How good is the new version of Google Translate?

Question

Google Translate recently (a few days ago I believe) updated its translation model for Latin and the results seem quite impressive to me. However, I am by no mean fluent in Latin and I would like to have the opinion of this community on the quality of its output.

Here are some examples of random sentences from English to Latin:

I have only two sons, one of which is a soldier. → Duos tantum filios habeo, quorum unus est miles.

My favourite Latin author is Cicero. → Meus ventus Latinus auctor est Cicero.

I think the man who said yesterday that you could not speak English never heard you. → Puto hominem qui heri dixit te Latine loqui non posse te numquam audivi.

Answer 1

This feels better than before, but there are still some errors. Just going by what you wrote above:

• Ventus as "favorite", for example, is wrong. I'm guessing they saw "favorable wind" as a meaning of ventus and went with that, but there is no precise word for favorite in Latin. So something weird happened.

• Latinus auctor would imply that Cicero is a "Latin," but I think, despite coming from Arpinum, he might have taken offense at that. Cicero is a Roman author who writes in Latin. I would expect it differentiate between ethnicity and writing, such as in English, "My favorite author who wrote in Latin is Cicero."

• audivi in the third sentence is wrong. The form as is 1st person perfect active, but it should have been an infinitive. The structure is "I think (puto) the man (hominem) never heard (numquam audivisse)." Since the main verb is present and the action is in the past, you want a perfect infinitive: audivisse.

• It's not wrong per se, but a relative clause inside reported speech is usually in the subjunctive, so dixerit instead of dixit. Cf. the example from Woodcock:

Dicit se librum quem heri legerit tibi daturum.

He said that he will give you the book which he read yesterday.

Overall Google Translate isn't as awful as before. The real key will be to test some unusual Latin and how English would render it, or even throw it a mistake or two. You can see how it treats this passage from the Historia Augusta:

Latin:

Occisus est eo tempore etiam Claudius quasi a latronibus, cuius filius cum pugione quondam ad Commodum ingressus est, et multique alii senatores sine iducio interempti, feminae quoque divites.

Google Translate:

At that time also Claudius was slain, as it were by robbers, whose son had once entered Commodus with a dagger, and many other senators had been slain without a trial, even the rich women.

Loeb:

At this time Claudius also, whose son had previously come into Commodus' prescence with a dagger, was slain, ostensibly by bandits, and many other senators were put to death, and also certain women of wealth.

Someone who would read the Google translate might have very different ideas about what was going on with Commodus and Claudius' son!

Oddly, Google translate entirely glosses over idicio and treats it the same as iudicio, but doesn't alert the user that it's an error that should be fixed.

I then tried feminae quque divites instead of quoque, and the translation then read, "and all the rich women." So it seems it still has some trouble figuring out mistakes.

---

Sentences aside, it seems it still has a major problem with word-for-word translations from the modern era, as evidenced by this horrible mistranslation of the word "chocolate" (for which also see this thread.

Answer 2

I included examples in a question on the old Google Translate. Let me repeat the exercise and have the new version (although I see no indication of an update on the site itself) try the same ones:

Original	Old translation	New translation
1. Puer canem amat.	✗ The boy dog loves.	✓ The boy loves the dog.
2. Canem puer amat.	✓ The boy loves the dog.	✓ The boy loves the dog.
3. Puerum canis amat.	✓ The dog loves the child.	✓ The dog loves the boy.
4. Canis puerum amat.	✓ The dog loves the child.	✓ The dog loves the boy.
5. Puer canis amat.	✗ The boy dog loves.	✗ The boy loves the dog.
6. Canis puer amat.	✗ The dog loves the child.	✗ The dog loves the boy.
7. Canis pueri amat.	✗ The dog loves children.	✗ The dog loves the boy.
8. Pueri canis amat.	✗ The dog loves children.	✗ The boy loves the dog.
9. The boy loves the dog.	✓ Puer canem diligit.	✓ Puer canem amat.
10. The dog loves the boy.	✓ Canis puerum amat.	✓ Canis puerum amat.
11. The boy of the dog loves.	✗ Puer canem amat.	✗ Puer canem amat.
12. The dog of the boy loves.	✓ Canis pueri amat.	✓ Canis pueri amat.
13. The boy and the dog walk together because it does not rain.	✗ Puer et canis ambulare quia non pluet simul.	✓ Puer et canis simul ambulant quia non pluit.

The translator still seems confused with canis not being just the nominative. It can be the singular genitive or perhaps the plural accusative¹; either would much make sense in 5 and 6 than a "dog-boy" with both canis and puer understood as nominatives. In 8 the translator takes pueri to be a singular nominative.

Based on these examples there is improvement but the system still struggles with the very basics. The most impressive improvement is 13, where the composition starts to work. The tool gets some things right, and more so than before, but it also fails on elementary tasks. I therefore have to conclude that Google Translate is still unreliable with Latin.

You can certainly gain more insight by studying how Google Translate fares with other sentences. My only point here was to compare the current system to the old one, and the only source on the behaviour of the old one I happened to know is the old question I linked to. I would be happy to change some of the sentences to something else, but I can't do it retroactively for the old translator.

---

¹ With some third declension words the plural accusative ending can be -īs in addition to the usual -ēs. I am not sure if canis behaves this way, and I believe it does not. My point is just that reading it as a plural accusative is more plausible than reading it as a singular accusative. Google Translate seems to read canis as a singular accusative in 5 and 8. My only point with this is that even if I allow the tool to see canis as a plural accusative, the translations of 5 and 8 still don't make sense.