expectedInputLanguages support in Proofreader API

[Önder Ceylan]

Hi folks,

When I provide expectedInputLanguages field with a value any other than ['en'], I get the response of "The requested language options are not supported" error on both availability and create methods of the API.

Moreover, the model correctly proofreads the text in multiple languages I provide if I don't set the expectedInputLanguages config at all with the base capabilities. 

I tested this behavior on nl, es, tr languages so far and all of them resulted with the same behavior - error when the config is set, properly proofread when the config is unset. 

I was leaning towards detecting the input language with the Language Detection API, and providing the detected language code to the Proofreader API interface. However, I can't build multillinguality in my app by using this approach, so I ended up removing the expectedInputLanguages config altogether. 

Is this the expected behavior for the Proofreader API? What am I missing when I do not set the expectedInputLanguages config? Do we have any supporting material, such as fine-tunings or safety-checking models set for English or any other languages as it's hinted on the spec? 

Best regards,
Önder

[Thomas Steiner]

Hi Önder,

You're hitting the corner case where the API officially supports English (and I assume Spanish and Japanese) and tells you so, and when you ask it if it officially supports other languages, it correctly also tells you that these are not supported, but if you omit the asking step (that is, when you don't explicitly specify the language), it happily gives you a typically even good result.

I wouldn't rely on this behavior in the long-term, as the API could reject anytime if the to-be-expected language isn't defined and/or supported. A defensive, sustainable, long-term programming approach would be to check for support, and then, in the case of non-support, fall back to other means, like a server-side call.  

Cheers,

Tom

Thomas Steiner, PhD—Developer Relations Engineer (blog.tomayac.com, toot.cafe/@tomayac)

Google Spain, S.L.U.

Torre Picasso, Pl. Pablo Ruiz Picasso, 1, Tetuán, 28020 Madrid, Spain

CIF: B63272603

Inscrita en el Registro Mercantil de Madrid, sección 8, Hoja M­-435397 Tomo 24227 Folio 25

----- BEGIN PGP SIGNATURE -----

Version: GnuPG v2.4.8 (GNU/Linux)

iFy0uwAntT0bE3xtRa5AfeCheCkthAtTh3reSabiGbl0ck

0fjumBl3DCharaCTersAttH3b0ttom.xKcd.cOm/1181.

----- END PGP SIGNATURE -----

--
You received this message because you are subscribed to the Google Groups "Chrome Built-in AI Early Preview Program Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to chrome-ai-dev-previe...@chromium.org.
To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/chrome-ai-dev-preview-discuss/8b8b5f4f-2f45-44f3-aeab-01a488f183aen%40chromium.org.

[Önder Ceylan]

Hi Tom,

Thanks for your quick response, and your guidance on this topic. I
faced the same problem when I've just used ['es'] too btw. My engineer
self would like to implement it as you suggested, but my PM self would
love to deliver more value to users by making the solution
multilingual so I'm in a dilemma. From the capability perspective, it
feels like it would be a waste of resources if the model can help in
many multilingual cases but is blocked from doing so.

Is there a plan to add official support for more languages in the future?

Best regards,
Önder

[Kenji Baheux]

Hi Onder,

We are working on adding official support for more languages. Just as a data point, are you mainly interested in Spanish? For which markets?

Beyond that, we are not blocking scenarios where the language isn't specified and the prompt and/or instructions are not in supported languages. However, this does mean a bit more responsibility on your end* (e.g. quality and safety evaluations). Just make sure to guard the calls with try catch and consider a fallback plan just in case the behavior changes in the future.

*: If that's not appealing or practical enough then Thomas' suggestion is probably for the best.

Kenji BAHEUX (my how-to)

Product Manager - Chrome

Google Japan

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/chrome-ai-dev-preview-discuss/CANyzJ%3DGwfpCanUOM_DcPDeR%3DVe2pH5h0p6TD8OYgXiuzqX%2BsRQ%40mail.gmail.com.

[Önder Ceylan]

Hi Kenji,

Thanks for your response. I was mainly looking at the languages in
regions that I'm familiar with, like Dutch, and Turkish. I'm also
interested in Chinese, Hindi and Spanish too, from the perspective of
market size and language use rate.

That's good to know, I haven't considered the evaluations. Thanks for
the guidance!

Best regards,
Önder