Hi All, I am building FAQ Bot using rasa but i am facing training issue with Response selector.Diet classifier training taking 40min with 20k FAQs and accuracy=0.99 but response selector training taking 1 day for 100 epoch with GPU and accuracy is very low 0.048.can you please guide me how can i improve this response selector part. Why its taking so much time for training. When i am using small data like 100 FAQs its working fine with same code. When i used 20k FAQ .for any question its giving only one answer.i am using RASA 2.1.3
Wow. That was huge I would say. One possible reason for poor performance is may be class imbalance. So, my question is are you maintaining same no. Of training examples to all the intents.
most of the intent having 1 question and few intents having 2 question max.but when i used same data for 50 faq that time it worked well. accuracy around 0.88.
Agree to your point but there are some similar questions in those 20K faqs. That’s why it is failing. You can expect increase in the accuracy if you provide atleast 10 training examples.
sure i will do it.