This paper introduces the concept of “translation mining”, a cross-lingual embedding approach that bridges close reading and large-scale quantitative analysis of historical translations. Focusing on eighteenth-century British (ECCO) and French (Gallica) corpora, we systematically identify semantically aligned passages across hundreds of thousands of documents. The approach provides a comprehensive AI-driven methodology for locating translations of various types. This enables us to build a multi-scalar, data-driven taxonomy of translation practices that offers finer granularity than prior dichotomies of “full” versus “partial” translation. By embedding all texts once using high performance computing, we can iteratively detect subtle textual connections across languages without re-running the entire process. This paper demonstrates how embedding models derived from large language models can capture cross-lingual translation pairs, challenge rigid classifications, and illuminate multi-level cultural transfer, offering an adaptable framework for historical research.