multi-layer feedback aggregation. For perspective span identification/classification, GPT-4o zero-shot achieves an overall score of 0.57, substantially outperforming the 0.40 score of the LLaMA baseline. With a 2-layer MoA configuration, we were able to improve LLaMA performance up by 28% to 0.51. For perspective-based summarization, GPT-4o zero-shot attains an overall score of 0.42 compared to 0.28 for the best LLaMA zero-shot, and our 2-layer MoA approach boosts LLaMA performance by 32% to 0.37. Furthermore, in few-shot setting, our results show that the sentence-transformer embedding-based exemplar selection provides more gain than manually selected exemplars on LLaMA models, although the few-shot prompting is not always helpful for GPT-4o.

YaleNLP @ PerAnsSumm 2025: Multi-Perspective Integration via Mixture-of-Agents for Enhanced Healthcare QA Summarization

Large Foundation Models have displayed incredible capabilities in a wide range of domains and tasks. However, it is unclear whether these models match specialist capabilities without special training or fine-tuning. In this paper, we investigate the innate ability of foundation models as neurodegenerative disease specialists. Precisely, we use a language model, Llama-3.1, and a visual language model, Llama3-LLaVA-NeXT, to detect language specificity between Alzheimer's Disease patients and healthy controls through a well-known Picture Description task. Results show that Llama is comparable to supervised classifiers, while LLaVA, despite its additional "vision", lags behind.

On Large Foundation Models and Alzheimer's Disease Detection

The proliferation of wearable devices and sports monitoring apps has made tracking physical activity more accessible than ever. For individuals with Type 1 diabetes, regular exercise is essential for managing the condition, making personalized feedback particularly valuable. By leveraging data from physical activity sessions, NLP-generated messages can offer tailored guidance to help users optimize their workouts and make informed decisions. In this study, we assess several open-source pre-trained NLP models for this purpose. Contrary to expectations, our findings reveal that models fine-tuned on medical data or excelling in medical benchmarks do not necessarily produce high-quality messages.

A Preliminary Study on NLP-Based Personalized Support for Type 1 Diabetes Management

Sepsis is a leading cause of death in Intensive Care Units (ICU). Early detection of sepsis is crucial to patient survival. Existing works in the clinical domain focus mainly on directly predicting a ground truth label that is the outcome of a medical syndrome or condition such as sepsis. In this work, we primarily focus on clinical time series forecasting as a means to solve downstream predictive tasks intermediately. We base our work on a strong monomodal baseline and propose multimodal transformers using set functions via fusing both physiological features and texts in electronic health record (EHR) data. Furthermore, we propose hierarchical transformers to effectively represent clinical document time series via attention mechanism and continuous time encoding. Our multimodal models significantly outperform baseline on MIMIC-III data by notable gaps. Our ablation analysis show that our atomic approaches to multimodal fusion and hierarchical transformers for document series embedding are effective in forecasting. We further fine-tune the forecasting models with labelled data and found some of the multimodal models consistently outperforming baseline on downstream sepsis prediction task.

Multimodal Transformers for Clinical Time Series Forecasting and Early Sepsis Prediction

Clinical coding is a critical task in healthcare, although traditional methods for automating clinical coding may not provide sufficient explicit evidence for coders in production environments. This evidence is crucial, as medical coders have to make sure there exists at least one explicit passage in the input health record that justifies the attribution of a code. We therefore propose to reframe the task as an entity linking problem, in which each document is annotated with its set of codes and respective textual evidence, enabling better human-machine collaboration. By leveraging parameter-efficient fine-tuning of Large Language Models (LLMs), together with constrained decoding, we introduce three approaches to solve this problem that prove effective at disambiguating clinical mentions and that perform well in few-shot scenarios.

