Bonten Poad: Generating fake Vermont placenames with a neural network
Y’all, we’ve entered a computing epoch when AI is available to us at nearly the level of commodity entertainment. Take this experiment in generative British placenames by Dan Hon; his workflow is 7 goddamn steps to use technology only the likes of Alan Turing could have dreamed up, until a few years ago.
Dan used a multi-layer recurrent neural net to come up with imaginary locales like “Buchraston-on-Ter-Sey”, so I thought I give it a shot and see what a complex algorithm could do with a state that includes “Ira”, “Lympus”, and “Satan’s Kingdom” IRL. Here are the choicest results of the several thousand I generated; they look more Welsh than I would have expected:
- Cellen Boneyboas
- Macih Pond
- Pendoil
- Ralas Port
- ClintCtow
- Wi
- Wont
- Cissham Bond
- Whortdlay Pond
- Sad Cahue
- Coal Bontewlu Porn
- Bent Pond
- Hicton
- Mash dank Mount
- Goryhy Pond
- Ha
- Dondslon
- Dorits Wat Poon
- Cold Bamronr
- Wot Ponde
- Claylelon
- Golen
- Cllen Boneyboas
- Rut anvill
- Ponk
- Chomtax Bastarr
- Lentrenston
- Wershrok Modanton
- Bonten Poad
- Ponkey
- Wasdesh
- Vontfalilina
- Kart Pond
- Poon Welton
- Hate Zonord
- Cowrelu
- Mu
Now, these don’t exactly meet my not-totally-formulated standard for “Vermontyness”, but that’s the nature of machine learning in this case: it’s about figuring out the question using a series of answers. And to be fair, Vermont’s placenames aren’t that distinct (at least by the standards of the Northeastern U.S.): mostly English place and family names, with a very small smattering of French and Abenaki. Also, the fact that both “Ponk” and “Ponkey” show up here tells me that my training set might be a tad small — only ~2000 placenames went into the model — but overall this is good fun.
Many thanks to Justin Johnson (whose work I’ve benefitted from before) and Cristian Baldi for making such accessible tools in the brave new world of machine learning.
Try it out yourselves: here’s a docker image with everything you need!
