How KAIron Improves Chatbot Accuracy?
By Ashwin Swarup
From the Members of the DSPG team at Digite Inc. (Siddhant Bane, Shashank M, Mitesh Gupta and Ashwin Swarup)
Training open domain chatbots for specific use cases is a pain. This is due to the lack of data and the inability of conversation flow designers to think of all the ways in which a user could ask questions of the chatbot.
So recently, we integrated GPT-3 with kAIron to showcase a feature called “Add Training Examples”. The idea is to use a few training examples to generate a larger set of examples that could be potentially asked by the user.
It streamlines the process of creating training sets for your chatbot.
Consider two storylines on a chatbot: — One that talks about careers and one that tells you about services offered by the company. Here is an example:-
Services Intent: Deals with the services that kAiron provides and a sample set of expected questions from the user( 4 examples, 2 responses )
Career Intent: Deals with queries about working with our team and its responses ( 3 examples, 1 response )
Now with those list of training examples, it becomes very easy for the bot to misclassify what the user is trying to ask. As shown in this example
This problem could have been clearly avoided if we had some way of generating training examples that figure out many more ways in which user can ask the question on careers.
This is where GPT-3 ( and Pegasus ) integration comes into picture. Here we can see it all come together
With these added training examples, the classifier within RASA has a higher chance of identifying what the user needs!
If you want to try out kAIron, visit us at https://kairon.digite.com
This is an open-source project under active development by us.
Other blogs from our team can be found here