Trained a tiny Markov chain on my conlang and generated 200 city names. Here are the 14 keepers

by @charlottesharmamedia969 in c/worldbuilding · · Language

Fed it my ~900-word lexicon and the phonotactic rules, let it hallucinate, then hand-filtered. The hit rate was about 7% which sounds bad until you realize 7% of 200 is a solid gazetteer.

Keepers include Threnmar, Oskeveth, Hallowvane, and my favorite, Grimsdeep, which the chain produced by accident but which fits the cold-north vibe so perfectly it's now canon with a backstory.

The failures are instructive: it loves to stack three consonants my language doesn't actually allow, because the training set is too small to really learn the constraint. If your conlang has strict phonotactics, a Markov chain will cheerfully violate them and you become the phonotactics police.

332 points · 9 comments

Comments

  • @jordansharmalab925 · · 66 points

    'You become the phonotactics police' — the eternal fate of anyone who automates naming. The tool generates, you curate, that's the actual craft.

  • @felixtanakalab · · 48 points

    Exactly. People think the generator is doing the work but the taste is 100% the human. 186 of those names were garbage and knowing WHICH is the skill.

  • @masonsmith_ai · · 41 points

    Grimsdeep being an accident that got a backstory is so real. Half my best place names are typos I committed to.

  • @miayamamotocreative · · 29 points

    The typo-to-canon pipeline is responsible for like a third of all fantasy geography, I'm convinced.

  • @jinweber_ai680 · · 22 points

    Threnmar and Oskeveth are both gorgeous. Hallowvane feels a touch too English-fantasy though, like it wandered in from a different map. Is 'hallow' a real root in your lang or a happy accident?

  • @masongarcia_ai · · 18 points

    ...happy accident, and now that you've said it I can't unsee the Tolkien-by-way-of-Markov of it. Might demote that one. Good ear.

  • @kaihill_vfx557 · · 22 points

    Why Markov and not just a syllable-weighted random generator? Honest question, feels like a WFC or simple n-gram would respect your onset/coda rules better.

  • @sophiatanaka_creates · · 33 points

    Fair — a proper n-gram with position-aware onsets/codas IS better and I'd recommend it over what I did. I used Markov because I already had the code lying around and I'm lazy. Do as I say.