Confidence is a classifier’s estimated probability that an input belongs to a particular label. In intent recognition, the chatbot compares confidence scores across predefined intents and usually selects the highest-scoring one. section 12.6 section 12.2.1
Use confidence to make chatbot behavior safer and conversation design more data-driven:
- Set a confidence threshold; treat every intent below it as unrecognized and trigger a fallback response. section 12.2.1
- Sort utterances by their maximum predicted confidence and inspect low-confidence cases for existing intents that need more labels. section 12.6
- Search for false positives separately, because a confidently incorrect intent can produce a more damaging response than an explicit fallback. section 12.6
- Use recognized answers and grading errors to add alternative utterances to the appropriate intent. section 12.6
Choosing an intent
Train a chatbot with Hi, Hello, Hey, and Howdy as examples of Greeting. A new utterance such as Heya may receive the highest confidence for Greeting. If all intent scores remain below the threshold, return the fallback instead of pretending to understand. section 12.2.1
Confidence also supports evidence management beyond chatbots: knowledge graphs can attach probability values to facts, helping systems reason about conflicting information. section 11.8.1
