trackerdb icon indicating copy to clipboard operation
trackerdb copied to clipboard

Descriptions in different languages

Open EvilWatermelon opened this issue 1 year ago • 2 comments

Is it possible to return the results in different languages? I want to show the results in german.

EvilWatermelon avatar Mar 20 '24 08:03 EvilWatermelon

Could be also interesting for Ghostery. The question how to do it best (without creating too much work)?

I would like to keep two aspects:

  • It is sufficient to provide an English text
  • The texts are human curated (*)

(*) We started to use AI to assist; but the lead is still the human and it should remain that IMHO. Generative AI is useful to help read through texts on the company website and come up with neutral and unbiased summaries; and it is helps with the language, especially for non-native speakers. But for multiple reasons, I'm skeptical that an AI agent can (or should) create the summaries on its own.

I could see an argument for creating releases where the English text is used as the source of truth and auto-translated. But it should be clear that it is an auto-translation. To not pollute the data set, perhaps best to keep it at a different location (only having some option to overwrite).

@EvilWatermelon Do you think auto-translations would be useful? If so, then I could see an argument to centralize it here.

philipp-classen avatar Mar 20 '24 10:03 philipp-classen

I think as long as its marked clearly it should be fine using generative AI like DeepL. One of the biggest issues seems to be languages that are very difficult like Mandarin, Russian etc.

But languages that use the alphabet (e.g. A-Z) are very good to translate. I think the usage of AI should be limited in terms of the number of languages. If you start with translating into one or two languages and test it then you could enhance it by time if it works well.

It would also be possible to work with a feedback system in order to improve the failed translations. But this requires Real-time ML and we do not know if a company like DeepL uses this learning technique.

EvilWatermelon avatar Mar 25 '24 11:03 EvilWatermelon