We’re all over the map — how we geolocated SMK artworks with the kind help of humans and machines
Burning churches, rolling hills, and ragged seascapes — we’ve placed them all on a map of Denmark. With a companion social media campaign, this may be the most fun we’ve had with data to date.
In March 2021 we placed 4.000 SMK artworks on a map of Denmark. We did it by combining existing data with machine learning and the help of kind human beings. And we did it by leveraging the SMK API which will ultimately ingest the refined location data to allow future re-use.
That, dear reader, was the abridged edition. For the hows, the whys and the what nows, please scroll on!
The agony of size
The SMK collection holds some 260.000 artworks. Inspiring? Sure, but also potentially terrifying. Where to begin? How to navigate? As a regular Danish person, you may have no obvious point of reference.
Enter: The map. A map makes things decodable and very personal. You were born somewhere specific. You grew up somewhere specific. You live somewhere specific. A painting portraying one of these sites is immediately identifiable. It’s something you may know very well indeed.
Now, seeing a well-known place rendered through the eyes of art may induce memories. But it may also inspire you to see a landscape afresh. And in the time of COVID, it may let you travel virtually or supply new inspiration for your physical weekend trip.
Putting paintings in their place
Having established the virtues of cartography, how to proceed? SMK’s collection is registered in the national Danish collections database SARA, an acronym referring to the prosaic “collections registration and management”. Some artworks have rudimentary location data about what they depict such as “Northern Zealand” or “Copenhagen”. Often not very precise, but a place to start!
To narrow things down we analysed artwork titles using a Named Entity Extraction technique with a Multilingual Bert model*.
Machines were tasked with teasing out place names which, combined with the museum keywords, were run through Google Maps API and the artworks were placed on a Google map. For instance, John Christensen’s Folkeliv på Kapelvej (“Street life in Kapelvej”) was now placed, not just in the area Nørrebro in Copenhagen, but on the actual street named Kapelvej.
Importantly, we limited the results to Denmark (including Greenland and the Faroe Islands) and excluded certain location types such as cafes (which are often named after places, bound to cause confusion).
The level of precision achieved? Useful, sometimes spot-on, but not always convincing. We found close to 4.000 places but often the location was very general.
For Vilhelm Hammershøi’s Landscape on the Island of Falster all we could find was “Falster” and so the painting was placed in the center of that fine island even though it actually seems to be painted by the sea.
Now, obviously the dataset wasn’t extremely daunting and we could have refined the locations manually ourselves. But while common sense might improve precision somewhat we are hardly experts on every specific hillside, park bench or cliffside. But out there, someone is bound to be! So we published the map and asked the public for help.
Artificial intelligence meets local expertise
On 17 March 2021 we published the map. At the same time we started up a dedicated Facebook group and sent out a general press release. In the Facebook group we invited members to join us on a trip across the country from East to West and as this trip (consisting of posts of highlights from the area) progressed we contacted local news outlets and relevant Facebook pages or groups. We emphasised two attractions of the map: See your hometown through the eyes of art (for the general public) and Help us improve the map (for the more dedicated user).
Together, these efforts stirred up considerable attention. A range of local media picked up the story, and as of 9 April 2021 we’ve had 16.000 visits. Even more interesting, I think, is the fact that our geographical reach is far more evenly distributed than we’re used to. We often tilt strongly towards Copenhagen but this time only 35% of users are in the capital (the home of 23% of the population).
Initial feedback was quite overwhelming. Also, it was chaotic. We’d — naively! — planned to manage dialogue either through the built-in features of the map interface or the dedicated Facebook group. A nice plan, but immediately emails started pouring in in the form of social media comments, DMs to our private accounts, emails to the museum’s main address and even to our director.
Sure, there were concrete suggestions for moving map pins. But also, of course, there were questions about digitisation practice, image resolution, requests for clarification and (sometimes kind, sometimes less kind) suggestions that we read up on geography. Honestly, they were all extremely welcome and I only hope that not too many fell through the digital cracks.
Many commented that locations with identical names had been mixed up. Many also commented that Google had sometimes guessed a little liberally when places no longer existed (“Sydstrup” no longer exists, so Google had guessed “Skrydstrup”). But perhaps the most interesting comments were ones that helped us clean up errors in our core data. Apparently, when transcribing venerable old paper cards, some museum worker — at some point in time — had assumed that “Femø” (south of Zealand) should be “Fanø” (west of Jutland). And somehow we’d typed in “Kalvø” instead of “Kalø” in the title of Janus La Cour’s A Showery Landscape. Kalø Vig, Jutland placing it around 160 kilometers too far South. We’ve essentially (and rather accidentally) crowdsourced the proof-reading of our metadata, and I do think this qualifies as something of an eye-opener.
Where do you want to travel next?
We’re continuing our journey Westwards (and learning tons about Danish geography as we go along). When we’ve reached absolute (Danish) west, we’ll work on importing all the new coordinates into our API, and perhaps also into the collections database itself.
Meanwhile, many have asked why on earth other museum collections aren’t on the map.
Not surprisingly, people don’t really care which museum semi-arbitrarily stores which artworks — they care about art. Including other collections is way beyond the scope of our project, but the logic is totally sound. In Denmark, we’ve recently struggled to merge most museum collections databases into one and while I realise that this task was momentous in itself, of course the next step should be linking up the data in new and useful ways. We hope our little map may provide some inspiration for how to work with the nation’s museum data in the future.
- I’ll be happy to pass on any technical questions you may have…
The team
- Christina Jensen
- Nikolaj Erichsen
- Merete Sanderhoff
- Jonas Heide Smith
- Michala Rosendahl
With very special thanks to Sofie Glargaard, Kim Brasen, Tanja Larsen and others.
