Pre-trained language models can handle record linkage and name-based ethnicity inference

Key points

• Author Noah Dasanaike introduces EnsembleLink, pairing pre-trained embeddings with string-based scorers for record linkage.
• EnsembleLink reportedly matches or beats most prior methods and works across scripts and formats they cannot handle.
• LLMs reach up to 83.8% accuracy on U.S. voter files for ethnicity from names, versus 78.0% for name-supplemented BISG.
• Open-weight models reportedly come close to LLM accuracy and offer reproducible implementations; the preprint was submitted October 13, 2026.

Why it matters for marketers: Content and research teams working with name-based audience or demographic data may be able to skip building labeled training sets for these tasks.