4 ms·
> as likely as marking up the parse tree and parts of speech for all your regular prose sentences That sounds completely doable with modern NLP like GPT-4. May
by gwern 4y ago
> as likely as marking up the parse tree and parts of speech for all your regular prose sentences
That sounds completely doable with modern NLP like GPT-4. Maybe we should start doing more design experiments with how one could exploit such semantics? There's no reason to think that syntax-free writing ought to be ideal, any more than space-free and punctuation-free writing was the ideal form of writing which could not be improved...
- edflsafoiewq 4y agoYes it does, so there's no reason to mark it up. The failure of the semantic web was trying to lower human text to the level the machine could operate on it. The success of LLMs is to raise the machine to level it can operate on ordinary human text.
- gwern 4y agoYou mark it up with the LLM so you can do something with it. A LLM like GPT-4 is still just emitting text tokens, it's not also generating pixels to push to your monitor. If you want to do something involving semantics like part-of-speech tagging, you'd obviously use the LLM to parse & annotate it, and then do stuff like use CSS to style nouns differently from verbs, or whatever. Maybe you don't distribute that, but run it locally, but you still have to do that at some point!