Paragraph-level Simplification of Medical Texts

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Computational Linguistics

Background:

• Biomedical information is often inaccessible to the public due to technical jargon.
• Manual text simplification is not scalable for the vast and growing body of medical literature.
• Lack of large-scale parallel corpora for medical text simplification hinders automated approaches.

Purpose of the Study:

• To develop automated methods for simplifying medical texts.
• To create a new parallel corpus of technical and lay medical summaries.
• To propose and evaluate a novel metric for assessing text simplification quality.

Main Methods:

• Introduction of a new parallel corpus of technical and lay summaries in English.
• Proposal of a new metric using masked language models pretrained on scientific texts.
• Evaluation of baseline encoder-decoder Transformer models with a novel augmentation penalizing jargon production.

Main Results:

• The proposed metric effectively differentiates between technical and lay summaries, outperforming existing heuristics.
• Automated simplification models show improved readability.
• The novel augmentation for penalizing jargon production yields significant improvements in simplification quality.

Conclusions:

• Automated simplification of medical texts is feasible and necessary for public accessibility.
• The developed corpus and metric provide valuable resources for advancing research in this area.
• The proposed model augmentation offers a promising direction for creating more readable and understandable medical content.