3
$\begingroup$

As it is discussed here, and I saw it on other Latin language forums too, everybody complains about how Google Translate fails to translate the Latin language. From my personal experience, it is not that much bad on other languages, including romance languages.

So, what makes Google Translate fail so much to translate the Latin language? Is it about its syntax and grammar or lack of data?

$\endgroup$
4
$\begingroup$

I don't know what model Google is using for their translations, but it's highly likely that they're using one of today's SOTA deep learning models.

The latest NLP models are trained on data scraped from the web, e.g. OpenAI's GPT-2 was trained on a dataset of 8 million web pages, Google's BERT was trained on the BookCorpus (800M words) and English Wikipedia (2.500M words) pages.

Now think about the amount of latin web pages and notice that there are over 6 million english wikipedia articles but less than 135.000 in latin (see here).

As you can see, massive amounts of data are crucial for neural machine translation and I assume there is simply not enough out there for latin. Plus latin is one of the most complex and complicated languages, this makes the task not easier. Maybe Google and Co also focus less on a 'dead' language which is not spoken anymore and has it's right to exist more for educational purposes.

$\endgroup$
3
  • $\begingroup$ Is Latin actually that complex and complicated compared to other languages? In terms of grammatical complexity Latin doesn't appear high in the ranks. $\endgroup$ – conscious_process Sep 1 '20 at 10:49
  • 2
    $\begingroup$ @MightestDuck, one type of complexity that Latin has is that it's more morphologically rich than English– you can produce a lot of distinct word forms by changing the word endings. Effectively it has a larger vocabulary than English, and this makes it harder to learn the patterns. Some word forms might be very rare in your training data, like imperative verbs in Wikipedia. Many researchers have thought about workarounds, but I think the most common one is to throw even more data at it. $\endgroup$ – Jetpack Sep 1 '20 at 17:43
  • 1
    $\begingroup$ Here I briefly describe what Google is apparently using for Translate now. $\endgroup$ – nbro Sep 2 '20 at 23:11
1
$\begingroup$

Simple old Latin is different from Latin and in the language words are added to written language that are not spoken as well as reverse order of words to have forward meaning.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.