EVENT DOSSIER
The research team released the CorpFam benchmark dataset on September 7, 2026. This dataset incorporates 6,638,350 records from U.S. federal funding records, including 54,864 candidate pairs and 10,307 corporate families. The study evaluated the performance of the matcher by stratifying based on name visibility and found that the strongest matcher achieved a recall rate of 100.0% for identical pairs, while only 4.2% for invisible pairs. Since 93.2% of real links were never included in the candidate set due to filtering, the matching stage alone was insufficient to solve the problem of parsing corporate families. The study indicates that this task is essentially a retrieval problem rather than a simple string matching issue, and the key intervention point lies in the candidate generation stage rather than the ranking stage. The team has made the benchmark dataset, judgment logs, and source code available for subsequent research.
研究团队发布 CorpFam 基准，包含来自 6,638,350 份美国联邦拨款记录的 54,864 对候选配对及 10,307 个企业家族。该基准按名称可见性分层评估匹配器性能：最强匹配器在完全相同配对中召回率为 100.0%，在不可见配对中为 4.2%；而 93.2% 的真实链接因被过滤从未进入候选集，导致仅靠匹配阶段无法解决。研究指出企业家族解析本质是检索问题而非匹配问题，干预点在于候选生成而非排名，并公开了基准、裁决日志及代码。