EnsembleLink: Accurate Record Linkage Without Training Data

Noah Dasanaike

cs.CLJan 29, 2026

EnsembleLink: Accurate Record Linkage Without Training Data

Authors: Noah Dasanaike

Organizations: PhD Candidate, Department of Government, Harvard University

Abstract

Record linkage, the process of matching records that refer to the same entity across datasets, is essential to empirical social science but remains methodologically underdeveloped. Researchers treat it as a preprocessing step, applying ad hoc rules without quantifying the uncertainty that linkage errors introduce into downstream analyses. Existing methods either achieve low accuracy or require substantial labeled training data. I present EnsembleLink, a method that achieves high accuracy without any training labels. EnsembleLink leverages pre-trained language models that have learned semantic relationships (e.g., that "South Ozone Park" is a neighborhood in "New York City" or that "Lutte ouvriere" refers to the Trotskyist "Workers' Struggle" party) from large text corpora. On benchmarks spanning city names, person names, organizations, multilingual political parties, and bibliographic records, EnsembleLink matches or exceeds methods requiring extensive labeling. The method runs locally on open-source models, requiring no external API calls, and completes typical linkage tasks in minutes.

Figures & tables

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

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	Query	Matched Reference
City	South Ozone Park, NY	Queens, NY
	East Elmhurst, NY	Queens, NY
	OKC, OK	Oklahoma City, OK
	Newbury Park, CA	Thousand Oaks, CA
Candidate-Voter	Tim Flynn	Timothy Bryant Flynn
	Tony Daysog	Anthony Honda Daysog

Table 2: Example matched pairs from each task, drawn from correct predictions. Examples are chosen to maximize string distance between query and reference, demonstrating the method’s ability to handle semantic similarity beyond lexical overlap.

Reranker	City	Candidate	Organization	Party
Jina v2 Multilingual	0.900	0.990	0.961	0.843
Jina v3	0.881	0.985	0.961	0.837
BGE Reranker v2 M3	0.903	0.990	0.951	0.777
MiniLM (English-only)	0.701	0.990	0.971	0.306

Table 3: Ablation: reranker comparison across tasks (Qwen3-0.6B embeddings,

k=30

, top-1 accuracy).

Task	Zero-shot	25	50	100	200	300
City	0.989	0.973	0.988	0.988	0.987	0.988
Candidate-Voter	0.977	0.964	0.968	0.977	0.976	0.977
Organization	0.902	0.878	0.898	0.902	0.902	0.892
Party	0.949	0.949	0.931	0.949	0.949	0.945

Table 4: Fine-tuning experiments: F1 scores at different label counts, evaluated on held-out pair samples. Fine-tuning provides no consistent improvement over the zero-shot baseline.

Corpus	Embed	Index	Retrieve	Rerank	Total	Acc
5,000	22.0s	2.0s	49.8s	52.3s	132.8s	0.984
10,000	35.5s	3.2s	111.5s	65.2s	221.8s	0.989
20,000	71.2s	7.5s	227.8s	59.2s	372.4s	0.985
40,000	142.7s	13.3s	437.6s	57.9s	658.5s	0.982
64,263	216.8s	19.4s	413.1s	35.3s	729.5s	0.979

Table 5: Scalability on DBLP-Scholar (963 test queries, 40% split)

Method	Training Data	F1	Accuracy
Supervised (thousands of labels)
Ditto ( Li et al., 2020 )	Full training set	94.3
DeepMatcher ( Mudgal et al., 2018 )	Full training set	94.7
RoBERTa fine-tuned	Full training set	93.9
Zero-shot LLMs
GPT-4 ( Peeters et al., 2024 )	None	89.8

Table 6: Comparison with published results on DBLP-Scholar. Published methods use pair-level F1 on candidate pairs. For EnsembleLink, I report both top-1 accuracy and pair-level F1 at three score thresholds

\tau

Model	Device	Pairs/sec	ms/pair	Speedup
Jina Reranker v2 Multi (278M)	GPU	748	1.3	30 $\times$
Jina Reranker v2 Multi (278M)	CPU	25	40	1 $\times$
MiniLM-L-6 (22M)	GPU	581	1.7	1.2 $\times$
MiniLM-L-6 (22M)	CPU	478	2.1	1 $\times$

Table 7: Inference speed comparison: GPU (NVIDIA RTX 4090) vs CPU

Entity Resolution

Entity resolution links records that refer to the same real-world entity across datasets despite differences in names, attributes, formats, and languages.

See all 38 papers

Pre-Trained Language Models as Zero-Shot Tools for Social Science Research
Oct 13, 2026Noah DasanaikeEntity Resolution Zero-Shot Learning
PhD Candidate, Department of Government, Harvard University.
Entity Resolution in Practice: Lessons from a Self-Serve Pipeline
Jul 28, 2026Kaushik Pavani, Ganga Aluri, Pravin Jadhav +2Entity Resolution
Walmart
Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking
Sep 9, 2026Parinthapat Pengpun, Simran Khanuja, Graham NeubigMultimodal Grounding VLM Reasoning
Carnegie Mellon University

Oct 13, 2026J/K move · Enter open · S save

Pre-Trained Language Models as Zero-Shot Tools for Social Science Research

Noah Dasanaike

PhD Candidate, Department of Government, Harvard University.

Entity Resolution Zero-Shot Learning

Open paper AskPDF arXiv

Record linkage and demographic inference from names are routine demands in social science. Most approaches to these problems require labeled training data that are difficult or costly to acquire. I show that pre-trained language models perform both tasks accurately without requiring task-specific training. For record linkage, I introduce EnsembleLink, an algorithm that uses pre-trained embeddings alongside string-based scorers to retrieve a short list of candidates and then identify the likeliest matches. EnsembleLink matches or exceeds the performance of most prior approaches while generalizing across scripts and formats that those methods are unable to handle. For ethnicity inference from names, I show that large language models (LLMs) achieve up to 83.8% accuracy on U.S. voter files, compared to 78.0% for name-supplemented BISG. Open-weight models offer reproducible implementations that approach these models in accuracy. Moreover, LLMs are able to infer group membership for religious, caste-based, and other ethnic group categories outside of the United States, and reduce bias in downstream estimates of geographic racial gaps. Where pre-training encodes the world knowledge required by each task, task-specific labels add less information.

Abstract

Entity Resolution

Explore similar work

Pre-Trained Language Models as Zero-Shot Tools for Social Science Research

Entity Resolution in Practice: Lessons from a Self-Serve Pipeline

Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking