Each starts with finding a set of initial clusters, there are two primary methods for finding initial clusters:
- Using provided pretrained bootstrap intents
- Using semantic search to investigate topics you believe exist
Having found a first concept you want to build from it's very easy to get stuck in the weeds - Trying to guess forward what size of taxonomy you are trying to build, and then guess what rough size of granularity to create the intents from and under which groups.
You can rely on the hierarchy classifier in HumanFirst and the ability to very rapidly rearrange, merge or subdivide intents to avoid this, and continue to make progress having confidence the model is growing in a controlled way.
To help you do this we recommend a "divide an conquer" approach. Find a concept, check that we have captured the full width of the variations of that concept, then rapidly subdivide it into it's subconcepts, checking that the classifier can still clearly differentiate each of them.