CorpFam says 93.2% of the hardest corporate-family links disappear before matching

A new public benchmark argues that corporate-family resolution is being measured at the wrong stage. In CorpFam, the authors report that 93.2% of parent–subsidiary links in the hardest “invisible” name stratum never enter the candidate set, meaning no downstream matcher can recover them.

A benchmark built around visibility

CorpFam contains 54,864 candidate pairs spanning 10,307 corporate families. Its labels come from 6,638,350 US federal award records in which suppliers self-report their ultimate parent, and the paper assigns pairs to three strata: identical normalized names, different names sharing a distinctive token, and “invisible” names sharing none. The authors split data by corporate family rather than by pair to reduce family memorization across train and test.

That stratification changes the apparent difficulty. On the test split, the strongest overall matcher—a character 3-gram TF–IDF system—reports 76.8 F1 overall, with 89.6% precision and 67.3% recall. But its recall falls from 100.0% on identical-name positives to 49.9% on visible links and 4.2% on invisible links. Across all five evaluated matchers, none exceeds 4.7% invisible-stratum recall.

The paper deliberately reports recall by stratum rather than F1 because the strata have sharply different positive rates, from 10.2% to 97.3%. Its warning is useful beyond this dataset: a single aggregate can mostly reward the easy, name-identical majority while hiding failure on the relationships that actually require corporate knowledge.

The ceiling appears before classification

CorpFam also evaluates seven blocking schemes over 114,230 entities, a naive space of 6.52 billion pairs. The strongest single blocker, MiniLM semantic nearest neighbors, achieves 72.5% pair completeness overall but only 2.89% on invisible links. Attribute blocking that ignores names reaches 2.39%; character q-grams reach 2.79%. Combining every scheme raises invisible-link recall to just 6.8%, while expanding the candidate set from 1.53 million pairs under token blocking to 13.07 million.

The authors address the obvious concern that invisible-name labels could be bad registry data. They report that SEC Exhibit 21 subsidiary schedules independently corroborate 64.2% of invisible links, versus 0.16% with permuted parents and 0.41% when searching the true parent’s wrong filing exhibit.

The benchmark’s conclusion is narrower than “entity matching is broken.” It says corporate-family resolution often requires retrieval of evidence outside the two compared records. That makes candidate generation—not a more elaborate pairwise ranker—the first system component that needs to change. For benchmark designers, the broader lesson is equally sharp: report performance on the cases where the needed relationship is not already visible in the input strings.