Say what was not heard, not just “repeat”
Aliases: trouble-source repair · targeted re-prompt · open vs restricted repair
What it is
After a recognition miss, the system’s next sentence is not “please say that again.” It is a sentence that names the piece that did not land. “I missed the flight number” and “please repeat” both ask for speech; the first locates the trouble source, the second turns the whole utterance back into a burden. A bag-drop kiosk that already has the name and the date, and only lost the destination city, should not demand the whole sentence again. The wording is not aimed at the gap.
Why it happens
In conversation analysis, other-initiated repair can sit at different resolutions: an open “huh?” only flags that the last turn had a problem; a restricted “which city?” pins the problem on one slot. Resolution decides how much the other party must rebuild. Repeating the whole utterance sends already-good slots back through the recogniser; a new substitution can contaminate a value that was already grounded. Naming the gap lets the user redo only that piece; grounded content stays grounded.
“Please repeat” also has a reference problem: repeat what? The whole last sentence, the last word, or the stretch the system just mangled? People often repeat everything, or only the end, which does not line up with what was actually dropped. Naming the unheard slot puts the repair target on shared ground. That is a question of which sentence to say this turn. Whether the user should change one word or the whole line, and whether consecutive misses should change channel, is a different decision.
Studying it
Render the same miss as two recovery sentences: generic (“please say that again”) versus slot-named (“I missed the destination city”). Dependent measures: whether the next user utterance is whole or partial, whether already-filled slots get overwritten, whether the gap fills this turn, number of repair turns. Independent variable: whether the lost slot sat at the start, middle, or end of the original sentence (the end is more often treated as “the bit to repeat”).
On corpora, code system wording with the open versus restricted repair initiation distinction from the Schegloff line of work. In logs, after “please repeat,” which words does the user redo, and do they intersect the actual low-confidence slot. No intersection means the wording missed the trouble source. If the lab pops a transcript after the miss, the measure is reading repair, not spoken wording.
Where it stops holding
When the system does not know which piece was lost, naming a slot aims at the wrong target and people redo something that was already right. Better to admit “I missed the whole sentence” than to fake precision. Several slots low-confidence at once: naming one leaves silent errors in the rest; name every low slot, or ask for one at a time. When the user is already correcting (“not Paris — Lyon”), asking “please repeat” ignores a trouble source that has been handed over.
Applying it
- On a miss turn, read confidence first. One slot low: name that slot. The whole utterance low: say the whole utterance was missed. Do not default to “please say that again” when slot confidences exist.
- When naming a slot, anchor with what did land: “the flight number to Paris didn’t come through,” not a bare “flight number?”
- If the user is already doing a partial correction, follow that piece. Do not drag the floor back to a full repeat.
- How to check: sample “please repeat” wording, code the overlap between words the user redid and the low-confidence slot. If overlap is often empty, switch to naming the slot; re-measure overlap up and overwrite of already-filled slots down.