We conduct a preliminary evaluation of ChatGPT/GPT-4 for machine translation. [V1] arXiv
This repository shows the main findings and releases the evaluated test sets as well as the translation outputs, for the replication of the study.
We conduct a preliminary evaluation of ChatGPT/GPT-4 for machine translation. [V1] arXiv
This repository shows the main findings and releases the evaluated test sets as well as the translation outputs, for the replication of the study.
This was a very interesting read. I tried very basic translation myself with chat gpt before although I was weary of it’s reliably. This however makes me more hopeful.
I’m wondering about the randomness issue, are you only using the web interface and not the API? Couldn’t you set the temp=0 to get the output to be more reliable? (I’m assuming you most likely don’t have chat gpt 4 API access so I know you would have to wait for that.)
Translation to Chat GPT is really a loose term, because actually what you are telling it to do is create a story in language B that follows the plot of the story in language A. So naturally it is going to take some creative license with the translation, and you’ll get a more accurate response if you give it the context of what you’re translating.