Hacker News Reader

Comment by cubefox | Hacker News Reader

Related thought: Imagine we received a lot of text in an alien language with a radio telescope, with no "Rosetta stone" to decipher it. Say, 1 TB worth of text.

Now we add to that data another 1 TB of English text, and train an LLM on the 2 TB of data. Then we ask the model (in English) to translate some text from the alien language to English.

Would it work?

View on HN

No. You always need some kind of Rosetta stone or other relationship to a known language plus some context and 'plausible guesses' to understand an unknown language. Sure if I gave you III,IIII;VII — II,II;IIII — VI,II;VIII you would be able to guess that these are elementary number signs in what amounts to a rudimentary table of additions. That much would be true whether the snippet is from a potshard of an ancient civilization or received from outer space via a radio antenna. But outside of context—and nothing would be more out of context than an extraterrestrial culture—you cannot even tell with certainty whether I stands for 'one' or 'ten' or 'twelve' or 'thousand', and here we've already reached the end of what a text per se can tell you about its meaning if the signs are not clearly pictorial (and even pictorial scripts like early Chinese or Egyptian hieroglyphs are already conventionalized to the degree that for quite a number of signs in either script we are to this day not sure what they depict).

Your idea can not work unless the data that you feed the language model with correlated items. It can't. Imagine I feed a predictor with a long list of images on the one hand and, on the other hand, a long list of randomly ordered image descriptions that may or may not match the images. Do you think you could learn a foreign language that way? You absolutely need the image of a donkey be associated with the name for that animal in the foreign language, and the algorithm is no different.

Those are good reasons, yet the language model discussed above would presumably understand Alienese just as well as it would understand English. So if an LLM understands the meaning of an expression X and of an expression Y, wouldn't it be able to tell how similar those meanings are?

> here we've already reached the end of what a text per se can tell you about its meaning if the signs are not clearly pictorial

Note that language models today seem to be quite good at understanding English, even though they are only trained on symbolic text, not on any images.

The trained model would likely be able understand both Alienese and English equally well, but it never learned to translate even one word or context. It might have an internal representation for "eating food" in both languages, but since since no links exists between the languages the embeddings will not be close.

You could try it on earth with if you train a model on two separate languages, being careful that the traning data does not contain any mixed language. But even then, modern Human languages most likely have too much cross-contamination. Would be an interesting experiment nevertheless

That's the question, would the embeddings be close?

It's not clear that they wouldn't. Would an embedding of the Alienese word for "and" be close to the embedding of the English "and"? This does seem quite possible to me.

> You could try it on earth with if you train a model on two separate languages, being careful that the traning data does not contain any mixed language. But even then, modern Human languages most likely have too much cross-contamination. Would be an interesting experiment nevertheless

I agree. Though shouldn't we be able to answer this a priori? It sounds like a mathematical question.

Your understanding of 'understanding a language' is obviously different from mine when you write that "the language model discussed above would presumably understand Alienese just as well as it would understand English" and "language models today seem to be quite good at understanding English".

Language models don't understand any natural language, they're very good at manipulating it (and us!) in terms of continuing patterns across the scale from letter (orthography) to phrases and paragraphs of seemingly utility and correctness. In that regard, yes, the aforementioned model will likely have no difficulty in reproducing novel outputs that would appear likewise useful and correct to Alienese speakers as is the case for English. However this assumption, too, should come with the disclaimer that unless someone produces a reliable test for the utility and correctness of the same LM for a variety of natural and invented languages with divergent grammars (such as including e.g. polysynthetic languages which have a very different view of what constitutes a 'word') without having to tweak any of the many finnicky parameters of these models—we can't be sure the model won't produce garbage when trained on the next 'exotic' language. So who knows, in English you use very few infixes and a lot of grammar takes places between fairly constant, fairly short words; a model with a given set of parameters that works well for such languages may not be very good at languages that has words built from many specific prefixes, infixes and suffixes that are as expressive as entire phrases in English. Just like the current generation of text-to-image generators are pretty good at a lot of things but then screw up when asked to picture a cornfield.

> Your understanding of 'understanding a language' is obviously different from mine when you write that "the language model discussed above would presumably understand Alienese just as well as it would understand English" and "language models today seem to be quite good at understanding English".

> Language models don't understand any natural language, they're very good at manipulating it (and us!) in terms of continuing patterns across the scale from letter (orthography) to phrases and paragraphs of seemingly utility and correctness.

Come on, chatting an hour with GPT-4 should remove all doubt that it understands you quite well. Otherwise, what would be understanding? Lest it turns out that we are stochastic parrots, too!

https://www.bing.com/images/create/cornfield/64b58e89d412420...

Also, the assumption that math is universal so sharing vocabulary in math is helpful for bootstrapping language understanding is a fascinating assumption to question. Even if you can explain Pi and prove that you can mutually understand trigonometry that might give you some small portion of engineering insight, but it can't yield most of the rest of engineering such as design or aesthetics (or emotions) or any number of other things that make for useful project communication.

It's something I've often thought about in the way that the Voyager record was built and Sagan's Cosmos novel assumes it and many others. Even recently, the novel Project Hail Mary borrowed that assumption that math is enough shared language to bootstrap understanding. I think the movie Arrival did some of the best work of showing why that wouldn't necessarily work, but also had the language in question designed by a mathematician and still fell into some parts of the assumption/trope. I'm not saying any of these examples are bad for doing this, I certainly love them all. It's still a small something worth criticizing.

It's certainly not a bad thing to want to communicate math, and to hope that things like Pi are "constant enough" to provide bootstraps to other communications, but it's also such a fascinating thing how much science fiction thinking (and real world scientific thinking such as the Voyage Record) think that you can just sort of "yada yada yada" your way from "so we established communications of basic mathematical constants and concepts" directly as a straight line of some sort to "now we can communicate all sorts of other things".

Reply on news.ycombinator.com