Explainable ICD Coding via Entity Linking

Chronic Kidney Disease (CKD) is a global health challenge, affecting 5–10% of the population, with a significant burden on healthcare systems. Early prediction of CKD progression from stage III to stage V is crucial to enable timely interventions. Traditional predictive methods rely on biochemical markers and demographic factors, but are often limited by issues such as missing data and reliance on structured inputs. This study explores the potential of several encoder-based language models, to predict CKD progression using a cohort from the Clinical Practice Research Datalink (CPRD) GOLD database. We applied both Full Fine-Tuning (FFT) and Parameter-Efficient Fine-Tuning (PEFT) with LoRA to pre-trained models, comparing them against traditional machine learning algorithms such as Random Forest and XGBoost. Our results show that fine-tuned models, particularly dmis-lab/biobert-v1.1-FFT, outperform traditional models in predicting CKD progression, with an AUC of 0.7787, precision of 0.7261, and accuracy of 0.7045. Although LoRA-based models are more computationally efficient, they consistenly exhibit lower performance. These findings suggest that fine-tuned encoder models hold significant potential for improving CKD progression prediction. However, there is still room for further enhancement in their accuracy and applicability in clinical settings.

Predicting Chronic Kidney Disease Progression from Stage III to Stage V using Language Models

Chronic pain affects millions, yet traditional assessments often fail to capture patients' lived experiences comprehensively. In this study, we used a Motivational Interviewing framework to conduct semi-structured interviews with eleven adults experiencing chronic pain and then applied Natural Language Processing (NLP) to their narratives. We developed an annotation schema that integrates the International Classification of Functioning, Disability, and Health (ICF) with Aspect-Based Sentiment Analysis (ABSA) to convert unstructured narratives into structured representations of key patient experience dimensions. Furthermore, we evaluated whether Large Language Models (LLMs) can automatically extract information using this schema. Our findings advance scalable, patient-centered approaches to chronic pain assessment, paving the way for more effective, data-driven management strategies.

Capturing Patients' Lived Experiences with Chronic Pain through Motivational Interviewing and Information Extraction

In fields like healthcare and pharmacovigilance, explainability has been raised as one way of approaching regulatory compliance with machine learning and automation.
This paper explores two feature attribution methods to explain predictions of four different classifiers trained to assess the seriousness of adverse event reports. On a global level, differences between models and how well important features for serious predictions align with regulatory criteria for what constitutes serious adverse reactions are analysed. In addition, explanations of reports with incorrect predictions are manually explored to find systematic features explaining the misclassification.

We find that while all models seemingly learn the importance of relevant concepts for adverse event report triage, the priority of these concepts varies from model to model and between explanation methods, and the analysis of misclassified reports indicates that reporting style may affect prediction outcomes.

Explainability for NLP in Pharmacovigilance: A Study on Adverse Event Report Triage in Swedish

The healthcare industry has accumulated vast amounts of clinical data, much of which has traditionally been unstructured, including medical records, clinical data, patient communications, and visit notes. Clinician-patient conversations form a crucial part of medical records, with the resulting medical note serving as the ground truth for future interactions and treatment plans. Generating concise and accurate SOAP notes is critical for quality patient care and is especially challenging in specialty care, where relevance, clarity, and adherence to clinician preferences are paramount. These requirements make general-purpose LLMs unsuitable for producing high-quality specialty notes. While recent LLMs like GPT-4 and Sonnet 3.5 have shown promise, their high cost, size, latency, and privacy issues remain barriers for many healthcare providers.

We introduce SpecialtyScribe, a modular pipeline for generating specialty-specific medical notes. It features three components: an Information Extractor to capture relevant data, a Context Retriever to verify and augment content from transcripts, and a Note Writer to produce high quality notes. Our framework and in-house models outperform similarly sized open-source models by over 12% on ROUGE metrics.
Additionally, these models match top closed-source LLMs' performance while being under 1% of their size. We specifically evaluate our framework for oncology, with the potential for adaptation to other specialties.

Downloads

Next from NAACL 2025

WisPerMed @ PerAnsSumm 2025: Strong Reasoning Through Structured Prompting and Careful Answer Selection Enhances Perspective Extraction and Summarization of Healthcare Forum Threads

Similar lecture

DataHacks at PerAnsSumm 2025: LoRA-Driven Prompt Engineering for Perspective Aware Span Identification and Summarization

.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 NAACL 2025

WisPerMed @ PerAnsSumm 2025: Strong Reasoning Through Structured Prompting and Careful Answer Selection Enhances Perspective Extraction and Summarization of Healthcare Forum Threads

Similar lecture

DataHacks at PerAnsSumm 2025: LoRA-Driven Prompt Engineering for Perspective Aware Span Identification and Summarization

Downloads