Machine translation question

[Moshen]

Quote

Then, in 2017, new research showed that machines could translate between two languages without first being trained on bilingual texts. Google Translate had always mimicked the way a human might use a dictionary, just faster and at scale. But these new machine learning methods bypassed semantics altogether. They treated languages as geometric shapes and found where the shapes overlapped. If a machine could translate any language into English without needing to understand it first, Raskin thought, could it do the same with a gelada monkey’s wobble, an elephant’s infrasound, a bee’s waggle dance?

 

The above is from the current issue of Wired magazine.  It's from an article about efforts to decipher animal "speech."  (Noises/sounds they make that appear to represent communication.)

 

Does anyone have an idea what the boldface passage means?  What does it mean to treat a language as geometric shapes?

 

(Just curious.)

[sanchuan]

In principle, any properly-trained linguist is supposed to be able to 'back-engineer' the workings of a language unassisted. Sufficient exposure to context-rich language segments can help you figure out if it's an SVO/SOV/etc language, whether the copula is expressed or implicit, if words seem to inflect (and what factors that might depend on), if any frequently-used function words seem to work like articles, prepositions or what have you, and so on. Working out a grammar like that can come to look a bit like a flowchart: each question solved, however basic, will narrow down the range of answers that are possible/plausible for other questions. Once the fundamentals of grammar are clear, working out the meaning of content words is just a matter of inference and context, given comprehensible input (of the Krashen's "i+1" type). All of this can theoretically be accomplished without access to any Rosetta Stone material, such as bilingual dictionaries, textbooks etc. It would take the life of Methuselah (not to mention easy access to all sorts of context-rich language material and/or impossibly patient informants) to do that to completion, but that's the working principle.

 

I reckon the article was about machine leaning being able to do all that, and to do it in double-quick time. 

 

The metaphor of "overlapping geometric shapes" might have been used to mean that once the machine is capable of working out a language from the inside, as it were, it can be asked to do so on another language and then to translate between the two simply by factoring out any surface difference in syntax and morphology (ie the type of answer that each language gave to those exploratory questions) to expose matching "shapes" (ie the message underneath the words) that can be overlapped and hence "translated". 

 

Mine's just a poorly explained stab at guessing, never even having read the article. But if the technology is as good as I've understood it to be, it would likely be orders of magnitude better than what's currently on offer - by human or machine.

 

That said, I for one believe (and I may be in a minority) that the form and substance of a language are inextricably linked, to the extent that you can't just strip one out of the other without killing them both. But that, as it may well turn out, could just be my all-too-human perception of language, rather than how language actually works!

[Dawei3]

On 9/18/2023 at 12:02 PM, Moshen said:

Google Translate had always mimicked the way a human might use a dictionary, just faster and at scale.

This is not how Google translates originally worked nor how it works now.  The fact that the article got something this basic wrong suggests they don't know what they are talking about.