Has anyone tried automating work with ChatGPT?
I've been trying to automate simple tasks using the ChatGPT API lately—things like sorting emails or summarizing meeting notes. But prompt engineering is trickier than I expected. Especially when handling Korean, the nuance often doesn't come through well. Has anyone had a similar experience? I'm curious how you're using it. Any tips or things to watch out for would be great!
8 answers
I've had a similar experience! Email classification works surprisingly well, but meeting note summaries often miss the context. Korean seems especially tricky with distinguishing between formal and informal speech. I found that adding a few examples to the prompt helped improve things a bit.
Prompt engineering isn't easy, is it? I found that instructing the model to consistently use the '~입니다' form in Korean processing improved coherence. I've also tried automating customer inquiry responses via API, but due to nuance issues, I only generate drafts and have a human revise them.
Which model are you using? GPT-4 handles Korean much better. When I summarize meeting notes, I include timestamps, and it improves accuracy. Give it a try!
I built an automatic email classifier using the ChatGPT API with Python, but I got a lot of weird results at first. So I added conditions like 'ignore emails with short content,' and it got much better. I guess it's because there's not much Korean dataset available.
I agree! I've also struggled with cases where adding instructions like 'politely' to Korean prompts actually makes the output feel awkward. As a tip, processing in English first and then translating works well too. However, it does cost a bit more.
I use this for analyzing customer feedback. Positive/negative classification works well, but I get confused with neutral nuances. Adding 'Be careful with ambiguous expressions due to the nature of Korean' to the prompt helped a bit. Let's think about it together!
I'm still a beginner, but do you have any recommended tutorials or resources? I'd like to try email classification too, but I'm at a loss on how to structure the prompts. Thanks for sharing your experience!
I recently succeeded in automating meeting minutes summarization with ChatGPT! The key was specifying a concrete format like a '3-line summary.' It works quite well with Korean too. The only hassle is that long documents need to be split into chunks due to token limits.