Corporate-Family Resolution Is Not a String-Matching Problem: A Public Benchmark Stratified by Name Visibility
cs.DB · cs.IR · cs.LG · cs.LO
Quick verdict
Claims corporate-family linkage fails at blocking, not matching; benchmark is rigorous and unusually honest, but baselines are weak and ground truth noisy.
Summary
The paper argues that corporate-family resolution—deciding whether two supplier records share an ultimate parent—is a retrieval problem misfiled as entity matching. The author builds CorpFam from 6.6M FY2025 US federal award records whose recipients self-report an ultimate parent, yielding 54,864 pairs over 10,307 families after removing 86.3% self-parenting registrations and hand-adjudicating 28 anomalous families. Pairs are stratified by name visibility (identical/visible/invisible after normalisation), and per-stratum recall is reported instead of F1 because stratum positive rates span 10.2-97.3%. Results: the best of 5 matchers reaches 100.0% recall on identical, 49.9% on visible, and 4.2% on invisible pairs (MiniLM 4.7%); a union of 7 blocking schemes recovers only 6.8% of invisible links, so 93.2% never reach a matcher. Against SEC Exhibit 21 subsidiary schedules, 64.2% of invisible links are corroborated versus 0.16% (permuted parent) and 0.41% (wrong exhibit), supporting label validity. Registry self-disagreement (18.5% substantive) bounds attainable accuracy, and attributes cannot help because 94.1% of parents never transact. Limitations include weak baselines (no fine-tuned or LLM matchers), a single jurisdiction and fiscal year, one-hop parent relations, and non-random SEC corroboration coverage.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Corporate-family resolution is a retrieval problem misfiled as a matching problem, and the intervention point is candidate generation rather than ranking.
- On the invisible stratum the best of 5 evaluated matchers reaches 4.2% recall (MiniLM 4.7%), with every method at or below the majority-class floor on F1.
- The union of 7 blocking schemes recovers only 6.8% of invisible links, so 93.2% never enter any candidate set and are unreachable by matcher improvements.
- 64.2% of invisible links are corroborated by SEC Exhibit 21 subsidiary schedules versus 0.16% and 0.41% under two mechanistically unrelated null conditions, so the links are genuine relationships rather than registry noise.
- Attributes cannot rescue the stratum because 94.1% of parents never appear as award recipients, leaving the attribute vector empty as a matter of what a holding company is.
Strengths
Base-rate-honest metric design
Section 3.4 argues per-stratum recall over F1 because test-split positive rates are 97.3/10.2/16.9, and every table carries a majority-class floor (Table 5's 'answer yes to everything' row). The author candidly reports two of his own intermediate results (invisible-stratum F1 of 88 and 59.9) as base-rate artifacts he caught before publishing.
Dual-null corroboration architecture
Section 4.2/Table 3 validates labels externally: 64.2% invisible-stratum corroboration against SEC Exhibit 21 [60.8, 67.5] versus 0.16% under permuted parents (20 permutations, 15,795 links) and 0.41% under right-parent-wrong-exhibit (1/244). The floors land 0.25 points apart, Fisher p=0.34, and the paper explains why a boilerplate-keyed or company-token-keyed artifact would fail exactly one control each — neither happens.
Ground truth quantified rather than asserted
Reproducibility engineering with an a-fortiori check
Weaknesses
Thin matcher set behind a universal claim
Residual circularity requires careful reading
Label noise quantified but never propagated
Non-random corroboration coverage
Underpowered sub-analyses carrying design prescriptions
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
CorpFam (54,864 pairs; 10,307 families; 13,716 positives vs 41,148 negatives) is offered as a public benchmark for corporate-family resolution, plus a thesis: the task is a retrieval problem misfiled as entity matching, so the binding constraint is candidate generation, not ranking. Three headline claims are supported by the text: (1) the best of 5 matchers reaches 4.2% recall on invisible pairs and MiniLM 4.7% (Table 5); (2) the union of 7 blocking schemes recovers 6.8% of invisible links over a 114,230-entity roster, leaving 93.2% unreachable by any matcher (Table 4); (3) 64.2% of invisible links are corroborated by SEC Exhibit 21 schedules against null floors of 0.16% (permuted parent) and 0.41% (wrong exhibit) (Table 3). The universal phrasing 'no method exceeds 4.7%' is scoped to these five matchers; that caveat surfaces only in Section 10.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the CorpFam benchmark and where does its data come from?
CorpFam is a public benchmark of 54,864 labelled candidate pairs covering 10,307 US corporate families, built from 6.6 million FY2025 US federal contract award records. Labels come from the ultimate parent each supplier self-reports during federal registration (SAM.gov/USAspending UEI data), and every pair is stratified by whether the child and parent names are identical, share a distinctive token, or share none after normalisation.
Why do embedding models fail to link companies like BLUE AEROSPACE to HEICO?
Because the ownership fact is recorded in legal filings, not in the two name strings; nothing in the text of either name connects them. The paper shows a MiniLM sentence encoder reaches only about 3-5% recall on these name-invisible pairs, and unioning it with six other blocking schemes still leaves 93.2% of such links outside the candidate set.
What does the paper recommend practitioners do instead of string matching?
Treat corporate-family resolution as a retrieval problem: bring ownership evidence (filings, ownership databases) into candidate generation before any matcher runs. Once a link is absent from the candidate set, no ranking or matching improvement can recover it, which is why 93.2% of invisible links are currently unreachable.
How reliable are the benchmark labels?
Moderately. Two independent registry channels agree on the exact parent UEI only 64.9% of the time (73.8% on normalised names, with 18.5% naming substantively different parents), but 64.2% of the hardest name-invisible links are independently confirmed in SEC Exhibit 21 subsidiary schedules versus roughly 0.2-0.4% under two null conditions, so the links are real corporate relationships.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.