two thoughts on hofstadter’s take on AI translation in the Atlantic: 1. His examples are excellent and show ways in which languages is not just mastering syntax-lexicon-morphology code but requires real understanding of subject areas under discussion (and *not* under discussion)
-
-
I had the exact same intuitions about Monte Carlo methods for Go. You can find lots of statisical correlations by brute force but plateau bc you can’t find obviously correct moves that are unusual but forced by context I was wrong
Show this thread -
If neural nets allow the AI to apply enough processing power, it can brute force its way into the context, too
Show this thread -
DH points out intended audiences the AI fails to speak to, questions it fails to settle, emphatic distinctions it ignores Well with enough processing power you’ll be able to have a representational model of a text by audience, by implied questions, by semantic contrast
Show this thread -
If Google can’t pull it off it won’t be because they didn’t try to hard-code a knowledge base into the nets, but because they aren’t feeding it enough data, or data of the wrong type
Show this thread -
For example they will have very few translations of casual conversations and personal messages - especially not spoken convos Possible that modal written text has “haplax legomena” that are nonetheless frequent in verbal exchanges
Show this thread -
Could also be effects based on being unable to classify texts by generation - might be helpful to only use contemporary translations as targets, or at least distinguish between contemporary and modernized transl as two diff targets
Show this thread
End of conversation
New conversation -
-
-
Is it really false? What's your reasoning here?
-
did you see the rest of my tweets?
-
Yeah I saw them. Are you saying that Google aren't simply brute forcing connections, or that they are looking at meaning, or both?
-
well first did you read the article I mentioned? might be easier, since the sense in which eg MC methods are brute force as opposed to coding knowledge-based heuristics will be clear if you see Hofstadter’s contrast btw AI and “understanding”
-
I am not privy to AlphaGo and other deep AI research strategies but DH’s pt is that google transl doesn’t seem to understand anything about the goal of a sentence or understand the domain of objects it's referring to
-
I think he’s probably correct that they are not trying to teach it to understand those things separately, just looking at text-to-text statistics but that isn’t necessarily a problem
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.