How to solve Confusion Matrix?


I tested my Rasa NLU model using the following rasa test nlu -u data/ --config config.yml --cross-validation --confmat output --out output

It returned a confusion matrix as follows -


In this the contact_sales utterances were wrongly predicted as ask_faq_platform and enter_data. How should I resolve this conflict, do I have to manually tweak the utterances so that they donot clash or is there a better way to solve this.

My other query is that, when I tried these confusing utterances in rasa shell nlu, it predicted the right intent. Also the intents that get written in the intent_errors.json file on running the nul test are working fine for me on rasa shell nlu.

Thanks, Anurag.

in the test above, you do cross-validation which excludes some utterances from training data and tests on them. When you actually train the model for the bot, you train on the full dataset