Why can't we use Google Translate for every translation task?

Question

Once a book is published in a language, why can't the publishers use Google Translate AI or some similar software to immediately render the book in other languages? Likewise for Wikipedia: I'm not sure I understand why we need editors for each language. Can't the English Wikipedia be automatically translated into other languages?

Answer 1

Google has achieved significant progress in AI translation, but it's still no-where near a qualified human translator. Natural language translation is already very challenging, adding domain knowledge to the equation is too much even for Google.

I don't think we have the technology to translate an arbitrary book from one language to another reliably.

Answer 2

This is an good question and I have to wonder if someday we might. It may simply be a matter of formalizing all of the concepts conveyed by humans, which is emergent, but has to be finite. The present algorithms do not understand the content in a human sense of meaning, but are refining a statistical model to continually produce more accurate output.

You could do it today, and it actually has been done I am sure. There was a period back in the aughts of machine translated movie subtitles that were serviceable, but so bad as to be comical. (Humor that can be found in bad random translation is highly underrated;)

A key difficulty is linguistic. The same word or phrase (or symbol) can mean many things, [see: polysemy] so often the sense in which you mean it is required.

Another part of this is that words and symbols and idioms may not have a direct counterpart in another language. If the literal meaning is different from the figurative meaning, but the algorithm only knows how to translate the literal meaning, it's a bad translation.

As an example, try translating "hair of the dog" into Chinese and back to English on Google translate. The result is "dog hair" which is not the intended meaning.

The algorithm can learn that there is a figurative usage, but still has to know when to apply it as opposed to the literal. Without understanding the meaning of a phrase and the phrases leading to and from it (the context), the algorithm must rely on statistics.

So it may be a function of time, memory algorithms sufficiently complex, and a sufficient sample size. But, because language is continually evolving, there will always be instances in which the sample size is insufficient (new concepts or usage, new words.)

My sense is that the present algorithms will become very good at translating in certain contexts, especially where the meaning is literal (asking for directions, asking the price, customary social interactions, etc.) but will never reach fully human capability without the capability to grasp the meaning. Semantics vs. syntactics.

Whether algorithms will ever be able to understand in the sense of humans is still very theoretical and a subject of fierce debate, but until they do, human translators will likely be preferred.