Using large language models to create narrative events.

Area of Science:

• Artificial Intelligence
• Semantic Web Technologies
• Scientific Communication

Background:

• Narratives are vital for conveying complex scientific information and knowledge.
• Scientific communities utilize narratives to explain phenomena and share insights.
• Integrating AI into scientific workflows can enhance data interpretation and knowledge dissemination.

Purpose of the Study:

• To explore the integration of large language models (LLMs) with Semantic Web technologies for transforming raw scientific data into narratives.
• To investigate the capability of LLMs in automatically generating reliable narrative events from scientific texts.
• To evaluate the performance of various LLMs in maintaining the integrity and accuracy of original scientific narratives.

Main Methods:

• Conceptual definition of narrative events.
• Evaluation of smaller LLMs on a corpus of 5 scientific narratives.
• Performance assessment of LLMs on a larger dataset of 124 narratives.
• Application of prompt engineering techniques to optimize LLM performance.

Main Results:

• LLaMA 2 demonstrated superior performance in generating narrative events that accurately reflect the input scientific texts.
• The study confirmed the feasibility of using LLMs to automatically create narrative events while preserving text reliability.
• Prompt engineering further improved the quality and fidelity of LLM-generated scientific narratives.

Conclusions:

• LLMs, particularly LLaMA 2, can be effectively integrated into workflows to generate reliable scientific narratives.
• This approach enhances the process of knowledge discovery and communication within scientific communities.
• The methodology ensures the integrity of original narratives, crucial for scientific accuracy and expert review.