Entity Source Provenance

loading…

Most of the SciX entity graph is auto-extracted by open-weight NER (GLiNER, over titles and abstracts); a smaller curated slice comes from trusted authorities — SSODNet for asteroids, ASCL for software, VizieR for catalogues. This page maps how many entities each source contributes and, crucially, how trustworthy each is: the precision band an agent sees on the MCP entity tool, from the dbl.3 quality profile. V8 says what we have; V12 says how much of it is ground-truth vs best-effort.

Entities by source & type

Treemap of entity counts; each tile is a (source, entity_type) pair, area proportional to the number of distinct entities. GLiNER dominates; curated authorities are the small high-trust tiles.

Precision profile by source

Share of each source's mentions falling in each precision band, from scix.extract.ner_quality_profile. GLiNER bands are estimated from a representative sample of links (precision depends on entity type, paper era, and the INDUS classifier verdict); curated sources carry a single constant lexical precision.

high (≥0.85) medium (≥0.70) low (≥0.50) noisy (<0.50)

Ambiguity classes

Entities flagged by the ambiguity classifier (scripts/classify_entity_ambiguity.py). Only the classified slice is charted; the unclassified majority is noted below.