Skip to content

Evaluate the relevance of a NLP model

In Model Quality various views give insights about the relevance of the model:

  1. Model Stats: Monitores model perfomance in production:
    • self-evaluation of the model about its relevance in terms of recognition of intent and entities
    • number of calls and errors
    • average execution time
  2. Intent Distance: the distance between intents
  3. Model Builds: history and details about model builds
  4. Test Trends: evolution of the relevance of model tests
  5. Test Intent Errors: the list of intent errors found with model tests
  6. Test Entity Errors: the list of entity errors found with model tests

Model Tests 101

Model tests are used to detect qualifications errors.

Temporarily models are built from a random part of the whole sentence set of the model (90% for example) and then tested against the remaining sentences.

The process is repeated a number of times and the most frequent errors are pushed to an admin user.

Model tests are useful only with large models.

Test Intent errors

Click on the Intent Errors tab:

schéma Tock

Since the picture above is built from a very simple model, no real error has been detected. We can nevertheless note that in some cases the model is systematically wrong with a high probability.

Test Entity errors

These errors can be viewed via the Entity Errors tab.

schéma Tock