Published research behind the methods DarkMath uses. Summaries are ours; DarkMath is not affiliated with the authors.
Record linkage
Arora et al., 2024
An open-source package that applies transformer language models to record linkage, treating linkage as a retrieval problem over learned text embeddings, with support for fine-tuning on labeled pairs.
Why it matters here. The basis for semantic matching: compare what records mean, not how they are spelled.
Read the paperEntity resolution
Li et al., 2024
A framework that asks a language model a small number of well-chosen matching questions and uses the answers to reduce uncertainty in entity resolution results within a budget.
Why it matters here. Why DarkMatch spends language-model review only on the hardest, most uncertain cases.
Read the paperLookalike modeling
Peng et al., 2023
A two-tower deep learning system that builds customer embeddings from diverse data and uses them to find lookalike customers at scale for marketing campaigns.
Why it matters here. The pattern behind DarkMirror: embeddings of best customers, used to find more people like them.
Read the paperSee how these ideas hold up on your own records.
