Retrieval augmentation of large language models for lay language generation.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Scientific Communication

Background:

• Biomedical literature's complexity hinders public understanding.
• Existing automated methods focus on summarization/simplification, neglecting background explanation.
• Effective background explanation is crucial for lay comprehension but underexplored.

Purpose of the Study:

• Introduce Retrieval-Augmented Lay Language (RALL) generation for enhanced biomedical text simplification.
• Develop and evaluate methods for adding background explanations to lay summaries.
• Present CELLS, a large parallel corpus for lay language generation.

Main Methods:

• Developed RALL generation integrating information retrieval with text generation models.
• Utilized UMLS and Wikipedia for term definitions and explanation embeddings.
• Evaluated RALL models using state-of-the-art text generation models, including Llama 2 and GPT-4.
• Created and utilized the CELLS corpus (63k pairs) spanning 12 journals.

Main Results:

• Embedding-based RALL models significantly improved summary quality and simplicity.
• Wikipedia proved a valuable source for background explanations.
• Large Language Models (LLMs) show potential but require further refinement for optimal summary quality.
• RALL demonstrated effectiveness in enhancing background explanation for lay language generation.

Conclusions:

• RALL generation is a promising approach for making biomedical literature accessible.
• Integrating external knowledge sources like Wikipedia enhances lay language summaries.
• Further research is needed to optimize LLM performance in background explanation for scientific content.
• This work establishes a foundation for broader scientific knowledge dissemination.