MedSiML: A Multilingual Approach for Simplifying Medical Texts

Health literacy is crucial yet often hampered by complex medical terminology. Existing simplification approaches are limited by small, sentence-level, and monolingual datasets. To address this, we introduce MedSiML, a large-scale dataset designed to simplify and translate medical texts into the ten most spoken languages, improving global health literacy. This repository contains a sample of our 64k paragraph level medical text simplification dataset created from multiple sources using automated annotation process involving Gemini model and manual scrutiny. We also share our code for fine-tuning and evaluating the models. The entire dataset and model checkpoints were very large to be committed. They would be made available on DropBox post publication and the link would be shared here.

Table of Contents

• Introduction
• Dataset Overview
• Methodology
• Results
• Usage
• Contributors
• Citation
• License

Dataset Overview

MedSiML includes over 64,000 paragraphs from PubMed, Wikipedia, and Cochrane reviews, simplified into English, Mandarin, Spanish, Arabic, Hindi, Bengali, Portuguese, Russian, Japanese, and Punjabi. An additional super-simplified English version is available for those with learning disabilities.

• Sources: PubMed, Wikipedia, Cochrane Reviews
• Languages: English, Mandarin, Spanish, Arabic, Hindi, Bengali, Portuguese, Russian, Japanese, Punjabi
• Simplification Model: Flash-1.5 Gemini model
• Fine-tuning Model: Text-To-Text Transfer Transformer (T5) base model

Methodology

Data Collection

We compiled data from three main sources:

1. PubMed: 50,000 abstracts from biomedical articles.
2. Wikipedia: Biomedical articles sourced from Hugging Face 2022 Wikipedia corpus.
3. Cochrane Reviews: Derived from existing Cochrane datasets.

Data Cleaning

The collected data underwent rigorous cleaning to remove noise, ensure quality, and eliminate duplicates.

Annotation

Using the Flash-1.5 Gemini model, we simplified and translated the texts into ten languages and created a super-simplified English version.

Model Training

We fine-tuned the T5 base model on this paragraph-level, multilingual data, achieving significant improvements in readability and semantic similarity.

Results

Our fine-tuned model showed improvements over existing models in various metrics:

• ROUGE1: +10.61%
• SARI: +11.01%
• Semantic Similarity: +49.1%
• Readability Scores: FK +0.38, ARI +1.06

Usage

To use the MedSiML dataset and models, follow these steps:

1. Clone the Repository

   git clone https://github.com/nepython/MedSiML.git
   cd MedSiML

2. Installation

   pip install -r requirements.txt

3. Accessing the Dataset

   • Download the data.zip file from DropBox.
   • Move the data.zip into data directory.
   • Unzip the data.zip.
   • Then it can be loaded using datasets library

   from datasets import load_dataset
   
   raw_datasets = load_dataset("csv", data_files='data/data.tsv', delimiter='\t')

   • (or) Using pandas library

   import pandas as pd
   
   df = pd.read_csv('data/data.tsv', sep='\t')

4. Model Inference

   • Download the checkpoints from DropBox.
   • Move the checkpoints into notebooks directory.
   • Unzip the checkpoints.
   • Open the notebook T5_base.ipynb and run cells in all sections skipping Preprocessing, Finetuning.

Contributors

• Hardik A. Jain
• Chirayu Patel
• Riyasatali Umatiya
• Sajib Mistry
• Aneesh Krishna
• Amin Beheshti

Citation

If you use this dataset or model in your research, please cite our paper:

@inproceedings{Jain2025MedSiML,
  title={MedSiML: A Multilingual Approach for Simplifying Medical Texts},
  author={Hardik A. Jain, Chirayu Patel, Riyasatali Umatiya, Sajib Mistry, Aneesh Krishna, Amin Beheshti},
  booktitle={Neural Information Processing},
  year={2025},
  publisher={Springer Nature Singapore},
  address={Singapore},
  pages={165-179},
  isbn={978-981-96-6606-5},
  doi={https://doi.org/10.1007/978-981-96-6606-5_12}
}

License

This project is licensed under the MIT License. See the LICENSE file for details.