Improving translation of specific content

The GPT-3.5 model is doing good in translating between different languages, but when it comes to certain specific stuff, like song lyrics, the results become less then mediocre (due to slang for example).
How can I improve the results? Can I use fine-tuning for the task, assuming that I want to translate to multiple target languages? Can I use in-context learning by providing definitions and examples? Or should I stick to a different solution?

Hi and welcome to the Developer Forum!

My gut feeling is this will only get significantly better when the models are larger with bigger training sets, or until a continuous learning system is adopted, slang is ever changing and huge in scope, I don’t see a simple way to fine tune for all of it.

1 Like

Thank you for the reply! I want to add more context here.

In the “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” paper i’ve read that one of the method to improve reasoning of llm is to use fine-tuning. From this i conclude that fine-tuning can be used to integrate new knowledge in the model itself. The problem with the fine-tuning is that it requires you to use prompt-completion pattern. So in my case I can’t decide what should be the best option:

  • to use english lyrics example as prompt and corresponding translation to one of the target languages as completion;
  • to use poorly-translatable statement as prompt and corresponding definition in english as completion.

Here i poorly understand the difference between this variants in terms of the quality of possible results. Can i use only one target language as completion in first variant, and after that model “understands” the pattern and will work good for other languages? Would the second variant improve result at all?

As far as AI and song lyric understanding goes… you may well be the first person in the world to try it, one of the issues we face is that there are no books to reference to look for best practice, we are the ones building the knowledge.

We can apply some other domains limited understanding and say that, if you teach the model in English then it will be best at that task in that language but that some of the underlying rules learned will propagate to other languages… how that works for slang? I just don’t know, you would be the first to try this I think.

Thank you! Sounds interesting and inspiring :rocket: I’ll come back if i get any results.

1 Like