Now that Google Translate has begun to approach Star Trek universal translator status (I can speak into my Android phone in English and have it speak the translation in almost any Western language…), it begs the question: why are there no good Latin translators? All those I can find do a very tentative and partial translation job, putting much on the user to select the correct word order, verb tense, etc. and making them a tool for someone already partially fluent in Latin.
It would seem that Latin lends itself to more precise translation than any evolved language. Why are the translators so “incomplete”?
My first thought is that structures like the ablative absolute defy simple translation. They can be a challenge for a human, so a translator program is no match for them. As to word order, it’s often an apples vs. oranges thing between Latin and English.
My GUESS: a lot of Google Translate’s algorithm depends upon analysis of websites that have been translated elsewhere (like by a human). For instance, Coke.comCoke.de Coke.au Coke.fr could be examined and expected to have similar information and used to compare the languages in order to develop the correlation of words and concepts. However, there just isn’t much Latin on the internet compared to “real time” languages so there isn’t enough data to form a good training set to teach Google Translate.
I have to confess to almost no understanding of Latin as a language, however much “dog Latin” I’ve picked up from the translations of common phrases. (I once again regret that I attended my high school in the first year three years of a foreign language could be substituted for the formerly required 2-of-Latin and 2-of-modern language… although I have found my rough French useful from time to time.)
It doesn’t seem possible that a translator cannot be crafted to be at least as accurate as a two- or three-year student, no matter how many conditions, cases, tenses etc. there might be. I guess it’s a case where I don’t even understand what I don’t understand.
I’m guessing low demand has something to do with it. You get a lot more “bang for the buck” spending a year getting French working than spending a year getting Latin working… think of how many French speakers there are in relation to the number of Latin speakers/scholars.
The translators for Latin suck, and I advise my students not to use them bc
the translators tend to go word by word and miss the meaning
Latin is structured differently from English.
E.g. - a simple English sentence "we had come to town to make money " comes back in Latin as “Nos venit ad oppidum ad pecuniam” which means “we/us he came to(wards) town to money”
The correct phrase, in Latin, “veneramus ut pecuniam faceremus” comes back as “come in order to make good money”.
Whereas English uses auxiliary verbs like "has/ have, had " to indicate the tense, Latin puts different endings on different parts of the word. It’s the same deal with indicating the person, Latin uses a different ending, English uses a pronoun.
Also, a lot of things covered by prepositions in English are covered by putting different endings on the word in Latin.
Then there are numerous structural differences. The infinitive “to make” can be “facere” if it’s as in “I am able to make …” but not if it’s “in order to make” because then it can be "faciam / facerem "
Gotta go, a real life friend just showed up, but hopefully you get the gist
Thanks for the further grist. It still seems that adequate AI could be developed to handle Latin’s structure - I mean, if twenty generations of schoolboy could do it… - and as for demand… there are a lot of deeply-crafted apps out there with a very small body of users. Ah, well. Tempus fuggit.
You are assuming that Google Translate is particularly bad for English-Latin, but you may be overestimating how good it is for other languages. For Japanese-English (among other pairs I am familiar with), it does an absolutely atrocious job. Sometimes not even good enough to allow you to guess the original meaning.
Google Translate doesn’t use an AI that knows anything about the structure. That’s not how it works. It is based on finding matches with previous translations.
Lack of demand is not the main factor preventing the development of an English-Latin translator. Translation of natural languages is a really, really hard problem, and for the moment, schoolboys are better at it than computers.
To sum up my earlier post with an analogy, it’s like looking at a series of pictures and having someone say “what is this a picture of?” Or "where is the cat in this picture? ". Very easy for a human to do, but virtually impossible for even the most advanced computer, as all it sees are a bunch of pixels.
I actually wasn’t aware that GT handled Latin - the last time I was laboring over a phrase, it was using other translator tools. I am aware that GT is as much rules and comparison based as any actual analytical “translating.”
