4 Results

Table [1] summarizes performance across the four benchmark tasks. I compare against fastLink (Enamorado et al., 2019), a probabilistic record linkage method implementing the Fellegi-Sunter framework, and a re-implementation of fuzzylink (Ornstein, 2025) using GPT-4o-mini for LLM labeling. All methods are evaluated using identical methodology: for each task, I randomly select 40% of unique queries as the test set, then each method selects its top-1 prediction from the full corpus. Accuracy is the fraction of queries matched correctly. Without any labeled data, EnsembleLink outperforms both baselines on all four tasks.

Table 1: Accuracy across benchmark tasks (40% test split, top-1 matching). Exact match, fastLink, and EnsembleLink require no labels; fuzzylink uses LLM-assisted active learning. Blocking column indicates whether hierarchical blocking was applied (with blocking variable in parentheses).

| Task | Method | Blocking | Accuracy |
|------|--------|----------|----------|
| City Matching | Exact match | — | 0.000 |
|      | fastLink | — | 0.713 |
|      | fuzzylink | — | 0.717 |
|      | EnsembleLink | None | 0.901 |
|      | EnsembleLink | State (Fuzzy) | 0.900 |
| Candidate-Voter | Exact match | — | 0.000 |
|      | fastLink | — | 0.230 |
|      | fuzzylink | — | 0.965 |
|      | EnsembleLink | — | 0.990 |
|      | EnsembleLink | Last name (Fuzzy) | 0.960 |
| Organization | Exact match | — | 0.000 |
|      | fastLink | — | 0.282 |
|      | fuzzylink | — | 0.922 |
|      | EnsembleLink | None | 0.961 |
| Multilingual Parties | Exact match | — | 0.102 |
|      | fastLink | — | 0.096 |
|      | fuzzylink | — | 0.391 |
|      | EnsembleLink | — | 0.843 |
|      | EnsembleLink | Country (Fuzzy) | 0.926 |
|      | EnsembleLink + LLM | — | 0.894 |
