SciGraphSelect is a reference‑grounded framework for auditable scientific data selection. It arranges domain‑specific knowledge, reusable reasoning operators, and reference‑attested relations into a typed knowledge‑reasoning graph. Candidate documents are scored by measuring concept coverage, co‑support of relations and paths, and directional alignment. To avoid over‑interpreting document embeddings, the rich relational structure on the reference side is preserved, while evidence on the candidate side is treated conservatively without assuming entailment. By exposing the reference‑grounded evidence behind each selection, SciGraphSelect adds traceability to the upstream construction of scientific training corpora and complements downstream verification.
Review: The approach balances semantic richness with transparent scoring, offering a trustworthy audit trail for large‑scale scientific data curation.