zlacker

I'm not understanding the praise this is getting. The words I've seen are very clearly wrong and do not match how English words are made. Some examples:

> méxis: an obsessive or revelatory pursuit

No comment....

> heelbark: a red braid fastened to a man's hat so as to prevent heeling

Unless you put your hat on your shoes, you're on the wrong end of the body.

> transgate: raise the value of (something, especially money) by expanding its capacity to become transactions or funds.

What's that even mean?

>noress: a unit of electric charge equal to one nanosecond

Where's the Coulombs? Who is J̶o̶h̶n̶ ̶G̶a̶l̶t̶ Noress?

Additionally I'm seeing words that either exist or are natural permutations/mispellings. Example:

> monucleotides, but mononucleotides are a real thing.

Additionally, the example sentences are just as crazy. Maybe I'm having bad luck. There are some good hits, but the majority of them appear pretty tashy (this is a crazy difficult problem!)

>>godels+(OP)
I’ll add it’s also getting syllables wrong:

- incineratory split into in•cin•e•ra•tory which have the suffix wrong (right is tor•y)

- re•tro•genic instead of re•tro•gen•ic

>>godels+(OP)
How are English words made?

>>Cathed+jp
I'm not a linguist, but typically words evolve as memes and/or follow etymological patterns made up of root words. It's very rare that they're plausible sounding gibberish attached to plausible arbitrary meanings. This generator seems like it's in the "uncanny valley".... They're all somewhat plausible immitations of words, but the fact that they're not natural can be felt.

>>yosito+jR
Totally agree with all of that. It's probably a matter of outlook; I expect and enjoy some weird uncanny valleying from a humorous ML generator.

>>tasoga+Il
I think people here are missing what you're saying because it is subtle. Which is correct. That "troy" is different from "tor•y". "y" should be the suffix. Just like how "fix•ed" would be different from "fi•xed". "y" is the suffix like "itch" vs "itchy".

What this means, building off of what the evidence I gave, is that this model is not learning the morphemes (smallest root meaning). This exact characteristic is part of why these words sound weird. It is the same problem as the one brought up by tasogare.

[0] https://en.wikipedia.org/wiki/Morpheme

>>Cathed+jp
Typically they have a root to them. There are words that don't and are made up, like yeet (which I'll consider a word because of its usage and common knowledge), but other words like "microscope" are are derived from Latin or something else. The example here is from microscopium. There's a lineage and things modify more slowly (slang typically moves faster but also rarely stays in the lexicon long term). Many words are portmanteaus or compounds, like heelback (heel + back). How words are composed is called Morphology[0]. I mentioned in another comment morphemes. Let's look at transgate. We have trans+gate. Trans is a loan word from Latin meanings “across,” “beyond,” “through,” “changing thoroughly,” “transverse". We know what a gate is, but it can also be like a block (gated) or in a circuit (which is like a door). Here the model is taking the morpheme "trans" and using it as if it is "transaction". But in "transaction" the word makes sense because it is through an action (the word started from the meaning to do business and because this often means exchanging money, that's how we now think of it).

So "transgate" also sounds weird because it has opposing ideas. "through" + "block". But we need to look at morphemes to see why. At least (IIRC) it made this word a verb.

[0] https://en.wikipedia.org/wiki/Morphology_(linguistics)

>>godels+MS1
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