Gene–disease extraction: our method vs PubTator3

Same 12 benchmark papers. Our method reads the parsed full text with the decision card (variant v1_card_lean_medium), using one initial request and at most one repair. PubTator3 results are precomputed annotations fetched by PMID and aligned locally; its relations are primarily abstract-derived. Neither system is used as the answer key. Output counts and disagreements do not establish relative accuracy or recall.

Head to head

Each row is scaled to its own larger value.

Relation overlap

Relations matched through aligned entities. Every segment is a disjoint count; the two bars use different totals.

Output fields and cost

Selected examples

These examples were selected by hand to show different disagreements. They are not a random sample.

Overlap by level

Each bar is one level, scaled to 100% of the items either system found; labels show counts. PubTator also tags symptoms and general terms as Disease;

Gene–disease relations per paper

Bars share one scale. Our method also marks pairs as uncertain or explicit_negative; PubTator has no such states (counts in the table below).

Why the relation sets differ

Every relation disagreement

Per-paper table

Entity differences

Names as PubTator normalised them, with the text it tagged in brackets. Several PubTator-only "diseases" are abbreviation mis-normalisations (e.g. ECM, EMC, PC) or general words; others are phenotypes our method did not list.