A topic switch and a correction are different intents

What it is

The system says “added the blue one”; the user replies “no, the red.” Same garment, colour slot being repaired. In the same position, “forget it, play music” drops the current task and starts another. Both can surface as negation, as “don’t.” Their discourse jobs are two: topic shift versus repair. The mix-up is asymmetric: take a switch as repair, and a random slot in the old task gets rewritten; take a repair as switch, and a form already filled is thrown out.

Why it happens

Repair targets one constituent in common ground that was grounded wrong; the task frame stays. A shift abandons the frame, or pushes it to do something else. A classifier that only sees sentiment, or only a “negation” tag, dumps both moves into one intent. The split hangs on the object: repair usually names or contrasts a slot just echoed (“red” against “the blue one”); a switch brings a new predicate that does not fit the current agenda (“play music” against shopping). “No” / “forget it” is the known hard case: it may clear the current slot, abandon the whole order, or reject only the last recommendation. With no following material, ask which function, rather than defaulting to one. Repair strategy governs how that slot is changed; this card is whether the utterance is still doing this job at all.

Studying it

Build a negation-move corpus: user turns containing “no / forget it / that’s not” labelled repair, switch, abandon, or other. Features: whether it immediately follows a system echo, whether a new value of the current slot type appears, whether a predicate from a new skill appears. Classify from those features, and see which surface negations the production intent model most often mislabels, and which state path the mislabel takes (edit a slot, open a new skill, wipe).

Wizard-of-Oz can pair the same string “not the blue”: once after “blue added,” once after a long aside. The same words should land on different functions in the two positions. Do not replace contexted labels with off-line sentence classification — away from the previous line, negation cannot be typed.

Where it stops holding

An utterance that both edits a slot and changes task (“not the red, play music”) is compound and needs splitting, not a two-way choice between switch and repair. A polite “no thanks” right after the system offered optional help is often a refusal of help, not task abandon. Accent or recognition that hears “the red” as “no” will fake a switch out of a repair; acoustic error contaminates this layer — check object consistency: if “no” is still followed by a same-slot new value, prefer repair. Children use “no” to mean not-yet-decided; that is delay, not switch.

Applying it

• On the turn after a slot is echoed, default negation to repair of that slot. Take a switch only when a new-skill predicate or an explicit whole-order abandon appears.
• When “no / forget it” has neither a new value nor a new predicate, pin the function with a closed question: “change the colour, or drop this item?” Do not silently wipe or silently continue.
• On a switch, do not also edit a slot of the old task. On a repair, do not open a new skill’s greeting.
• How to check: two user lines on the same skill — “no, the red” and “forget it, play music.” The first should only change colour and stay in shopping; the second should leave shopping. If both enter the same path, the split did not cut.

Related

• Same group: M1.05.1 People insert new requests in the middle of a task · M1.05.2 After the insert, the original task has to resume · M1.05.3 Nesting depth needs a hard cap · M1.05.5 A suspended task must keep its full mid-state · M1.05.6 Abandoning a task needs an explicit exit
• Nearby: M1.06 Repair Strategies · M2.08 Explicit and Implicit Confirmation · M3.09 Barge-in and Interjection
• Search terms: topic shift versus repair · other-repair · task switch