[INDOLOGY] Sanskrit in ChatGPT

Tue Nov 28 15:01:23 UTC 2023

Dear Harry,

If your examples are from google translate, I can maybe say just a little
bit more about how they work than I can with something like ChatGPT. There
is a linguist by the name of Martin Benjamin who is doing testing of how
(or how well) google translate works. Whenever a new language is added, he
asks google to translate a number of phrases into that language and then
asks natives of that language (or in the case of literary languages:
scholars of that language) to comment on the quality of the translation.
The Sanskrit I looked at was mostly gobbledygook because everything had
been translated word by word, a method that could of course never render
anything even vaguely idiomatic. The expression for 'having a fun time
going shopping or to bars/restaurants' was thus rendered as  नगरस्य उपरि
बहिः. (If you're interested, a rather discursive piece about his GT
research involving 'out' phrases is here:
https://www.teachyoubackwards.com/conclusions/)

>From my understanding of language as a linguist and a teacher of language
and literature, I'd say that if AI is at some point able to generate
human-quality texts, it will also be able to translate these into any
language it has been trained on; but I don't think that it will ever be
able to do decent work on human-created language that is even the tiniest
bit literary in quality. There are just so many nuances and so much context
(that is not recognisable as relevant context for most people, and thus
would most likely not be involved in how the software is trained) that
needs to be considered to understand (and thus translate) a text; I'm happy
to give examples of this.

Ananya, I think we'll be fine:-). But as McComas rightly points out, the
faultiness of these results makes them excellent teaching tools. (At the
very least because you can dissect those mistakes as much as you want,
while humane teachers would be kinder than that when correcting student
mistakes.) Oliver, I very much agree that it would be interesting to see
where the simple grammatical errors come from; but I suspect that this has
to do with who google has working for them, and how, and that is something
they will probably (?) not divulge.

--Antonia

On Tue, 28 Nov 2023 at 03:42, Harry Spier via INDOLOGY <
indology at list.indology.info> wrote:

> Minor clarification.  The examples I gave are from Google Translate not
> ChaptGPT but clearly what you say makes sense to that also.
>  I'm wondering how does an AI application learn how to translate a
> language.  Do human beings program in a bunch of translation rules of how
> to translate language x to language y  and then these human beings refine
> the rules over time.  Or is there a kind of general artificial intelligence
> programmed into a computer that is just fed thousands of sentences and
> their translations and from that it learns how to translate language x to
> language y and with more sentences fed in, it itself refines its
> translation ability.? In other words learning language translation almost
> like a human being, by practice.
>
> Harry Spier
>
>
> On Mon, Nov 27, 2023 at 5:39 PM Antonia Ruppel <antonia.ruppel at gmail.com>
> wrote:
>
>> The use of the past active participle to render the English past active
>> is to be expected: it’s the standard/most common way to render the past
>> tense in modern/spoken Sanskrit as taught eg by Samskrta Bharati, and I
>> assume that that’s the sort of Sanskrit that ChatGPT is trained on. Not
>> applying external sandhi also is not uncommon in modern Sanskrit, at least
>> as used by those who aren’t complete masters of the language the way eg
>> Madhav is.
>>
>> Antonia
>>
>> On Mon 27 Nov 2023 at 23:29, Harry Spier via INDOLOGY <
>> indology at list.indology.info> wrote:
>>
>>> Madhav wrote:
>>>
>>>> I hear that students are already beginning to use Google-Translator to
>>>> do their Sanskrit homework.
>>>>
>>>> I just did a little experiment.  Taking a few of the english
>>> translations in Apte's  "The Student's Guide to Sanskrit Composition" and
>>> comparing what Google Translator gave as a sanskrit translation of these,
>>> and comparing to the original sanskrit quotes .  A couple of surprising
>>> things stood out.  Surprising because these are fundamental things nothing
>>> subtle.  Google translator seems to use sanskrit past active participle to
>>> translate english simple past.  It doesn't seem to apply visarga sandhi, a
>>> completely mechanical process.
>>>
>>> In these examples, the yellow highlighted sanskrit is the citation from
>>> Apte, the blue highlighted sanskrit is the google sanskrit translation of
>>> Apte's english translation given below.
>>>
>>>  Rama saw govinda
>>>
>>> rāmo govindamapaśyat
>>>
>>> rāmaḥ govindaṁ dṛṣṭavān
>>>
>>>
>>> I Salute the parents of the universe, Parvati and Paramesvara.
>>>
>>> jagataḥ pitarau vande pārvatīparameśvarau
>>>
>>> viśvasya mātāpitarau pārvatīṁ parameśvaraṁ ca namāmi
>>>
>>>
>>> He washed his hands and feet.
>>>
>>> hastau pādau cākṣālayat
>>>
>>> saḥ hastapādau prakṣālitavān।
>>>
>>>
>>> She shut her eyes
>>>
>>> sā locane nyamīlayat |
>>>
>>> sā netrāṇi nimīlitavatī
>>>
>>>
>>> So says the revered Shankara
>>>
>>> iti śrīśaṁkārācāryāḥ |
>>>
>>> tathā vadati pūjyaḥ śaṅkaraḥ।
>>>
>>>
>>> Thou art, therefore, a friend.
>>>
>>> tasmāt sakhā tvam asi
>>>
>>> tena tvaṁ mitram asi
>>>
>>> _______________________________________________
>>> INDOLOGY mailing list
>>> INDOLOGY at list.indology.info
>>> https://list.indology.info/mailman/listinfo/indology
>>>
>>
> _______________________________________________
> INDOLOGY mailing list
> INDOLOGY at list.indology.info
> https://list.indology.info/mailman/listinfo/indology
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://list.indology.info/pipermail/indology/attachments/20231128/6b37be91/attachment.htm>