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

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

Paper 3 is open access on Zenodo: https://linkedculture.org/research/semantic-clusters-language-partitioning

Most multilingual AI models don't actually connect languages.

They search within them.

I discovered this the hard way while building LinkedCulture, a cross-institutional cultural heritage search platform indexing over 300k records across eight museums.

The setup seemed solid. Multilingual embedding model. Hybrid retrieval. A "galaxy" visualization showing semantic clusters across the collection. Then I searched for "blue bird" in English.

The system returned nothing from the French collection. No Jean Paul Gaultier dress titled L'Oiseau Bleu. No French records at all. The concept was identical. The language was different. That was enough.

So I ran the numbers.

French-language records had nearest neighbors that were 99.9% French, despite French comprising only 17% of the corpus. English records were equally siloed. The embedding space wasn't a unified semantic map. It was two separate continents with a wide ocean between them.

Switching to a better multilingual model didn't fix it. I tested five alternatives. Every one failed the same way. Records got slightly closer across the language boundary, but not close enough to actually surface in results.

What worked was simpler: translate the query, search again, merge the results. Not a model problem. An orchestration problem.

This is the central finding of Paper 3 in the LinkedCulture Research Paper Series: Semantic Clusters and Language Partitioning in a Multi-Institutional Cultural Heritage Index.

The implication for anyone building multilingual retrieval systems is direct: don't evaluate your system only by what it returns. Evaluate how the underlying space is structured. A system can appear to work at the surface while remaining partitioned underneath.

Multilingual is not cross-lingual.

The research paper is part of a series: Paper 1: Discovery architecture for cultural heritage: layered retrieval, institutional authority, and the limits of keyword search https://linkedculture.org/research/discovery-architecture-for-cultural-heritage

Paper 2: Discovery architecture for cultural heritage: layered retrieval, institutional authority, and the limits of keyword search https://linkedculture.org/research/embedding-cultural-heritage-metadata

Paper 3: Semantic Clusters and Language Partitioning in a Multi-Institutional Cultural Heritage Index https://linkedculture.org/research/semantic-clusters-language-partitioning

Learn more about LinkedCulture and the work behind it: https://linkedculture.org/

See the galaxy interactive: https://galaxy.linkedculture.org/