Trimmed information still needs a way back in

What it is

Speech reports only “the Lujiazui Starbucks, eight hundred metres”; hours and whether there are seats were trimmed. The user follows with “what time do you close.” If the system can only search again or say it does not know, trimming has become deletion. Retrieval path for trimmed content: fields withheld from the first turn must still be nameable in the next — “more detail,” “the address,” “opening hours” — not discarded with the short readout.

Why it happens

Trimming compresses a full record to the one or two fields this decision needs. The other fields leave the turn but remain on the object at the system side. Listeners will want them shortly; they do not want them now. With no retrieval path, the gains of a short readout (comparison still in the buffer, less barge-in) are cancelled by a missing constraint, and people ask for the whole passage again — trimming might as well not have happened. Retrieval is addressing unspoken slots on the current object, not opening a new search: the shop is still in context, and “what time do you close” should hit its hours field.

The path itself has to be short and expected. Re-bundling every trimmed field into the next monologue is delayed non-trimming. The right shape is named retrieval (one field, one sentence) or a “more detail” that opens the next layer, not a one-shot dump of the file. When a screen is present and people are looking, retrieval can be that visual row; when gaze is away, the path still has to work by mouth.

Studying it

The same restaurant record, first turn name and distance only, then elicit follow-ups on unspoken fields (closing time, seating). Conditions: object still in context, field retrievable versus context gone, search again versus field never on the object. Dependent measures: follow-up success, whether the whole passage is reread, whether people re-speak the shop name. Independent variables: lag from first turn, whether another skill was inserted in between.

In logs, align “more / address / what time” follow-ups after short readouts. Failures that cluster on “slot not on the object” rather than on ASR mean the cut left no door. A lab that prints the full record on paper in the first turn will not measure retrieval, only reading.

Where it stops holding

If the trimmed field is one the system never had (no seating data), the path should say it is absent, not pretend retrieval. Fields that must not be spoken again (full card number, someone else’s phone) should not have a voice door after the cut — change channel or refuse. If the user has switched objects (“what about the other one”), retrieval on the old object needs a check that the referent is still this shop. Treating retrieval as “so the first turn should not have trimmed” stuffs every field back into the opening; the point of the path is that the opening may be short.

Applying it

• Behind every short readout, keep the current object and a table of unspoken slots. Map “more detail / address / phone / closing time / what else” onto those slots. Do not rerun search.
• “More detail” opens one layer (two or three fields), not the whole file; further drill waits for another name.
• Before a new task evicts the object, if an unspoken slot is safety- or irreversibility-related, ask whether to hear it rather than drop it silently.
• How to check: after a short shop readout, ask “what time do you close.” The system should give the time without rereading name and distance. If it searches again, claims ignorance, or rereads the card, the path was not built.

Related

• Same group: M3.11.1 Answer length should track how certain the answer is · M3.11.3 Preamble spends the auditory budget before the payload
• Nearby: M1.04 Context Retention · M3.08 Difficulty of Reading Long Lists Aloud · M1.02 Memory Load of Screenless Interaction
• Search terms: retrieval path for trimmed content · progressive disclosure · follow-up slot