Health, vitality, and trust: embedding-based cross-lingual retrieval of medicinal advertisements in European and colonial press

This paper presents a case study of medicinal advertisements published in the early twentieth century across European and colonial press. It offers a critical reflection on the retrieval and evaluation of multilingual corpora within large-scale media archives, with particular attention to their validity and relevance for historical research. The study develops a reproducible workflow for embedding-based retrieval, focusing on keyword-level queries as a first stage of cross-lingual corpus construction. Retrieved materials are evaluated through a two-step protocol combining cosine similarity ranking and expert close reading, enabling a comparative assessment of semantic precision and discursive coherence across selected keywords. The paper argues that embeddings provide a powerful alternative to string-based search in multilingual and orthographically unstable corpora, while demonstrating that quantitative cluster stability is a necessary but insufficient condition for meaningful retrieval. This finding has particular implications for languages such as Vietnamese, where compound terms are prone to lexical and semantic drift in cross-lingual vector spaces.