4

Google Translate is notoriously bad at Latin. I am very bad at Latin myself and when I tried to converse with the GPT chatbot in Latin, it had to frequently correct the mistakes in my questions. From the point of view of a complete beginner, like me, it seems, that it is perhaps bettar than Google Translate in the quality of texts it produces. Are there some reviews of its Latin capabilities by some authorities or can you judge whether it is significantly better than what Google Translate used to do?

Here is some short example:

Potestne GPT recto Latine scribere?

Sic, potest GPT recto Latine scribere. Sum a lingua model amplum aperta quae Latine et aliis linguis et docta sum scribere et respondere questionibus. Sed ut in omnibus disciplinis, potest contingere ut non semper accurate scribam sine erroribus, praecipue in casibus ubi petitio non est satis clara aut ubi non habeo recentissimam informationem.

or

Quomodo Sanctus Wenceslaus regebat?

Sanctus Wenceslaus (Václav in lingua Bohemica) fuit princeps Bohemiae saeculo X. Dicitur eum moderatum et iustum regem fuisse qui populum suum amavit et eis carus erat. Legenda quoque ferunt eum Christianam religionem fideliter sectatum esse et populum suum ad fidem convertisse. Historia tamen de modo exacto quo regebat non satis clara est.

Here I am not sure, among others, why ferunt is plural and legenda singular (nominative or ablative?)?

Vladimir F Героям слава
  • 1,063
  • 5
  • 16
  • 2
    Legenda is actually plural neuter here. It's not bad, very readable, but a few mistakes (and barbarisms) in the first response. It's immediately clear this is not actual Latin. – cmw Feb 02 '23 at 22:54
  • 1
    @cmw but the word GPT is looking for here is legenda, -ae, f (not classical, probably mediaeval and later), so this is plain wrong. – Sebastian Koppehel Feb 03 '23 at 08:32
  • ... or it is the ablative, and ferunt is impersonal: "they say, by way of a legend ..." this could make sense. – Sebastian Koppehel Feb 03 '23 at 08:35
  • @SebastianKoppehel Ah, I never dug too deeply into Medieval Latin, so I didn't know that was fem. sing. – cmw Feb 03 '23 at 14:07
  • @cmw Far be it from me to claim a particular familiarity with (or affinity for) mediaeval Latin, but speaking a gendered language like French (la légende) or German (die Legende) also helps ;-) – Sebastian Koppehel Feb 09 '23 at 01:28
  • I didn't even think to connect it to French or German. Très intelligent. – cmw Feb 09 '23 at 01:34

1 Answers1

6

As a test, I wrote down in English what I thought each reply said, and let Google translate it:

First question

Yes, GPT can write correct Latin. I am a large, open language model trained to write and respond to questions in Latin and other languages. However, as in all areas, it can happen that I won't always write accurately and without errors, especially in cases where the question is not sufficiently clear, or where I don't have the latest information.

Ita, GPT recte Latine scribere potest. Sum magna et aperta exemplar linguae Latinae et aliis linguis ad scribendum et respondendo eruditum. Sed, ut in omnibus locis, fieri potest ut non semper accurate et sine erroribus scribam, praesertim in casu ubi quaestio non satis clara est, aut ubi novissimas notas non habeo.

Observations:

  • Google doesn't parrot the incorrect recto from the question and writes recte.
  • Google translates “model” as exemplar, which is at least better than not translating it at all, I guess. It is difficult to say how this technical term should be translated, perhaps simulacrum?
  • But then unfortunately proceeds to misgender the word exemplar.
  • I don't even know what to say about linguae Latinae et aliis linguis ad scribendum et respondendo eruditum. I thought surely Google would at least beat the trainwreck at the beginning of the GPT answer, but not really.
  • in casu ubi is Latinglish, that's not how you say “in cases where” in Latin. GPT's in casibus has the same problem.
  • nota, -ae, f for “information” seems a poor choice, it means something like sign, letter. (GPT's informatio is also not correct.)

Second question

Saint Wenceslaus (Czech: Václav) was a ruler of Bohemia in the tenth century. He is said to have been a moderate and just king who loved his people and was dear to them. Legends also report that he faithfully pursued the Christian religion and converted his people to the faith. Regarding the precise way in which he ruled, the history is not very clear, however.

Sanctus Wenceslaus (Czech: Václav) erat princeps Bohemiae saeculo decimo. Rex moderatus et iustus fertur, qui suos amavit et illis carus fuit. Fabulae quoque referunt se fideliter Christianam religionem persecutum ac suos ad fidem convertisse. Quoad modum praecise quo rexit, historia non satis clara est.

Observations:

  • “Czech” was not translated.
  • Google uses the NcI for “is said to have been,” which is generally better, but note that the AcI with dicitur is occasionally found in classical Latin too.
  • Actually it is more of an NsI because the infinitive fuisse is missing, which I think is not correct.
  • qui amavit etc., Google and GPT both use an indicative relative clause in indirect speech, which is wrong.
  • suos amavit sounds like he loved his family, suum populum is definitely clearer.
  • illis should probaby be eis (similar to how GPT does it), because they were just mentioned.
  • Google uses se in indirect speech to refer to someone who is not the subject of the main clause, which is wrong.
  • Again with the missing infinitive, but in this case the missing esse seems more excusable to me.
  • quoad + acc. for “regarding” is extremely unusual and doubtful.
  • praecise is wrong, it says “the way in which he ruled accurately.”

In summary, I would say both are pretty rough, and I do not get the impression that GPT is significantly better than Google, but it is also certainly not significantly worse.

Sebastian Koppehel
  • 34,011
  • 2
  • 58
  • 110