The review, published on 12 October 2026, makes the case that the choice between SQL and NoSQL architectures is not a mere implementation detail. It fundamentally shapes which entity resolution techniques are feasible, how they must be optimized, and how well they perform at scale. To capture this, the authors construct a three-dimensional classification framework that organizes the sprawling literature along three axes: the characteristics of the underlying data model, the system architecture on which resolution runs, and the target workload the technique is designed to serve. This database-centric lens, they contend, bridges a persistent gap between algorithmic innovation in research labs and practical deployment in real data infrastructures.

The intellectual roots of entity resolution stretch back decades. A landmark statistical theory of record linkage published by Ivan Fellegi and Alan Sunter in 1969 established the probabilistic foundations still used today, framing the matching decision as a formal hypothesis test between the possibilities that two records are a match or a non-match. Early systems such as the merge/purge work of Hernández and Stolfo in 1995 tackled the problem on relational databases, where rigid schemas, structured tables, and declarative query languages offered a relatively tidy environment. In that world, rule-based methods flourished: engineers wrote matching dependencies and similarity thresholds over well-defined attributes, and techniques such as sorted neighborhood indexing and canopy clustering kept the quadratic explosion of record comparisons under control through blocking, the art of cheaply grouping likely matches before expensive comparisons are made.

Relational databases, however, assume that data arrives clean, typed, and schema-conformant. The explosive growth of web-scale, heterogeneous, and semi-structured data shattered that assumption. NoSQL systems—document stores, key-value stores, column-family databases, and graph databases—emerged precisely to handle flexible schemas, horizontal scaling, and high-velocity ingestion. But that flexibility is a double-edged sword for entity resolution. When the same customer may appear as a nested JSON document in one store and a flat row in another, when attributes are optional, inconsistently named, or embedded in free text, the tidy rule-based machinery of the relational era strains. The review documents how techniques adapted: schema extraction methods, some using BERT embeddings to infer structure from schema-less document stores, attempt to recover the semantic scaffolding that relational systems get for free, while schema-agnostic approaches such as the MinoanER framework resolve web entities without any schema knowledge at all.

Scale is the second axis of the framework, and here the review traces a dramatic architectural migration. As datasets outgrew single machines, entity resolution moved to distributed paradigms. MapReduce-based systems such as Dedoop demonstrated how deduplication could be parallelized across Hadoop clusters, and later Spark-based approaches, including a probabilistic entity resolution pipeline published in this same journal, pushed throughput further. Parallel and distributed NoSQL platforms changed the calculus of blocking as well: probabilistic signatures, meta-blocking strategies that organize candidate pairs into a block graph, and progressive blocking techniques that surface the most promising comparisons first all reflect the need to make resolution tractable when billions of pairs are in play. The review also highlights real-time and on-demand variants, from inverted-index approaches for streaming resolution to systems that resolve entities lazily, only when a query actually needs them.

The third axis, workload, distinguishes batch deduplication from streaming, interactive, and progressive scenarios. Interactive deduplication using active learning, first explored in the early 2000s, lets a human oracle adjudicate the most ambiguous cases, and later work on online entity resolution formalized how to query that oracle efficiently. Incremental record linkage addresses the reality that databases change continuously, updating resolution results without recomputing from scratch. These workload distinctions matter enormously in practice: a nightly batch job consolidating a data warehouse faces very different latency and consistency constraints than a fraud-detection system that must flag a duplicate account the moment it is created.

Perhaps the most vivid thread in the review is the rise of deep learning. A 2018 design-space exploration for deep entity matching marked a turning point, and the field has since raced through sequence-to-sequence models for heterogeneous records, BERT-based matchers optimized for efficiency, transformer architectures for data integration, graph convolutional networks that exploit relational structure, and hierarchical graph attention networks. Pre-trained language models and their embeddings have become a lingua franca for comparing messy textual attributes, with experimental analyses cataloguing which embeddings work best for which matching problems. The frontier has now reached large language models: systems such as Booster use them to enhance resolution pipelines, in-context learning offers cost-effective matching without task-specific training, and ChatGPT has even been evaluated directly as an entity matcher. Zero-ER demonstrated matching with zero labeled examples, and ensemble methods squeeze further accuracy from transformer stacks. In a striking sign of the times, the review also covers hybrid quantum neural networks, which recent experiments suggest can dramatically reduce the need for hand-tuned free parameters in entity matching.

