Better Retrieval, Limited Clustering Gains: A Controlled Study of Multilingual Company Entity Resolution

1 Introduction

• Missed links fragment evidence, while false merges combine unrelated entities and can propagate into downstream analytics.The task links alternative company names across scripts and sources.Source: §1 · p. 1
• Candidate recall alone cannot establish final entity quality because existing classifiers may not recognize new matches and transitive clustering can propagate an incorrect edge.The concern is especially relevant when a deployed retriever changes.Source: §1 · p. 1
• The study compares frozen and adapted multilingual E5 encoders [3] under matched positives, country eligibility, candidate budgets, and decision rules.It asks whether adaptation improves coverage, whether hard negatives outperform random negatives, and where new candidates fail to improve clusters.Source: §1 · p. 1
• The empirical contributions include an 11-arm comparison under two fixed matchers, pair-level tracing to complete partitions, and an assisted audit recording evidence dependence.A revision-stage temperature sweep extends the analysis; the work uses established components.Source: §1 · p. 2

2 Related work

• JRC-Names-Retrieval benchmarks multilingual name search, while DeepBlocker and UniBlocker [6] motivate deep and contrastive blocking approaches.The paper uses established supervised contrastive learning without claiming a new loss or state-of-the-art blocking performance.Source: §2 · p. 2
• Ditto fine-tunes language models for pair classification, whereas this study keeps classifiers fixed to isolate retriever replacement without evaluating joint retraining.RRF, CSLS, and a capped union provide ranking controls and a stronger frozen multi-view baseline.Source: §2 · p. 2
• Embedding studies on short reviews and financial reranking relate representations to task outcomes but do not study company partitions.Efficiency studies [13], address other workloads; equal candidate budgets here do not measure those optimizations or serving cost.Source: §2 · p. 2

3 Method and workflow

• The five-step workflow moves from supervision and exposure exclusions through encoder adaptation and development-only selection to equal-budget retrieval, fixed confirmation, clustering, and partition evaluation.The analysis traces gained pairs to their decisions.Source: §3 · p. 2, Figure 1 · p. 2

3.1 Records and evaluation targets

• Silver positives are pairs of records sharing an identity, and candidate recall is the fraction of those pairs retrieved; cluster metrics include transitive relations.Identifier equality supplies silver rather than independently adjudicated gold labels.Source: §3.1 · pp. 2–3

3.2 Contrastive encoder adaptation

• Contrastive adaptation maps query-prefixed names to normalized embeddings and trains same-identity names to be closer relative to competitors.Negatives share source and country but differ in identity/source group; random and hard sampling use permitted records and original-encoder top-20 neighbors, respectively.Source: §3.2 · p. 3

3.3 Candidate selection at equal volume

• Retrieval arms rank candidate edges using lexical or dense similarities and ranking combinations, with each country contributing a capped edge budget.Within-selector comparisons isolate adaptation, while between-selector comparisons also change retrieval policy.Source: §3.3 · p. 3

3.4 Pair classification and guarded Union-Find

• Fixed classifiers score candidate pairs, then greedy constrained Union-Find merges components only when country compatibility and the cross-pair score guard are satisfied.The guard limits erroneous bridge propagation, but the greedy procedure can depend on edge order, is not globally optimal, and cross-pair checks can be expensive for large components.Source: §3.4 · pp. 3–4

4.1 Data provenance and coverage

• The evaluation snapshot contains 7,800 records, 2,880 identities, 3,633 names, and 882 silver-positive pairs drawn from sampled GLEIF pages.Filtering reduced leakage but did not ensure semantic family separation; all countries and GLEIF were historically seen, limiting inference to held-out-record comparison rather than worldwide or unseen-country generalization.Source: §4.1 · p. 4
• The sampled names include primary, alternative-language, and ASCII variants, while country imbalance, bounded sampling, and source-generated transliterations limit coverage.The audit examines label reliability and its evidence dependence.Source: §4.1 · p. 4

4.2 Training parameters and temperature control

• Both adaptation arms use the same pinned multilingual-e5-small initialization, seed, optimizer schedule, and positive-pair count; the revision adds a post-hoc temperature sweep.Development RRF recall selects epoch 1 or 2, while the original 0.05 checkpoints remain the reference and the evaluation pool is reused without selecting a new temperature.Source: §4.2 · pp. 4–5

4.3 Frozen confirmation and analysis controls

• Confirmation uses fixed classifier settings and historical thresholds, with θ = 0.4 targeting 95% development precision and γ = 0.05 retained as the guard.Weak development RRF results motivated pre-acquisition UNION controls; checkpoints were not test-selected.Source: §4.3 · p. 5
• The paired bootstrap conditions on models, data, and silver labels, excluding seed and label uncertainty and multiplicity adjustment.Sixty checks and independent recounts cover 22 original partitions; all revision partitions and guards are also checked.Source: §4.3 · p. 5

5.1 Retrieval and final clusters

• HARD-UNION raises direct recall from 474/882 to 679/882, while GLOBAL-HARD cluster recall rises from 32.54% to 33.33% and precision falls from 95.99% to 95.45%.Table 1 reports equal budgets of 72,660 edges per arm and silver labels.Source: Table 1 · p. 5, §5.1 · p. 5
• The original RRF contrast improves retrieval by 18.48 points and cluster recall by 7.03, but its frozen baseline is weaker than frozen dense and lexical retrieval.Source: §5.1 · p. 6

5.2 Where retrieval gains stop

• Of 208 gained direct positives, GLOBAL-HARD scores 202 below 0.4; six pass and co-cluster, while the seventh additional correct relation is transitive.LOCAL-HARD admits three gained direct positives.Source: §5.2 · p. 6
• Positive-edge-only closure connects 718/882 positives (81.41%); cross-script coverage rises from 385/691 to 553/691, and same-script coverage from 89/191 to 126/191.These are heuristic script strata, and the uneven country mix limits language-specific interpretation.Source: §5.2 · p. 6

5.3 Temperature sensitivity

• UNION retrieval recall ranges from 73.47% to 76.98% across three temperatures, but the post-hoc sweep establishes neither an optimal temperature nor multi-seed robustness.Table 2 uses fixed confirmation, 72,660 candidates per arm, and 882 silver-positive pairs.Source: Table 2 · p. 6, §5.3 · p. 6

• The 137-pair sample includes every false co-cluster pair and all three lost positives, plus up to 30 sampled pairs per remaining stratum.The 30-pair cap was a workload limit, not a power calculation.Source: §6 · p. 6
• The audit recorded 92 SAME, 44 DIFFERENT, and one UNSURE judgment; all 136 resolved judgments agree with silver labels.Table 3 reports the assisted single-reviewer judgments.Source: §6 · p. 6, Table 3 · p. 7
• All 30 sampled gained positives remain rejected, and reviewers judge all 14 false co-cluster pairs, including two new errors, DIFFERENT.Source: §6 · p. 6
• For 107/137 entries, including 24/30 gained positives, the cited evidence consists only of GLEIF/LEI pages; other URLs need not provide independent evidence either.Source: §6 · p. 6
• Single-reviewer, source-dependent judgments cannot establish gold accuracy or agreement; unequal sampling prevents population-precision estimates, and 30 residual negatives cannot tightly bound hidden matches.Source: §6 · p. 7

• Fixed-confirmation results do not establish behavior under joint retraining.Source: §7 · p. 7
• One seed, seen sources and countries, heuristic filtering, source transliterations, and silver labels restrict generalization; production search, missing-country cases, and incremental updates remain untested.The post-hoc sweep also does not resolve optimality or seed variability.Source: §7 · p. 7
• Future work proposes training matchers on out-of-fold deployment-retriever candidates, separating feature, training-distribution, and threshold effects, and assessing thresholds on a fresh test sample.Independent dual annotation, broader country coverage, multiple seeds, and guard-cost measurements are also proposed.Source: §7 · p. 7

• Protocols, acquisition responses, checkpoints, schedules, features, scores, partitions, and annotation records are retained locally, but no public archive is yet available.Reproduction requires the saved snapshot because live pagination changes.Source: §8 · p. 7
• AI tools were used only to polish implementation code; the authors claim the research ideas, methodological choices, analyses, and interpretations as their own.Source: §8 · p. 7
• One reviewer verified the assisted annotations, and the authors accept responsibility for manuscript accuracy.Source: §8 · p. 7

9 Conclusion

• Contrastive adaptation improves retrieval under equal budgets, but fixed confirmation adds few correct relations and introduces new errors.Source: §9 · p. 8
• Pair tracing locates most lost retrieval gains at the classifier threshold.Source: §9 · p. 8
• The study does not establish whether jointly retrained systems share this limitation; future evaluation should test confirmation changes and independent identity labels.Source: §9 · p. 8