LinkedCulture

This paper describes the implementation behind LinkedCulture, an experimental cross-institutional cultural heritage retrieval prototype. It explains how structured cultural heritage metadata is translated into embeddable prose, how records from eight institutions are normalized into a canonical schema, how OpenSearch and Qdrant support keyword and semantic retrieval, and how Reciprocal Rank Fusion combines the two ranking signals.

The paper documents a transition from English-centered to multilingual embedding prompted by an observed cross-language retrieval failure, and reports preliminary observations on cross-lingual retrieval behavior in the extended index. It also examines where pipeline design choices introduce potential bias, where embedding models fail in predictable ways, and how metadata density, institutional imbalance, licensing constraints, and local infrastructure choices shape retrieval behavior.

This is the second paper in the LinkedCulture research paper series.

The first paper in the LinkedCulture series argued that keyword search alone is insufficient for cultural heritage discovery, because the vocabulary of the cataloger and the vocabulary of the searcher often do not overlap. The second paper described the implementation of a hybrid retrieval system combining keyword search and semantic similarity across a multi-institutional cultural heritage corpus. That implementation adopted a multilingual embedding model after an English-centered embedding setup produced a visible retrieval failure when non-English records entered the index.

This paper reports what happened after that change was tested at larger scale. The central finding is that multilingual retrieval and cross-lingual representation are not the same thing. The multilingual embedding model used by LinkedCulture made records retrievable within multiple languages, but it did not produce a unified cross-language semantic space. French-language records overwhelmingly clustered with other French-language records, while English-language records overwhelmingly clustered with English-language records, even when records from different languages described conceptually related material.

The finding emerged from analysis of a corpus of 302,798 cultural heritage records, including a French-language bloc large enough to make language partitioning visible as structure rather than noise. Nearest-neighbor analysis showed near-total language siloing: French records' nearest neighbors were 99.9% French despite French records comprising only 17.0% of the corpus. Model bake-off experiments further showed that alternative multilingual text embedding models reduced measured cross-language distance but did not eliminate the partition or produce reliable cross-language ranking. Query translation, rather than model substitution, produced the strongest cross-lingual retrieval improvement.

The paper also documents a related operational evolution in LinkedCulture's hybrid retrieval layer: the replacement of Reciprocal Rank Fusion with normalized-score fusion and a coordination factor after production use revealed that RRF could elevate weak candidates in sparse-metadata environments. Together, these findings suggest that practical cross-institutional cultural heritage retrieval depends less on selecting a single superior model than on designing orchestration layers that compensate for the structural behavior of models, metadata, and ranking systems. For multilingual cultural heritage indexes, cross-language discovery is presently an orchestration problem, not a solved representation problem.