Okay, so I thought sure, I’d try generating names using ChatGPT.
Here’s what I told it: “Generate names that are three, four, five, or six letters long; that alternate consonants and vowels; that use only the letters w, e, r, t, y, u, i, o, p, a, s, d, g, h, k, l, v, b, and n; that end in a, o, or u.”
Can anybody immediately see what I did wrong? I shouldn’t have used the word “names” because naturally ChatGPT generated lists of actual names. Of course it did! Totally my fault. I immediately said, “Same criteria, but the words should be meaningless in English.”
Too many of the resultant words looked silly, and also ChatGPT had trouble keeping all the rules in mind, so I got names like
- Nivak
- Lurty
- Hosea
And isn’t it interesting that ChatGPT couldn’t keep track of the rules? I reminded it that the names must end with a, o, or u. Then I reminded it that vowels and consonants must alternate. I actually got a handful of names that might work, maybe with tweaking. But here’s what happened when I asked for names that fit the same criteria, began with vowels, and were five letters long:
- Arova
- Utieu
- Esofi
- Irala
- Ovito
- Eniwo
- Uruwa
- Isiou
- Ovula
- Aneia
ChatGPT could not remember that words cannot end in “i” or that vowels must alternate with consonants. Isn’t that interesting? I think that’s interesting.
A few of the names I got might be okay, but honestly, it’s probably easier just to generate names myself and then fiddle with them.
Please Feel Free to Share:
I don’t think ChatGPT is able to really recognise the uncommon rules you set it AS rules, and program itself to follow them while generating text. With common formulations of rules it will have seen them in its training materials, and have learned how to respond ‘correctly’; but it doesn’t UNDERSTAND the content of what you are saying.
So questions like “how much is 2 plus 2” it will recognise and answer (though I saw an example where an early user kept telling it that it was wrong to answer 4, the answer should be 5 (trolling it), and after about 20 back-and-forths it consistently changed to answering 5).
Programming language has clear rules, so it could be taught to respect those, but human language is not so consistent. And when you make up a whole new set of rules like that, it really cannot follow them. It is just a text generator, where the next bit of text is triggered by the previous bit of text, and where it’s choice for the next word is influenced by the frequency of it occurring after the previous ones in its training data. So even when asking for words that are meaningless in English, it will have a tendency to gravitate towards existing words and names – I recognise at least three probable existing non-English words or (tribal) names in the few you’ve listed above.
I know that, Hanneke, but it’s amazingly difficult to remember that ChatGPT has zero ability to comprehend anything and is strictly a text predictor. Then you do something like this and it’s suddenly much more obvious.
Something like this might work better: https://www.vulgarlang.com/
It generates a “language” but you can define precisely the parameters of what words look like, and then ignore the paradigms and definitions it puts them in and just pick the ones that have the right feel.
So for example, I used the letters you listed to define “phoneme classes”
C = w, r, t, p, s, d, g, h, k, l, v, b, n
E = a, o, u,
V = a e i o u, jo
and “word patterns”
CE
VCE
CVCE
VCVCE
CVCVCE
You can make more complicated rules as well, to specify that certain patterns should appear more frequently than others, and so on.
Very interesting, Litotes, thank you for the pointer!
Litotes, that is SWEET! A conlang generator! Thanks so much for the pointer!
… if you use “Ovula”, I shall die with laughter.
Yeah, not that one, Heather!
Classic machine learning seems better in this case. Feed in all the (male) names and words you can find, and see what comes out. I doubt there are enough female names to determine a learning set.
Lurty.
*snort*
Yeah, Lurty is bad, but I have to agree that Ovula is worse!
Hi Rachel – I had been trying to find the website that I remembered having a pretty good, customizable generator for names specifically and I finally found it:
http://www.rinkworks.com/namegen/
Unfortunately, there’s no way to define syllables, and n, but for Ugaro names, it’s pretty easy to define the sets of vowels and consonants and make the rule to alternate them, e.g. the formula (|e|o|u|i|a)(w|r|t|y|p|s|d|g|h|k|l|v|b|n)(e|o|u|i|a)(w|r|t|y|p|s|d|g|h|k|l|v|b|n)(o|u|a) generated 3,4, and 5 syllable names:
Obipa Ubetu Agako Sowu Oloyo Ilora Inasu
Bipa Ovegu Oketo Kano Sewo Pulo Iwoho
Dilu Awiyo Yahu Ohuwu Ukero Ukaho Anoko
Awibu Awogo Unoga Tewu Oyuwo Gatu Ohobu
Keyu Ipopu Ihuvu Isasu Itonu Iraku Iliwu
Idara Isonu Okuro Dido Erulu Adogu Evila
Iwaha Oyika Ipusu Adena Bidu Piga Inula
Boru Egodu Yuvo Aniha Uporo Oharu Alibu
Ulano Ereto Pewa Rida Alodu Hula Eweho
Ihatu Ipiwa Abalu Sabu Ukuno Otupu Oyeho
Anobo Gavu Iruno Edeyo Okuwu Alowa Avigu
Urono Uwoka Ivuru Aveko Heru Awuvu Ebawo
Uruvo Asusa Ikika Akena Ehaso Isoso Udaso
Ekiyu Isoku Iposo Odeku Luwu Vada Pilu
Pubu Adeno Awovo Aketo Opeya Ogiba Iteka
Uroru Utiso Iduso Idosu Okuro Otupa Iyono
Eteyu Osaga Agewo Suva Behu Otolo Uyaku
another tool that could be fun to play with!
Wow, Lydia, thank you! Some of those are definitely usable!