When building training data for your chatbot model, it can be a challenge to balance specific keywords across each intent you want to return against them. How do you ensure you don’t over represent one specific word in a single intent? You will typically want to see important words of your training data utterance at an equal or higher proportion for the correct intent than the incorrect one. You will also want to make sure it is equal or higher proportion than the global model average.
It’s also just as important to be aware of any possible over-use of the insignificant “noise” words (like THE, AND, IS etc) within your intents, which could cause unintentional incorrect intent predictions. You will want to see these less significant words having lower proportions of representation compared with the more significant keywords in your training data utterances.
But how do you get a quick snapshot of this information?
The answer: QBox Word Density feature.
Our new 'Word Density' feature will indicate the level of representation that each word has in your training data, both at intent level and global model level. The information is presented in an easy-to-understand way and works as a complementary feature to Explain, which uses explainable AI, and is designed to demonstrate how each word in a single utterance influences the prediction. Using these two features in conjunction, you can quickly and easily find out any problem areas and have the information needed to fix the issues.
Would you like to see the feature in action?
We’ve recorded a short video to show you how this works.
Get in touch with us to see a demo of QBox