search

microsoft / DialoGPT

Large-scale pretraining for dialogue
MIT License
2.34k stars 341 forks source link

doesn't take context into consideration. gets confused easily. #73

Open alan-ai-learner opened 3 years ago

alan-ai-learner commented 3 years ago

image Pls, help to remove this unwanted behavior, or suggest some ways to do it. Thanks

chiranshu14 commented 3 years ago

image Pls, help to remove this unwanted behavior, or suggest some ways to do it. Thanks

@alan-ai-learner You can try fine tuning it with custom dataset to improve the quality of responses. But don't expect it to be perfect. Can you tell your use case? So that I can suggest better option if I know?

alan-ai-learner commented 3 years ago

I just want to make a chatbot, that is like, two friends were talking to each other, it not some domain based but a simple and usual conversation that will held between two friends, so that my aim.

chiranshu14 commented 3 years ago

If you have to use GPT, then best thing would be to fine-tune with super clean dataset.

If you always want clean responses, try rasa, it will for sure give you clean results. But you will have to build your own training dataset and response sets. It won't be as creative as GPT. But you can maybe integrate GPT inside rasa (I'm doing this). For some basic intents, use your predefined responses, and for fallback_intents get GPT response. It will get heavier though.

alan-ai-learner commented 3 years ago

@chiranshu14 thanks a lot for your response, btw my chatbot wasn't domain specific, I don't know any predefined intent as of now, I just need a chatbot that can do a casual conversation with the user and remember the context for a reasonable time.