Research
Written work from operating LinkedCulture: a peer-visible paper series documenting what a live multi-institutional retrieval system actually does, and shorter working notes along the way. By Edward Monk.
RSS feed · Papers are archived on Zenodo with DOIs; the pages here are the canonical web editions.
September 17, 2026
The print that every museum owns.
Hokusai's Great Wave sits in at least four LinkedCulture institutions at once, and Rembrandt's Three Crosses in five. That is not redundant collecting. It is what a print actually is, and the record says less about it than it should.
September 5, 2026
If your records are not in English, the model will not notice.
I generated 3,468 image descriptions for a French national collection and every one came back in English, because the prompt was written in English and nothing else was. How that happens, how to store both languages, and why the English ones are worth keeping.
September 4, 2026
Describing 5,228 museum objects from the photograph alone.
I generated a description for every image in the Museo Egizio's online collection using only the photograph, no title and no catalog record. For about 900 objects the museum had already written something better. For 3,310 there was nothing at all.
September 2, 2026
What automated image descriptions cost, at 5,000 objects and at 500,000.
The same accessibility job costs nothing at one scale and hundreds of dollars at another, and the right tool is completely different at each. Measured prices and throughput as of September 2026, for collections managers deciding whether this is affordable.
August 30, 2026
What it took to describe half a million photographs.
The first full attempt at generating image descriptions across the whole corpus: which model, what it cost, where the output failed, and the parts of the process I got wrong. Superseded on cost by a later note, kept for the working notes.
August 28, 2026
Mapping 5,525 places museums never recorded as coordinates.
Museum records often know where they are without saying so: a caption in period French, a spelling nobody uses anymore, an excavation hierarchy. Turning 5,525 of those into map points, and the places the process went wrong.
August 14, 2026
The Blue Bird probe.
A Jean Paul Gaultier dress from Paris Musées revealed that multilingual embedding and cross-lingual retrieval are not the same thing. This is how that finding changed LinkedCulture.
August 11, 2026
Three ways a spelling correction can be confidently wrong.
Building Google-style spelling correction for LinkedCulture surfaced three separate bugs, each one a plausible-looking answer that was actually wrong. None of them showed up until real misspellings were thrown at it.
August 10, 2026
Semantic search across institutions.
The Met's semantic search launch is a landmark moment for the field. Here is what becomes possible when that same capability crosses ten institutions, languages, and borders.
July 21, 2026
The museum that broke our search.
Museo Egizio's onboarding exposed two separate retrieval bugs that had been sitting latent in LinkedCulture's hybrid search since before it joined. Neither was its fault.
July 19, 2026
Museo Egizio joined LinkedCulture.
Museo Egizio is now live in LinkedCulture: 5,228 image-backed Egyptian archaeological records indexed across Postgres, Qdrant, and OpenSearch.
July 18, 2026
Why LinkedCulture exists.
LinkedCulture has crossed 500,000 objects across nine institutions. This is why it exists, what it is trying to do, and who it is for.
July 17, 2026
LinkedCulture just crossed 500,000 objects.
The unified semantic index now spans 500,000+ open-access records across nine museums, from the Rijksmuseum to the Minneapolis Institute of Art.
July 17, 2026
One list, three museums, no account required
LinkedCulture now lets anyone build their own cross-institution collection, no sign-up, just a private link that's theirs alone, and a public link they can share once the list is real.
July 17, 2026
The photos that weren't actually missing
A batch of CC0 Getty photographs looked permanently broken on LinkedCulture. They weren't. A single character in an image-sizing request was the difference between a working photo and a blank box.
July 14, 2026
A Bridge in the Rain: How a Dying Man's Print Crossed the World
Van Gogh's copy of a Hiroshige print is the famous case, and one of three he made that year. Degas's hundred-print collection, Cassatt's turn to a new medium after seeing 700 Japanese prints in one room, and Toulouse-Lautrec's Japanese signature are the same story, several times over, across institutions on two continents.
July 13, 2026
Every object page has a printable QR code
Two QR codes on every LinkedCulture object page, one back to the object, one out to the museum's own record, for anyone who wants to point a phone at something instead of typing.
July 12, 2026
Same artist, cataloged differently at every museum
A review queue for merging duplicate artist records across museums when the match is a name, not a shared authority ID.
June 29, 2026
Behind the scenes, LinkedCulture.org represents cultural knowledge in three complementa...
• Vector embeddings power semantic discovery, helping you find related objects even when they don’t share the same keywords.
June 29, 2026
Where in the world should LinkedCulture go next? 🌍
If you could add one museum, gallery, archive, or cultural heritage institution to LinkedCulture’s cross-institutional search, which would it be and why?
June 24, 2026
Semantic Clusters and Language Partitioning in a Multi-Institutional Cultural Heritage...
Most multilingual AI models don't actually connect languages. They search within them.
June 24, 2026
Semantically derived collection clusters cross institutions effortlessly. Across langua...
That is the finding in Paper 3 of the LinkedCulture series, and it overturns something I assumed in Paper 2.
June 22, 2026
Museum Collections Are Digitized, So Why Can't Anyone Find Anything?
Cultural heritage collections have never been more accessible. Digitized records. Public interfaces. Decades of metadata work.
June 21, 2026
Two more world-class collections have joined LinkedCulture.
They join the Getty, the Art Institute of Chicago, the Rijksmuseum, the Cleveland Museum of Art, Harvard Art Museums, and The Met, bringing the total to 300,000+ objects across 8 institutions.
