SciDaSynth: Interactive Structured Data Extraction From Scientific Literature With Large Language Model

Area of Science:

• Scientific data extraction
• Information retrieval
• Knowledge discovery

Background:

• The rapid growth of scientific literature necessitates efficient data extraction.
• Existing tools struggle with multimodal and inconsistent data formats.
• Structured data is crucial for evidence-based decision-making.

Purpose of the Study:

• To introduce SciDaSynth, an interactive system for automated structured data table generation.
• To enable data integration from diverse sources like text, tables, and figures.
• To support efficient data validation and refinement for cross-document consistency.

Main Methods:

• Utilizing large language models (LLMs) for data extraction and structuring.
• Developing an interactive system for user-guided data table generation.
• Implementing visual summaries and semantic grouping for data validation.

Main Results:

• SciDaSynth effectively generates structured data tables from multimodal sources.
• The system demonstrated higher efficiency in producing high-quality structured data compared to baseline methods.
• Researchers confirmed the system's utility in resolving cross-document data inconsistencies.

Conclusions:

• SciDaSynth offers a novel approach to structured data extraction from scientific literature.
• The system enhances the efficiency and quality of data synthesis for researchers.
• Design implications for human-AI collaborative data extraction systems are discussed.