Semantic Clusters and Language Partitioning in a Multi-Institutional Cultural Heritage Index

This paper presents observations from LinkedCulture, an experimental cross-institutional cultural heritage retrieval prototype. It examines what happened after the system adopted multilingual text embeddings to support discovery across collections described in different languages. The central finding is that multilingual retrieval and cross-lingual representation are not the same thing. In a 302,798-record corpus, French-language records clustered overwhelmingly with other French-language records, while English-language records clustered with English-language records, even when records described conceptually related material. Model bake-off experiments showed that alternative multilingual embedding models reduced cross-language distance but did not eliminate the partition or produce reliable cross-language ranking. The paper argues that cross-language discovery in cultural heritage indexes is presently an orchestration problem rather than a solved representation problem. Query translation, hybrid retrieval design, and ranking strategy remain necessary to bridge language-partitioned embedding spaces.