AI and translation

[Frabjous]

I recently checked out an old German book about tanks from the library. There was a paragraph that I wanted translated so I fed it into an AI and got a decent translation.

For fun, I decided to scan the page to a pdf and feed it to the translator.
1) The section I had previously translated was different
2) The level of english of the entire page was significantly lower than the previous translation
3) It gave the page a title
4) It gave the page a subtitle
5) It said that the page was “confidential”

What is the status of AI translation?

 

[jedishrfu]

A mixed bag. Chatgpt paid seems pretty good. But others not so good.

Also it depends on the language and developers. Nuances in a language not inderstood by a developer have greater mistranslations than when developer teams know source and target languages.

 

[Greg Bernhardt]

AI translation is very good. All the frontier models handle it well. At work we translate into more than 70 language with native reviewers using Sonnet 5 confidently and that's not even frontier. The challenge atm is still localization. Language variance can be regional and hyper local. A coursebook on Japanese is not exactly how they speak in everyday life.

 

[FactChecker]

I recently checked out an old German book about tanks from the library. There was a paragraph that I wanted translated so I fed it into an AI and got a decent translation.

You typed this in directly?

For fun, I decided to scan the page to a pdf and feed it to the translator.
1) The section I had previously translated was different

That sounds like the translation might be context-specific. I wonder if there would be a difference if only the original part was in the PDF?

2) The level of english of the entire page was significantly lower than the previous translation

Can it be that the larger amount of text caused it to do a worse job in the translation. It may be limited in the number of parameters or memory allowed for the entire task, therefore leaving less for the test part.

3) It gave the page a title
4) It gave the page a subtitle
5) It said that the page was “confidential”

That might not be surprising if the PDF contained more information.

 

[Frabjous]

You typed this in directly?

That sounds like the translation might be context-specific. I wonder if there would be a difference if only the original part was in the PDF?

Can it be that the larger amount of text caused it to do a worse job in the translation. It may be limited in the number of parameters or memory allowed for the entire task, therefore leaving less for the test part.

That might not be surprising if the PDF contained more information.

Yes, but did not include diacritical marks.

I’ll give that a try.

It was only a single page.

It was still content not included on the page.