I still can’t see how a language could be beyond computer translation. I concede that it might be beyond simple word-substitution, rules and comparison, but a language is inherently simpler than visual imagery; if nothing else, it uses far simpler, machine-interpretable symbols of fixed meaning. Just because an algorithm to find the kitten with the blue eyes might be on the fringes of impossibility, I don’t follow how a grammatical, correctly-spelled sentence in one language cannot be translated to another, with any ambiguities of meaning, colloquialism, style etc. noted.
ETA: Especially one of the root languages of most modern Western languages - the shared culture and etymological history would seem to simplify the task.
ETA2: And I mean good, working translation, not necessarily fully automated publication-grade conversion. I’d accept two or three interpretations, alternatives, etc. to allow my judgment to come into play. While there are a few translators that do this, they produce so many variables that the result is all but meaningless. There must be a basis for greater precision.
En-Es/Es-En is the most frequent pair (and one for which there can be several English and several Spanish versions for the same text) and some of the results from Google Translate are… ehrm… not something I’d like to use in a delicate situation. You can get the gist, but things such as gender matching or choice of dialect are off more times than not.
Expecting a reliable machine translator for any relatively infrequent pair is overly optimistic.
Language translation is a very difficult problem. Good translation needs real understanding of the work. There are better and worse translations of books. But basically everybody can find the faces in a photo.
There has not been a lot of work put into finding the cat in the picture. There has however been a lot of work put into where is the human face in the picture and who is it. Computers are getting very good at it. I use Picassa the google photo organizing program. It is really good at finding faces in the pictures and is OK at guessing who it is with initial training.
I wasn’t aware of this, but it’s absolutely true: Google Translate for Latin is unusable. It doesn’t even work for some well-known Latin proverbs (I tried to have them translated from Latin to German).
I sometimes run Latin passages through Google Translate, and then have my students correct Google and analyze where and why it went wrong. It’s a great exercise. Typically, the reasons are that Google is unable to supply the necessary implied subjects and other pronouns, and it is easily misled by word order.
To translate Latin involves a lot of analysis of the forms of words and their relationships as step one: where is the verb, which words are in prepositional phrases or relative clauses, etc. As far as I can tell, Google Translate makes no attempt to analyze, simply to match words with meanings. It does a pretty amazing job all things considered.
I want to add another post saying that ‘Star Trek universal translator status’ vastly overstates what (for instance) google translate can do. OP, I’m guessing that you don’t actually speak any of those western languages that you claim your phone does, because if you did, you’d know.
The reason that translating is so difficult is that phrases don’t always exist in all languages and if they do, their meanings often don’t overlap. Translating is not a word-based thing (indeed it’s clear how to define ‘word’), but a sentence-based or text-based thing, you have to look at the whole to understand the parts.
Humans deal with that by looking at context, and using cultural knowledge and understanding and finally by taking licence while translating to approximate meanings; google deals with it by finding frequent commonalities in bilingual pairs of texts and compiling those. This works sketchily at best even for the most common pairs of similar languages.
One of the things that consistently confuse google is that it cannot handle different case forms of a word, or sometimes even plurals. This would make latin tricky but it also messes with a whole bunch of modern languages. You need to know grammar to figure that out, and google doesn’t.
I suppose that would mean that Latin to English would be tricky for a computer more so than say Dutch or French to English - but that is not because of Latin *per se *but rather about the language pair, ie the extent to which Latin is phormologically and syntactically like or unlike English and vice versa. Still, I’m not sure that there might not be other issues; the key reason (as others have pointed out) is going to be that there is simply not enough bilingual material available to google to base translations off of.
The structure and vocabulary of Latin is similar to, though not identical with, the languages that it eventually evolved to such as Spanish, Italian, and French. I would guess that a rules-based translation from Spanish to Latin would be more accurate than a rules-based translation from English to Latin, what with English being a Germanic language with more profound grammatical differences.
In terms of a corpus of writings to use for heuristics, what about Catholic Church documents? The official versions are normally in Latin and I would expect that finding Spanish or Italian translations wouldn’t be terribly difficult.
GT won’t be able to use those for a given sentence unless the specific sentence appears in them and it happens to know that the texts are parallel ones; as for Babelfish, that one is word-based and worse than your average “human being with a dictionary” (it always picks the first word in the dictionary).