Authors:
Liliya Makhmutova
1
;
Giancarlo Salton
2
;
Fernando Perez-Tellez
1
and
Robert Ross
1
Affiliations:
1
Technological University Dublin, School of Computer Science, 191 North Circular Road, Dublin, Ireland
;
2
Unochapecó, Servidão Anjo da Guarda, 295-D - Efapi, Chapeco, Brazil
Keyword(s):
Medical Texts Simplification, LLM Evaluation.
Abstract:
Doctors and patients have significantly different mental models related to the medical domain; this can lead to different preferences in terminology used to describe the same concept, and in turn, makes medical text often difficult to understand for the average person. However, getting access to a good understanding of patient notes, medical history, and other health-related documents is crucial for patients’ recovery and sticking to a diet or medical procedures. Large language models (LLM) can be used to simplify and summarize text, yet there is no guarantee that the output will be correct and contain all the needed information. In this paper, we create and propose a new multi-modal medical text simplification dataset with pictorial explanations following along the aligned simplified and use it to evaluate the current state-of-the-art large language model (SOTA LLM) for the simplification task for the dataset and compare it to human-written texts. Our findings suggest that the curre
nt general-purpose LLMs are still not reliable enough for such in the medical sphere, though they may simplify texts quite well. The dataset and additional materials may be found at https://github.com/ LiliyaMakhmutova/medical texts simplification.
(More)