AI and translation

  • Thread starter Thread starter Frabjous
  • Start date Start date
Join the discussion
Registration is free. Ask a follow-up in this thread, or start your own.
3 replies · 153 views
Frabjous
Gold Member
Messages
1,974
Reaction score
2,442
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?
 
Reply
  • Informative
Likes   Reactions: FactChecker
Physics news on Phys.org
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.
 
Reply
  • Like
Likes   Reactions: Frabjous
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.
 
Frabjous said:
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?
Frabjous said:
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?
Frabjous said:
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.
Frabjous said:
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.