Graph databases receive particular attention as both a challenge and an opportunity. Because graph stores encode relationships as first-class citizens, they invite collective entity resolution, where decisions about one entity inform decisions about its neighbors. Techniques range from latent Dirichlet models and collective resolution in familial and multi-relational networks to graph differential dependencies and knowledge-graph completion pipelines that perform incremental multi-source resolution. The review shows how the data model itself becomes an algorithmic resource: in a property graph, the structure around a node is evidence about its identity, something a flat relational table cannot offer. Conversely, graph workloads introduce their own costs, since traversals and neighborhood computations can be expensive at scale, motivating fast on-demand resolution methods for property graphs.

The review does not shy away from the harder constraints now bearing down on the field. Privacy-aware deep entity resolution and federated resolution address the reality that sensitive records often cannot be pooled centrally, with encrypted-data deduplication pushing matching into the cryptographic realm. Vector databases, a rapidly maturing class of systems built for similarity search over embeddings, are emerging as a natural substrate for entity resolution at scale, and the review surveys their management systems and their use for identity deduplication. Knowledge bases and ontologies add another layer, with formal frameworks now combining entity resolution with query answering so that cleaning and consumption happen in a single coherent pipeline.

What elevates this review beyond a catalogue is its insistence on actionable guidance. By mapping each family of techniques—rule-based, probabilistic, machine learning, deep learning, and LLM-based—onto the data models, architectures, and workloads where it thrives, the authors give practitioners a decision framework rather than a leaderboard. A team running a strictly relational warehouse with stable schemas and moderate volumes may still be best served by declarative rules and sorted neighborhood indexing; an organization ingesting heterogeneous documents across a distributed NoSQL estate will likely need schema-agnostic blocking, embedding-based similarity, and Spark-scale parallelism; a graph-centric knowledge base opens the door to collective, structure-aware resolution. The authors, who received no external funding for the work and declare no competing interests, frame their contribution as closing the loop between algorithmic research and database-aware deployment. In an era when data is routinely called the new oil, their message is blunt: the refinery matters as much as the crude. Entity resolution is not one problem but a family of problems, and the database underneath is not passive plumbing—it is an active determinant of what can be matched, how fast, and how well.

Subject of Research: Entity resolution techniques across SQL and NoSQL database management systems

Article Title: Entity resolution across SQL and NoSQL database management systems: a comprehensive review

Article References: Mekki, N., Berrabah, D., Malki, A., & Boufarès, F. (2026). Entity resolution across SQL and NoSQL database management systems: a comprehensive review. Knowledge and Information Systems, 68(1), Article 281. https://doi.org/10.1007/s10115-026-02878-z

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02878-z

Keywords: entity resolution, entity matching, SQL databases, NoSQL databases, data integration, data quality, deep learning, large language models, blocking techniques, graph databases, record linkage, database management systems

Brenda Betancourt, a Term Associate Professor of Statistics in George Mason’s College of Engineering and Computing, has received $286,130 from the U.S. Food and Drug Administration for a project titled Regulatory Grade Frameworks for Assessing Record Linkage Uncertainty in Real-World Evidence. The funding began in September 2026 and runs through September 2028. Over those two years, Betancourt will evaluate how uncertainty in record linkage affects the reliability, interpretability, and evidentiary confidence of analyses based on linked real-world data that are used to support FDA regulatory decision-making. In other words, the project asks a deceptively simple question: if the links between records are not certain, how certain can the conclusions drawn from them be?

The stakes are considerable. Record linkage is the invisible plumbing of real-world evidence. When a researcher wants to know whether patients exposed to a medication later experienced a particular adverse event, the exposure records typically live in one database, perhaps pharmacy claims, while the outcome records live in another, perhaps a hospital discharge registry. Joining those datasets requires deciding which rows refer to the same individual, often without a shared unique identifier such as a national patient number. Probabilistic matching algorithms compare names, dates of birth, addresses, and other fields, assigning scores that indicate how likely two records are to describe the same person. Every one of those assignments carries a probability of error, and those errors propagate silently into every downstream estimate.

