turtlesoup

You are on to something there. For the syllables I'm actually using a rule-based model from Python's "pyhphen" library: https://pypi.org/project/PyHyphen/

I am not totally happy with the results but have not had a chance to train my own

You flatter me by thinking so; it is a bot! The source code is open here: https://github.com/turtlesoupy/this-word-does-not-exist

No purpose!

There are some subtleties (e.g. hyphens, derived forms, bigrams, etc.) but the biggest problem is that most English dictionaries don't have entries for every scientific word / piece of internet slang. I ended up tokenizing Wikipedia for a blacklist and still missed a lot :(

Yup, the line for the blacklist lookup is here: https://github.com/turtlesoupy/this-word-does-not-exist/blob...

For the latter, press the "Write Your Own" button and it'll do exactly that

In general, I don't fix a random seed so in general you can get different definitions (sometimes: I cache data)

That ones a little more fuzzy; intermodulate doesn't occur very much in discourse (e.g. not in the wiki article at all) even though it would naturally be related

Got a few comments about that one and a few others! I just updated the existing word blacklist and re-deployed.

Not currently, but I should probably reject samples that do! Filed a task here: https://github.com/turtlesoupy/this-word-does-not-exist/issu...

I have it on my desk!

Yep! It's a sampling procedure; I could fix a random seed or put the results somewhere. "somewhere" ended up being the browser URL here (with a signature to prevent tampering)

That's an accurate description of how I was feeling when deploying the site and learning kubernetes for the first time.

Just AI, it's getting... creative.

That's a cool idea! I know nothing about DnD; do people usually write the campaigns down and/or compose with a tool?

It worked! Thanks for the tip

I have some starter code for urban dictionary here if you want to give it a go: https://github.com/turtlesoupy/this-word-does-not-exist/blob...

The early results were that it works, but noisier datasets are tough. The urban dictionary corpus also has a ton of racist definitions

That's really unfortunate; results are straight out of the algorithm but it's reflective of a bias in the training data set.

Oh no, it must have missed my blacklist. It's supposed to give a little "this word probably exists" when that happens. It's surprisingly hard to determine if a word already exists: https://www.thisworddoesnotexist.com/w/tissue/eyJ3IjogInRpc3...

This game is great; I've been trying to think through ways of using these NLP models in a competitive game but the mechanics aren't obvious. It would be awesome to do something like an AI rap battle

How about embiggening that to incromulentness: https://bit.ly/2zFhVaD

This project is one huge yak shave starting from that idea! I was trying to pick a company name in the AI space and thought it would be appropriate if it was generated by ML. I tried a few datasets for training, and using the Oxford English Dictionary led me here

That's not ideal and probably picking up some of the original training set from GPT-2 (this model is bolted on top of it)! How about DUOLINGOLOGY instead https://bit.ly/3fPGP8q

I'm using a blacklist to reject "real" words but it's surprisingly hard to build for rare words. I'm up to ~600K items after parsing Wikipedia tokens and it still doesn't capture everything.