intentPar

intentPar is a term used in conversational AI and natural language understanding to refer to the probability distribution over a set of user intents for a given utterance. It arises from an intent classifier that assigns scores or probabilities to predefined intents such as book_flight, weather_query, or greeting. The resulting probability vector, sometimes called the intent distribution, is used by the dialogue manager to select the most likely intent, trigger corresponding actions, or decide on fallback behavior when confidence is low.

In typical deployments, the classifier outputs either probabilities or logits that are converted into probabilities by a softmax function. The values should sum to one, representing how likely each intent is given the input. These probabilities can be calibrated to reflect real-world frequencies, and training generally relies on labeled data with a cross-entropy loss. Some systems also maintain a separate confidence score or rank multiple intents to handle ambiguity or multi-intent scenarios.

Applications of intentPar include routing to the appropriate skill or handler, deciding whether to ask a clarifying question, and supporting analytics on user interactions. Practical considerations include handling class imbalance, avoiding overconfidence on out-of-scope utterances, and applying thresholds to trigger fallback or escalation.

Variations exist, such as storing log-probabilities instead of probabilities or applying temperature scaling to adjust confidence. An example would be an utterance like “Find a flight to Paris next Friday,” which might yield intentPar values such as {book_flight: 0.72, check_status: 0.04, unknown: 0.21}. The term is closely related to concepts like confidence scores and probability vectors.