MaTableGPT: GPT-Based Table Data Extractor from Materials Science Literature.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Extracting data from diverse tables in scientific literature is crucial for database creation.
• Existing rule-based methods are ineffective due to the varied formats of tables in materials science papers.

Purpose of the Study:

• To develop MaTableGPT, a GPT-based system for extracting table data from materials science literature.
• To improve GPT comprehension and filter hallucinated information for accurate data extraction.

Main Methods:

• Utilized GPT-based architecture with specialized strategies for table data representation and splitting.
• Implemented a follow-up question mechanism to filter hallucinated information.
• Evaluated zero-shot, few-shot, and fine-tuning learning methods for cost-effectiveness and accuracy.

Main Results:

• MaTableGPT achieved high extraction accuracy (up to 96.8% F1 score) on water splitting catalysis literature.
• Few-shot learning demonstrated the optimal balance between high accuracy (>95% F1 score) and low cost (GPT usage: $5.97, labeling: 10 examples).
• Statistical analysis of the generated database provided insights into catalyst properties like overpotential and elemental utilization.

Conclusions:

• MaTableGPT offers an effective solution for extracting structured data from complex materials science tables.
• Few-shot learning presents a cost-effective and accurate approach for data extraction using large language models.
• The generated database facilitates deeper understanding of materials for water splitting catalysis.