Model Review

45 mins + discussion

This consists of a detailed review of the current model peformance by the HumanFirst team against the loaded data with recommendations to improve.

Generally covering for an NLU system recommendations for

  • Subdividing overly greedy intents
  • Strengthening weak intents
  • Highlighting key intents to disambiguate with example problematic utterances
  • Highlighting key intents to merge with example overlapping concepts
  • Fixing entity labelling to reduce matching failures
  • Reducing clashing entity concepts in labelling
  • Specific NLU tuning for labelling based on native eval
  • Handling of mistranscriptions and mistyping