Statisticians have long understood that linkage errors can bias results in either direction. False matches, in which records from two different people are merged, can dilute genuine signals or manufacture spurious associations. Missed matches, in which records from the same person remain split, can fragment a patient’s medical history and undercount exposures or outcomes. The direction and magnitude of the bias depend on the nature of the analysis, the matching variables, and the population being studied. What has been missing, particularly in the regulatory context, is a standardized framework that quantifies this uncertainty in a way that meets the evidentiary bar regulators demand when a decision affects public health.

That is the gap Betancourt’s project is designed to fill. The phrase regulatory grade in the project title signals the ambition: not merely to measure linkage uncertainty in an academic sense, but to develop frameworks rigorous and transparent enough to be used in submissions that inform FDA decisions. Regulatory science operates under constraints that academic research often does not. Methods must be reproducible, their assumptions explicit, their outputs interpretable by reviewers who must weigh evidence under statutory standards. A framework that cannot communicate how much of a reported effect might be attributable to linkage error is of limited use to an agency deciding whether to approve, restrict, or warn about a therapy.

The mathematical core of the problem is rich. Probabilistic linkage models, descendants of the classical Fellegi-Sunter framework, estimate the likelihood that a pair of records matches based on agreement and disagreement patterns across comparison fields. More modern approaches bring in Bayesian modeling, clustering, and machine learning classifiers, each producing not a single deterministic assignment but a posterior distribution over possible linkages. The challenge is propagation: how does uncertainty in those assignments flow through data cleaning, cohort construction, statistical modeling, and ultimately into confidence intervals and p-values for the regulatory question at hand? Ignoring that propagation can yield analyses that appear far more precise than they actually are, a phenomenon statisticians describe as understated uncertainty.

Several strategies exist in principle for handling this propagation. One is to condition on the most probable linkage and then adjust estimates using sensitivity analyses that vary the assumed error rates. Another is multiple imputation over plausible linkages, generating several alternative versions of the linked dataset and combining the resulting estimates so that the final inference reflects linkage ambiguity. A third embeds the linkage model and the analysis model in a joint Bayesian framework, allowing uncertainty to be integrated out rather than fixed. Each approach carries computational costs and modeling assumptions, and each performs differently depending on data quality, record volume, and the discriminative power of the matching fields. A central task for the project is evaluating when these methods deliver trustworthy inference and where they fall short.

The timing of the work reflects a broader shift in how evidence reaches regulators. The FDA has invested heavily in real-world evidence programs, including the RWE framework established under the twenty-first Century Cures Act, which directed the agency to clarify how real-world data can support approvals and label changes, particularly in areas such as rare diseases, oncology, and post-market safety surveillance where randomized trials are impractical or unethical. As the volume of linked real-world analyses grows, so does the need for standards governing every step of the pipeline. Linkage is among the earliest and least visible of those steps, which makes its errors especially insidious: they are baked in before most analysts ever see the data.

For patients and clinicians, the practical payoff of this research is confidence. If a safety signal emerges from linked claims and hospital data, clinicians and patients deserve to know whether that signal could plausibly be an artifact of mismatched records. Conversely, if a real-world analysis finds no elevated risk, regulators should know how robust that null finding is to plausible linkage errors. Quantified linkage uncertainty turns these questions from matters of intuition into matters of measurement, allowing reviewers to weigh evidence with a clearer sense of its fragility or strength. In an era when real-world evidence increasingly shapes drug labels, coverage decisions, and clinical guidelines, that clarity is not a technical luxury but a foundation of trustworthy medicine.

The project also highlights the growing role of statisticians in regulatory science and the position of universities near the federal policy apparatus in shaping it. George Mason, Virginia’s largest public research university, enrolls more than 40,000 students and sits near Washington, D.C., placing its researchers in close proximity to the agencies whose standards define the evidentiary landscape. Over the next two years, Betancourt’s work will contribute to a question that touches every linked dataset behind every regulatory submission: how sure can we be that the records we joined belong together, and how should that residual doubt be reflected in the evidence we act upon. The answer may help determine how much trust real-world evidence ultimately earns.

Subject of Research: Quantifying record linkage uncertainty in real-world evidence for FDA regulatory decision-making

Article Title: Betancourt studying regulatory grade frameworks for assessing record linkage uncertainty in real-world evidence

Article References: Betancourt studying regulatory grade frameworks for assessing record linkage uncertainty in real-world evidence. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: record linkage, real-world evidence, FDA, regulatory science, statistics, data uncertainty, George Mason University, probabilistic matching, pharmacovigilance, electronic health records, Bayesian methods, drug safety