Restricting the scope of intents

In some cases, intent detection can be complex, especially when the model cannot be trained on the whole range of possible answers. For example, to ask for the last name of a user during a conversation: it is not possible to train an intent on every existing last name.

Intent restriction limits the intents that can be detected for the next user sentence only. Each eligible intent comes with a modifier added to its probability: a positive modifier makes the intent more likely, a negative one less likely. Intents that are not listed cannot be detected for this sentence.

In a Kotlin story (integrated mode)

Set the nextUserActionState of the bus:

nextUserActionState = NextUserActionState(
    listOf(
        NlpIntentQualifier("ask_last_name", 10.0),
        NlpIntentQualifier("cancel", 0.0),
    )
)

The eligible intents for the next sentence are ask_last_name and cancel, the latter being less likely to be detected because of its lower modifier.

In a configured story

Configured stories also support intent restriction (nextIntentsQualifiers in the story configuration, for instance in a story dump). The intents targeted by the quick replies of the story are automatically added to the eligible intents (with a modifier of 0.5), so that the restriction does not prevent the quick replies from working.

The current version of Tock Studio does not provide a screen to edit the intent restriction of a story.

See also

The mandatory states of an intent (see Intents) restrict an intent to the requests made in a given state.

Chat with Tock

×