MIT Technology Review > Intelligent Machines > AI’s Language Problem

It’s not only about language, and it’s not off-topic (at least not totally), lots of references to AlphaGo, etc.

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

From the same source is a review of the AlphaGo movie: https://www.technologyreview.com/s/604273/finding-solace-in-defeat-by-artificial-intelligence/

It was a good read, thanks for sharing.

AlphaGo learned to play go at a very high level by taking in a large number of human-played games Why couldn't a program learn human language the same way? Humans don't learn language by learning and applying rules. The explicit grammatical rules come after the user is fluent in the language. People use grammatical rules to learn as adults but rules alone do not lead to fluent speaking. Most people do not use grammatical rules to judge whether an utterance is acceptable or not; if it "sounds" right it is acceptable. Humans learn their first language(s) by hearing people speak then trying themselves and being corrected. Sounds a lot like how AlphaGo learned. No human is perfect at any human language. There are always unfamiliar words and even unfamiliar syntactic patterns. I wonder whether an AI could learn a language by "reading" thousands of books and "listening" to years(?) of speech via TV.

Actually, humans do learn language rules on the path to fluency. A classic example has to do with irregular verbs in English. For instance, the verb, go has the past tense, went. Children first learn went, presumably by imitation, but later go through a phase where they say goed, forming the past tense the standard rule. Later still they go back to saying went.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

This example of children's learning language is somewhat supporting my thesis that an important part, maybe the most important part, of language learning is through imitation with correction. The imitation most easily comes through interaction with fluent speakers but might also come from books. Certainly many of us acquire much of our vocabulary from reading. Even as adults we sometimes misuse a word we have read but not heard spoken.

interesting that the MIT article doesn't mention the most insightful work on AI of language being done at MIT, that of Pat Winston, its former director. Follow the money.

_________________
i shrink, therefore i swarm

I agree with this. It's important to remember that communication through language is a cooperative process. We are not just corrected but queried - even with a raised eyebrow - as we speak, so we learn to correct ourselves by rephrasing, so that redundancy in language becomes important. Furthermore, most of the time utterances do not have long-term goals. We seek instant feedback (either in language or behaviour).

Go differs in all three respects. It is a contest. Efficiency is prized. Long-term goals are paramount.

In that respect, AlphaGo is not likely yet to offer much insight into language learning. But maybe language learning can offer a little insight into playing against AlphaGo.

Like, I think, most amateur games players over a certain age, I tend to play by rote and with the barest of thought. I am playing purely for distraction. I do the same when playing a chess program on my iPhone. Because chess is so tactical, I routinely miss a simple trap which I know I know. Some time ago I decided to cheat. If it was indeed a trap or combination I did indeed know - in other words, I would have expected to have seen it if I was playing carefully - I take my move back (but still play quickly). This program has a decent Elo grading algorithm and I discovered that the difference in my grade without cheating and with cheating is of the order of 400 points. That's massive, but seems to equate what I have noticed in go - that older players playing by rote drop by about 4 dan/kyu.

What I suggest, therefore, is Matta Go - an experiment in which a pro can take back any last move once he sees how AlphaGo responds. A pro is very unlikely to be playing four grades below his best, of course, but the "correction" process might make enough of a difference for the most human aspects of go (long-term strategies?) to re-assert themselves. We might then see even better go.

My understanding is that linguists who say we don't learn rules of language mean that we do not learn by learning and applying rules that we can state. A large chunk of linguistic work is trying to state the syntactic rules that native speakers follow, and how those rules develop for the individual speaker during childhood.

It's also the case that some things, like punctuation, really are learned by explicit instruction and rules that the learner can roughly state. But the capacity to utter a grammatical sentence of English is not one of them.

_________________
Occupy Babel!

what's the Matta with you people? Can't you read?! i proved 10 years ago that there's no such thing as syntax - it's all made-up by Bishop Wilberforce - well, Bishop something or other
https://ssrn.com/abstract=2205530
and if that isn't bad enough, Noam Chomsky had the cheek to pinch my idea and publish it 40 years before me..., i mean, really!
But what is even worse, is that i had no idea what Matta means and had to Google it. Yes, that kind of sums up Go very aptly, except that Go isn't about Matta-ing the other one, it's really about Matta-ing yourself, which is a very queer way of Go-ing about life, if you could get round to it, which you can't. Even Alpha male dogs can't, but they keep trying.

How about being able to Matta recursively in real life? i would love to rewind the clock and have another Go at all the Mattas i missed out on, except that it would be pointless, for like all humans, i never learn anything and keep Go-ing on making the same mistakes over and over again, such as writing this.

On the other hand, Bishop Jawn may be on to something, for wouldn't butter Go be so much more exciting, if stones could slip one in, or slide across the bored?