Re:Domains for the Rest of Us
by dwrodri on Sunday September 20, 2020 @10:31PM
(#24539207)
Very curious to hear about the training set for these. If I had to guess, someone scraped GoDaddy/Google Domains/Namecheap for the "premium domains" that are being squatted on and then trained a language model on that corpus. Hopefully the OP can provide details!
[ Parent
| on HN ]
Re:Domains for the Rest of Us
by irrational on Sunday September 20, 2020 @11:48PM
(#24539602)
It looks to me like it is finding nonsense words that are pronounceable to English speakers and then checking if that nonsense word is available. I assume a linguist programmer could create an algorithm to come up with pronounceable words.
Is there a word for a nonsense word that is pronounceable? A potential word?
There's more to it than just pronounceable. e.g. I got tranclitic, mysothelium, gurnt,
1. tran- looks like reanalysis of words that begin with trans-, and we often pronounce the -s with the start of the root word. Clitic is a word in and of itself.
2. myso- is a rare Greek prefix, thelium means nipple. Unfortunately, together it would suggest nipple of dirt.
3. There's a low-scored urbandictionary entry for gurnt, but that's about it.
I think there's some knowledge of morpheme-like objects in the AI.
Markov chains, perhaps?
A jabberwocky word perhaps [1]?
Side note. Searching for a pronounceable password generator [0] is how I discovered the excellent, historical multicians.org site.
It may be based on https://www.thisworddoesnotexist.com/ ?