SciDaSynth: Interactive Structured Data Extraction from Sci Literature with LLM (onlinelibrary.wiley.com)

SciDaSynth is a new system that uses large language models to extract and synthesize structured, machine-readable data from scientific literature through an interactive, human-in-the-loop workflow. Rather than treating papers as opaque blobs of text, SciDaSynth frames extraction as a schema-guided generation task: LLMs are prompted (and optionally fine-tuned) to output relational records, nested fields, and table-like entries in JSON/CSV form, while interactive validation tools let users correct or confirm outputs. A key innovation is generating synthetic annotation examples and constrained prompts to reduce hallucination and improve robustness across diverse manuscript styles and layouts. This work matters because structured data locked in text and figures is a major bottleneck for reproducibility, meta-analysis, and building domain knowledge graphs. By combining synthetic training data, schema constraints, and human verification, SciDaSynth aims to substantially lower manual curation costs and produce cleaner datasets for downstream ML tasks (model training, benchmarking, systematic reviews). Technical implications include improved extraction of experimental parameters, measurements, and methodologies, better integration with data pipelines (JSON/CSV export), and a pragmatic approach to mitigate LLM errors via iterative feedback—though domain shifts and edge-case formatting still require oversight. Overall, SciDaSynth showcases a practical path for scaling high-quality scientific data extraction using LLMs.