June 21, 2026
What do 300,000 museum objects look like when you map them by meaning? A galaxy.
I embedded every object in LinkedCulture (300,000+ works across 8 institutions) and let them arrange themselves by what they actually are, not by who made them or when. The result is an explorable...
June 13, 2026
I’ve been building LinkedCulture — a project that finds connections across museum colle...
Full transparency: two of the co-hosts (Eulalia and Shay) are AI-generated avatar personas — I’m also testing AI video pipelines as part of this. The art history itself is fact-checked and sourced...
June 11, 2026
Before the research papers. Before the architecture arguments. There are the objects th...
The Edward Monk YouTube channel has been quietly publishing narrated descriptions of public domain museum objects from the LinkedCulture collection. Over 100 so far. Each image described by AI and...
June 11, 2026
Cultural heritage collections are more digitized than ever.
And researchers, students, and curious members of the public still leave without finding what they came for.
June 11, 2026
Cultural heritage collections are more digitized than ever.
And researchers, students, and curious members of the public still leave without finding what they came for.
June 11, 2026
Cultural heritage collections are more digitized than ever.
And researchers, students, and curious members of the public still leave without finding what they came for.
June 11, 2026
The first LinkedCulture paper made the argument. The second paper shows how the system...
I've published Paper 2 in the LinkedCulture Research Paper Series on Zenodo: Embedding Cultural Heritage Metadata: Pipeline Design, Multilingual Retrieval, and Hybrid Search in LinkedCulture.
June 3, 2026
LinkedCulture ahora narra su colección en español.
El proceso completo es automatizado. LinkedCulture obtiene la imagen del objeto y los metadatos institucionales, Claude escribe una descripción de cuatro frases: nombra la técnica, describe lo que...
June 2, 2026
LinkedCulture now narrates its collection.
The entire process is automated. LinkedCulture pulls the object image and institutional metadata, Claude writes a four-sentence description, naming the technique, describing what is visible, ground...
May 28, 2026
A perfectly cataloged museum record can still be practically undiscoverable. Not becaus...
That gap is structural. No amount of metadata improvement closes it. Neither does fuzzy search.
May 24, 2026
LinkedCulture slideshows are now embeddable on your website! 😎
A LinkedCulture topic or search result becomes a live, cross-institutional slideshow anywhere HTML runs: one script tag, one custom element.
May 20, 2026
Just shipped: LinkedCulture is now hybrid search with multilingual support.
Here's the problem I solved: If you search semantically, you find paintings that shouldn't be together. If you search by keyword, you miss connections nobody explicitly tagged.
May 17, 2026
Expériences du dimanche avec les collections françaises.
Voici "L'Oiseau Bleu," une robe de Jean Paul Gaultier conservée au Palais Galliera, musée de la Mode de la Ville de Paris. Elle est apparue ce matin comme premier résultat pour la recherche "oiseau...
May 17, 2026
Search. Click. Slideshow.
#LinkedCulture turns any search result into an instant slideshow. These are three different searches running at the same time, each one a different collection, each one a single click from a browsa...
May 17, 2026
Sunday experiments with French museum collections.
This is "L'Oiseau Bleu," a Jean Paul Gaultier gown from Palais Galliera, musée de la Mode de la Ville de Paris. It surfaced this morning as the first search result for "oiseau bleu" from a pilot te...
May 12, 2026
LinkedCulture has a new milestone.
The main LinkedCulture semantic search index now includes more than 220,000 open-access museum records across six collections, including the Art Institute of Chicago, Rijksmuseum, Cleveland Museum...
May 11, 2026
A thoughtful post about LinkedCulture from the French museum innovation community pushe...
LinkedCulture now has a first multilingual test page: semantic search in French across more than 50,000 Getty open-access records, indexed with a multilingual embedding model.
May 7, 2026
The Getty just joined LinkedCulture.
50,000 (was 10k) Getty objects are now indexed and searchable alongside the the other +220K Met, Harvard Art Museums, Cleveland, and the Rijksmuseum collections ingested into #LinkedCulture. Search...
May 7, 2026
What if you could search a museum collection by what's actually in the image, not just...
LinkedCulture Vision does exactly that. Using AI image captioning across Met and Harvard Art Museums collections, we've generated semantic descriptions for over 1,000 objects that previously had li...
May 5, 2026
Built a semantic search engine that searches across museum collections simultaneously,...
The same 17th-century etching of a cow, held by both Harvard Art Museums and the Art Institute of Chicago, surfaced side by side by semantic search across ~170,000 open-access works.
May 4, 2026
47 objects. 4 institutions. 250 years of Europeans putting extraordinary craftsmanship...
A LinkedCulture topic cluster on European decorative watches, spanning the 17th to 19th century across the Rijksmuseum, the Art Institute of Chicago, the Cleveland Museum of Art, and the Met.
May 2, 2026
60 objects. 4 institutions. One shared subject: Hindu deities and mythology.
A LinkedCulture topic cluster spanning roughly 1,500 years of South and Southeast Asian art, from a 2nd-century Kushan relief to a 10th-century Angkor bronze.
May 1, 2026
This is the original heavy metal.
An algorithm found 500 pieces of European Renaissance armor scattered across four museums with no shared catalog, keyword, or curator connecting them.
April 30, 2026
I asked no one for this.
An algorithm grouped Japanese kabuki actor prints from three museums that catalog them differently, connected by nothing in the metadata and everything in the meaning.
April 23, 2026
I've been bad at thinking out loud publicly. This is me fixing that.
First video. Channel introduction. Two minutes.