Loading...
Enhancing the Comprehension: Text Simplification Approaches and the Role of Large Language Models
Yang, Ziyu
Yang, Ziyu
Citations
Altmetric:
Genre
Thesis/Dissertation
Date
2024-05
Advisor
Committee member
Group
Department
Computer and Information Science
Subject
Permanent link to this record
Collections
Research Projects
Organizational Units
Journal Issue
DOI
http://dx.doi.org/10.34944/dspace/10208
Abstract
Radiology reports are highly technical documents aimed primarily at doctor-doctor communication. There has been an increasing interest in directly sharing radiology reports with patients to improve doctor-patient communication. However, radiology reports are primarily created for communication of imaging findings among medical professionals and are difficult to comprehend by laypeople. Thus, it would be useful if a patient-friendly version of a radiology report could be provided to patients in addition to the original report. Addressing this gap by creating patient-friendly versions of these reports not only enhances doctor-patient communication but also empowers patients by providing them with accessible information about their health conditions. However, a conundrum arises as requiring radiologists to augment traditional reports with patient-friendly summaries could exert negative influence on their cognitive load and productivity. Recent research worked on automatic simplification of health records, employing both lexical and semantic simplification methods, with recent advancements incorporating deep learning techniques. Yet, training these deep learning models for medical text simplification necessitates the collection of costly labeled data.
To address these challenges, we investigate the roles of Large Language Models (LLMs) in simplifying radiology reports to make them more accessible to patients. Firstly, we assess GPT-3's capacity to generate patient-friendly summaries of radiology sentences, specifically focusing on liver conditions. Our findings reveal that with appropriate prompting and fine-tuning, GPT-3 can produce high-quality simplifications, as evidenced by both automated metrics and manual evaluations by radiologists. Secondly, addressing the challenge of data scarcity in training models for medical text simplification, we introduce a novel data augmentation strategy. This approach leverages the generative capabilities of pre-trained LLMs to create simplified versions of unlabeled radiology sentences and employs paraphrasing techniques on labeled data, significantly enhancing the accuracy of our fine-tuned simplification model beyond baseline methods. Lastly, we explore the effectiveness of advanced prompting mechanisms, such as chain-of-thought and self-correction, in improving the quality of text simplifications. Through a dual-evaluation protocol involving both radiologists and laypeople, we demonstrate the superiority of self-correction prompting in producing simplifications that are both factually accurate and easier for laypeople to understand. Collectively, these studies underscore the potential of LLMs in bridging the gap between complex medical information and patient comprehension, providing valuable insights into the methodologies and evaluation frameworks that can facilitate the development of more accessible health communication tools.
Description
Citation
Citation to related work
Has part
ADA compliance
For Americans with Disabilities Act (ADA) accommodation, including help with reading this content, please contact scholarshare@temple.edu