Structured information extraction from scientific text with large language models

Area of Science:

• Materials Chemistry
• Computational Science
• Natural Language Processing

Background:

• Extracting structured knowledge from scientific literature is a significant challenge for machine learning.
• Existing methods may lack the flexibility to handle diverse scientific data types.

Purpose of the Study:

• To present a simple, accessible approach for joint named entity recognition and relation extraction.
• To demonstrate the effectiveness of fine-tuning pretrained large language models (LLMs) for scientific knowledge extraction.
• To create large, structured databases from scientific research papers.

Main Methods:

• Fine-tuning pretrained LLMs (GPT-3, Llama-2) for named entity recognition and relation extraction.
• Applying the method to three materials chemistry tasks: dopant-host linking, metal-organic framework cataloging, and composition/phase/morphology/application extraction.
• Processing text from single sentences to entire paragraphs.

Main Results:

• Successfully extracted structured records of complex scientific knowledge.
• Demonstrated flexibility in output formats, including simple English sentences and JSON objects.
• Showcased the potential for creating large, specialized scientific knowledge databases.

Conclusions:

• The proposed approach offers a highly flexible and accessible route to automated scientific knowledge extraction.
• Fine-tuning LLMs is a viable strategy for building structured databases from unstructured scientific text.
• This method can significantly accelerate the curation of scientific data for research and development.