Divide and Conquer

Subdividing larger intents into a variety of types the classified can understand#

Helpful core doc references:

🦸 SuperPower Objectives :#

  • Strengthen an existing large intent using the active learning loop from our unlabelled data
  • Use the similarity to check we've gone wide enough
  • SubDivide the large concept into more granular ones
  • Check we've included from the parent all the data we need
  • Check that all the data from parent is in a sub intent
  • Repeat subdividing children as much as you need
  • Strengthen any individual children short of examples to ensure have 6+ for each
  • At any point look again at your unlabelled data using any parent or child in the hierarchy
  • Show how we immediately have a strong f1 score even from the first k-fold because we built in harmony with our classifier.

Video Example#

17 minutes - yes a bit long, but really key!

Divide and conquer intents in HumanFirst