Interpretation results: Confidence

Interpretation results report confidence levels for both intent and entity interpretations.

Confidence levels can range from 0.0 to 1.0, and give an estimate of the probability that the candidate interpretation is correct.

Exact matches

The highest confidence levels will be for input that is detected as an exact match to the structure and entities of one of the training samples for one of the intents.

For example, suppose you have defined an intent named CHECK_ACCOUNT_BALANCE, and that this intent has a training sample of:

"Check my [ACCOUNT_TYPE]savings[/] account balance."

Suppose as well that the entity ACCOUNT_TYPE has other values defined—“checking,” “mortgage,” “joint,” and so on. A user input of “Check my mortgage account balance” would come back as an exact match.

Exact matches for text input come back with "confidence": 1.0 and "origin": "GRAMMAR". Exact matches for recognition text coming from ASR input will also come back with "origin": "GRAMMAR" and with a confidence that is high, but somewhat less than 1.0. The reduced confidence reflects the statistical uncertainty inherent to the speech recognition process.

Statistical matches

The other type of match is a statistical match. A statistical match will always have a confidence less than 1.0, and will be indicated with "origin": "STATISTICAL". If multiple statistical intent-level matches come back for a user input, the individual confidence values will be normalized so that the confidence values for the interpretation intents add up to 1.0.

For multiple exact matches, this normalization will not take place, whether the match comes from ASR input or text input.