Labeling Training Data for Entity Matching Using Large Language Models

New research investigates whether large language models can label training data for entity matching, a step that could reduce the manual effort required to build custom matching systems [1]. Entity matching — determining whether two records refer to the same real-world entity — is a core task in natural language processing (NLP) and data integration [3][4]. Traditional approaches rely on machine learning models or small language models (SLMs) such as RoBERTa, which require task-specific training data to perform well [1]. Recent large language models (LLMs) can handle entity matching without that data, but applying them to large candidate sets is slow and expensive [1]. A paper submitted to arXiv on 27 June 2026 examines knowledge-distillation workflows in which an LLM acts as a teacher, labeling training pairs that are then used to train a smaller student model [1]. The authors tested the approach across five standard benchmarks: Abt-Buy, Walmart-Amazon, WDC Products, DBLP-ACM, and DBLP-Scholar [1]. Language model benchmarks are standardized tests that provide datasets and evaluation metrics to compare model performance on tasks such as text classification and reasoning [2]. The experiments found that student models trained on machine-labeled data performed roughly on par with those trained on the original benchmark training sets; differences in either direction stayed below two F1 points [1]. Using GPT-5.2 to label the training sets for all five benchmarks cost between US$28.31 and US$40.88, while manually labeling the same sets was estimated to require 470 hours of work [1]. At inference time, the entity-matching tool Ditto proved 41.5 to 534 times faster than using an LLM directly [1]. The work sits within a broader push to make machine learning more accessible. Platforms such as Hugging Face Spaces now allow researchers to share interactive demos alongside arXiv papers, letting users test models without writing code [7][8]. The integration, launched in collaboration with arXiv, has hosted over 12,000 open-source demos since October 2021 [7]. Meanwhile, the cost of training large models has drawn industry attention: Chinese firm DeepSeek reported training its V3 model for US$6 million, far below the reported US$100 million cost of OpenAI’s GPT-4 in 2023 [10]. The paper’s findings suggest that combining LLMs with a suitable pair-selection method can substantially reduce or eliminate the manual labeling effort for entity-matching training data [1].