A single utterance can carry several intents

What it is

Waking the device is expensive, so people pack several different jobs into one breath: “turn off the living-room lights, put on some jazz, and wake me at seven.” That is not several slots of one form. It is three intents that could each be their own skill. A single-label classifier usually keeps the first, or the one that looks most like the training mode, and drops the rest as noise. The user thinks they said three things; the system did one.

Why it happens

The cost of opening (wake word, being overheard, holding the floor) makes people compress a to-do list into one intonational phrase. Conjunctions (and, then, also) are weak cues — they also appear inside one intent (“alarms at seven and seven-thirty” is still set-alarm). If every training sentence has one label, extra intents become slot noise or get eaten by the majority class. Decoding often cuts at the first high-confidence intent; the rest of the acoustics is still there, the semantics already gone.

The test: if these jobs were said on different days, would they take different skills. Yes → multi-intent. No → scrambled slots of one intent. Splitting the latter as the former turns “four people, window table” into book-plus-set-preference. Merging the former as the latter stuffs lights-off and set-alarm into one frame, and one of them will drop.

Studying it

Build a multi-label intent set: one sentence may have several intent spans. Compare single-label versus multi-label models on intent recall (whether the second and third intents survive) and on how many jobs the user afterwards believes were done. A field cut is intent count on first utterances: the share of live first turns that name more than one skill. Annotate spans; do not assign only a primary label.

Labs that instruct “say one thing at a time” design the phenomenon out of existence. In Wizard-of-Oz, ask people to speak everything they need before leaving the house in the morning, and see whether they pack spontaneously.

Where it stops holding

A strictly single-skill device (lights only) has no second intent to run; people will still pack. If you cannot execute the rest, at least hear that something else was asked — do not drop it silently. Keypad menus are one key per turn. Short “lights off” is one intent. People who pay a wake-word cost, or who will not speak twice in public, pack more. Anaphora in the same sentence (“turn that off too”) is a second intent or a second slot of the same one depending on whether “that” is another skill's object.

Applying it

• Label corpora with the “different day, different skill?” test. Report recall of the second intent; do not ship on single-label accuracy alone.
• If you cannot execute several yet, at least detect and speak: “lights are off — should I set the alarm next?” Silently dropping the second half is a defect.
• Do not dodge packing by prompting “one thing per command.” If logs show packing is the norm, a prompt will not change the cost of opening.
• How to check: run first utterances that name two or more skills. Executing only the first and never mentioning the rest is a detection miss. Splitting “table for four, window” into two skills is a slot/intent confusion.

Related

• Same group: M2.13.2 Execution order of a compound command changes the outcome · M2.13.3 Partial success has to be reported item by item
• Nearby: M2.07 Dialogue flow and state design · M1.05 Topic switching and task nesting · C7.14 Command syntax versus free expression
• Search terms: multi-intent · multi-label NLU · packed utterance