Use LLM to revise back translations

[johnml1135]

Use an LLM to revise a back translation to mimic the style of the source translation. The revised back translation might provide better training data for mixed source configurations.

[ddaspit]

What do you mean by "rearrange back translations"?

[Enkidu93]

I think John was referring to this slack conversation: https://sil-language-software.slack.com/archives/C8022MJ7M/p1735864021033209.

[ddaspit]

It probably makes sense to move this to silnlp, since it is a research topic.

[woodwardmw]

In case it's helpful here's the notebook I used to clean up a back translation using a very simple prompt with GPT-4o-mini: https://github.com/sil-ai/madlad-finetuning/blob/main/llm_BT_cleanup_async.ipynb

Took about 20 minutes, and I think ~$1 cost.

Example before and after texts:
https://github.com/sil-ai/madlad-finetuning/blob/main/data/eng-krrBT.txt
https://github.com/sil-ai/madlad-finetuning/blob/main/data/eng-krrBT-clean.txt