A neural network tries to generate townlands of Ireland
There’s a type of machine learning algorithm called a neural network, which learns to imitate the patterns it finds in datasets. When I’ve trained neural networks on names in English, they learn to match the sounds that make a word sound like a guinea pig name, or a name for a metal band (or, delightfully, both). They can also learn to mash together nouns to make Halloween costumes, or (with limited success), learn to tell a knock-knock joke.
This trick works in other languages as well - and thanks to a couple of collaborators, I had the chance to learn to see if they could speak Irish.
In Ireland, land is divided into townlands - about 65,000 in all. Townlands have been around since at least the 12th century, and most of the names are in Irish, a Celtic language largely unrelated to English. Mostly they describe some feature of the landscape. The townland name “Cornamucklagh”, for example, is an anglicization of the Irish “Corr na muclach”, which translates to “Round Hill of the Piggeries”.
I gave the full list of 65,000 anglicized townland names to a neural network to see if it could generate names that would make sense in Irish. This task is made a bit easier by the fact that the anglicizations weren’t done in any standard way - the word for meadow, “cluain”, has been anglicized as "clon", “clun”, “cloon”, “cloyn”, or “cloan”, among others. Without a standard spelling, there’s more room for the neural net to get away with errors. And thanks to generous translation help from Robert Chapple (who provided the dataset) and Martin O’Leary (who generated the gorgeous maps you see here), I had a chance to find out what the names mean.
The neural net’s early tries, however, generally didn’t mean anything in particular, although they did at least use some of the same sounds. This also settles the question of whether a neural net will come up with “fart” when working in a language other than English.
After it had trained for a while, however, its names started to make sense! These aren’t real places in Ireland but they could be translated. As an example, here’s one of my favorites: The neural net generated “Ballydoorock” which would probably translate to Baile an Dobhareach (town of the hippopotamus). Here are some more:
Knocknacooshy - Cnoc na gCuaisigh (hill of the hollows)
Inchadomeen - Inis an Tuaimín (island of the tumulus)
Tullywart - Tulach an Bharda (hill of the garrison)
Ballydoorock - Baile an Dobhareach (town of the hippopotamus)
One last thing I tried was to train the neural network with a much shorter memory (5 characters instead of 15) to see what would happen. Delightfully, the neural net would often lose track of when to stop writing a name, and would string together names of epic length.