search

bshao001 / ChatLearner

A chatbot implemented in TensorFlow based on the seq2seq model, with certain rules integrated.
Apache License 2.0
538 stars 212 forks source link

get wrong result when use trained model to infer #74

Closed QDPeng closed 5 years ago

QDPeng commented 5 years ago

the training stage is good，loss fall to 0.04，but when I use the trained mode to infer, the result is bad. I don't know where is wrong.!!!! I test BeamSearch and GreedySearch, this results is not good. Can you help me?

bshao001 commented 5 years ago

What did you mean by saying bad or wrong results? There are many reasons that impact the final outcome of a chatbot, way more than the model itself (yes, if you try Transformer model, you might get better result than using seq2seq/nmt).

Firstly, what training data did you use? The quality of the training data is at least as important as the amount. Secondly, based on the amount of training data, and a few other factors, you training may get over-fitted or under-fitted.

You can check all these or, give me more details so that I can help you. If you have wechat, you can add me, and I might be able to give you detailed assistance.

QDPeng commented 5 years ago

this is my wechat id:anay333, you can add me

QDPeng commented 5 years ago

hello, I have a question: Why reverse the source sentence?

bshao001 commented 5 years ago

With a simple google search, you should be able to find the answer. There are plenty of articles explaining this.