Overview

The Classification service takes a user phrase or sentence and classifies it into up to three QnAs or Intents. The classification model returns a list of QnAs or Intents and the probabilities associated with them. A selector then chooses which QnAs and Intents are most relevant to return, according to an absolute threshold which is a function of the total number of QnAs and Intents, and a relative threshold, which is calculated during training. Essentially, the selector returns the QnA or Intent with the highest probability if it is greater than the absolute threshold, and then for each subsequent QnA or Intent returns it if its probability is greater than the absolute threshold and its ratio with respect to the previously selected probability is greater than the relative threshold, returning a maximum of three QnAs or Intents.