LinkedCulture — the collection as a galaxy

The collection as a galaxy

Each dot is an object from a museum. Objects about similar things are grouped into clusters, and related clusters sit close together, so the distance between two clusters shows how alike their subjects are.

• DragTurn the galaxy around.
• ScrollZoom in and out. On a trackpad, slide two fingers or pinch. Get close to a cluster and its dots turn into photos.
• ↑ ↓Zoom in and out from the keyboard.
• ← →Turn the galaxy left and right. Hold Shift to pan sideways instead.
• SpacePause or resume the turning.
• RJump to a random cluster.
• Ctrl + DragPan sideways. Holding Shift or ⌘ works too, and so does dragging with the right button.
• ClickA topic name opens it. A dot that gets a circle when you point at it opens that object.

• DragTurn the galaxy around with one finger.
• PinchZoom in and out. Get close to a cluster and its dots turn into photos.
• Two fingersDrag with two fingers to pan sideways.
• TapA topic name opens it. A dot opens that object.

Search for anything and the matching topics light up. Switch the colours between cultures and museums at the bottom left. The gear button holds the legend and view options.

How this map was made

Each dot is a catalogue record from one of nine museums: the Art Institute of Chicago, Cleveland Museum of Art, Getty Museum, Harvard Art Museums, Minneapolis Institute of Art, The Metropolitan Museum of Art, Rijksmuseum, Paris Musées and Museo Egizio.

1. Embedding. The text of each record (title, type, culture, date, materials, description) was turned into a list of numbers with a multilingual model, nomic-embed-text-v2-moe. The pictures were not used.
2. Clustering. The numbers were reduced to 50 dimensions with UMAP and grouped with HDBSCAN into 2,499 clusters. About 46% of the dots belong to no cluster and show as faint points.
3. Names. A language model (Claude Haiku) wrote each cluster's name and description from a sample of its records and the clusters beside it. Names can be wrong; the objects are the evidence.
4. Layout. Clusters are placed in 3D from how close their centres are, so distance is a guide, not a measurement. Two objects that read alike can sit far apart.

French and English catalogues drift apart on the map largely because they are written in different languages, not because the objects differ. We left that visible on purpose.

The method, the measurements and the language finding are in the paper:
Semantic Clusters and Language Partitioning in a Multi-Institutional Cultural Heritage Index
More about LinkedCulture