Sentence-Aligned Simplification of Biomedical Abstracts.

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Health Communication

Background:

• Biomedical abstracts contain technical jargon, hindering public health literacy and informed decision-making.
• Existing simplification methods struggle with cohesive abstract-level adaptation and sentence-by-sentence accuracy.

Purpose of the Study:

• To adapt foundation language models for the Plain Language Adaptation of Biomedical Abstracts task.
• To develop a sentence-wise autoregressive approach for cohesive abstract simplification.
• To improve the accessibility and trustworthiness of biomedical information for healthcare consumers.

Main Methods:

• Utilized a sentence-wise autoregressive approach with foundation language models.
• Conducted experiments in zero-shot and fine-tuned settings using proprietary and open-source models.
• Introduced a stochastic regularization technique to prevent source-copying during inference.

Main Results:

• Achieved a 32-point increase in SARI and a 6-point increase in BERTscore over state-of-the-art.
• Outperformed recent open-domain and biomedical sentence simplification models on this task.
• Manual evaluation demonstrated human-level factual accuracy and simplicity.

Conclusions:

• Adapted foundation language models can effectively simplify biomedical abstracts, enhancing health information accessibility.
• The proposed methods improve information trustworthiness by retaining clear provenance for simplified statements.
• This advancement has the potential to significantly improve public health literacy and empower informed health choices.