Training Quality Check

Run a check of all your training data and see if there are issues with it.

Training quality check is available in the Training section of the NativeChat portal under the Problems Check menu.

An overview of the training quality check results screen.

Training data set

The check goes through every (marker A on the figure above):

  • Conversation trigger expression
  • entity of type Trait expression
  • question in Question answering

And then records if the bot responds with the correct conversation, entity, or answer.

Expected vs. actual response

There could be a mismatch between the expected (marker B) and actual (marker C) outcome if there are very similar expressions added as training data to two (or more) entities.

Confidence

The report also shows the confidence (marked under D) for every test utterance. The confidence is the certainty with which the bot has matched the user expression with the bot’s training data set. The highest is 100%. Matches with confidance bellow 65% are discarded (the bot responds to the user with the default invalid-replies message).

The problematic test utterances are marked with red cross when:

  • there is mismatch between expected and actual, or
  • the confidence is too low
  • no match is found

Resolution of issues

When there is a problem with the training data set, you can adjust it so that there are no similar (or missing) example expressions, or questions.

Test understanding is available under the Understanding tab in the Test console.

Then you can use the test console to check if there are still issues.