Week 5: Error Handling and Repair

“Meet you Thursday.”

“This Thursday?”

“Next Thursday.”

Three turns, one misunderstanding, fixed before anyone noticed there was a problem. That exchange is this week’s subject. It is the protocol’s error handling, and it is running constantly.

Repair is infrastructure#

Conversation has built-in machinery for trouble with speaking, hearing and understanding. Schegloff, Jefferson and Sacks called it the organisation of repair, and found a strong skew within it: speakers fix their own trouble far more often than anyone else fixes it for them (Schegloff et al., 1977).

So the protocol gives the speaker the first chance. “Meet you Thursday, I mean next Thursday” happens before a listener has to say anything.

How often does the listener have to say something? Dingemanse and colleagues went through 48.5 hours of conversation in twelve languages from eight language families. They found 2053 cases of a listener signalling trouble: one about every 1.4 minutes, in every language sampled (Dingemanse et al., 2015).

Misunderstanding isn’t an exception condition. It’s steady-state traffic.

Three ways to say you didn’t get it#

The same study found three kinds of repair initiator, the move a listener makes to flag trouble. Every language sampled had all three (Dingemanse et al., 2015):

And the rule that decides between them: people choose the most specific initiator available. It costs both sides least. A restricted offer (“This Thursday?”) means the speaker only has to say yes, or give one word. An open request means they have to repeat everything.

When you’ve half-understood, say the part you got. “Sorry, you want the draft by Friday, or on Friday?” shows the speaker exactly which part you did catch.

Kendrick timed how soon 169 other-initiations of repair began: the ones that located the trouble specifically arrived after shorter gaps than the ones that didn’t, and the rare ones that simply corrected an error arrived without delay (Kendrick, 2015). The most frequent gap before a repair initiation was about 700 ms — so the move that flags trouble is itself one of the slow turns, which is worth knowing if you have ever felt you were too late to ask.

And part of what that delay is doing, on Kendrick’s analysis, is giving the first speaker a chance to fix it themselves. Silence before a repair is not only hesitation. Sometimes it is room.

“Huh?” is a word#

When you haven’t understood any of it, the open request is correct. It is a real word, not a failure to produce one. Dingemanse, Torreira and Enfield found “huh?” in much the same form and function across ten languages, far more alike than chance would explain. It is learnt, not an involuntary noise, and it exists because every language needs a way to say “I didn’t get that” (Dingemanse et al., 2013).

“Sorry, what?” is a repair request. The “sorry” is optional.

The repair subroutine#

The course’s diagram of the pattern. The sources describe the moves; the state machine is ours.

Accept a repair in one turn (“got it”, “ah, next Thursday”) and carry on. Adding an apology for the misunderstanding can add further turns and move the conversation’s attention from Thursday to the two of you.

In the room — the repair drill (10 min)#

One minute: choose who reads first. The sender reads a line below, mumbling the bold word; the listener looks away from the prompt and asks the most specific repair question available. The sender then supplies the word, and the listener accepts the repair. These are planted trials, not a natural conversation sample.

1. “Meet you Thursday in the library.”
2. “Send the draft to the tutor by Friday.”
3. “The room is 214.”
4. “Bring the printed worksheet.”
5. “The deadline is noon.”
6. “Use Update as the email subject.”

Six minutes: one round per line, swapping after each. Two minutes: compare two repairs: what part had been heard, what did the initiator locate, and what had to be repeated? One minute: each person records an initiator they can use again. Written rehearsal can replace mumbling: cover the word.

The drill practises specific initiators. It does not compare them against open requests or silent waits. An open request remains valid when none of the message was understood.

Counting a repair#

For these scripted examples, count from the first turn containing a repair move. Count the final acceptance separately. A self-correction inside the original speaker’s turn is one turn, even if that turn began with trouble.

Specificity changes what needs repeating even when the turn count stays the same. These counts describe the written scripts; they are not expected prices for every repair, and waiting silently does not guarantee a self-correction.

When the error is an offence#

Mishearing a day is cheap. Letting someone down isn’t, and the protocol switches to a different subroutine: the apology.

Apologies have parts. Lewicki, Polin and Lount tested six components drawn from earlier research, alone and combined. Apologies with more components were rated more effective, and some components counted for more than others. Apologies for failures of competence (“I got the time wrong”) also landed better than apologies for failures of integrity (“I lied about it”) (Lewicki et al., 2016).

We read their abstract, not the full paper, so the course won’t claim which component matters most.

Kirchhoff, Wagner and Strack found the same direction in two experimental vignette studies, and added a mechanism. More verbal elements increased forgiveness, especially after a more severe offence, and the link between an apology’s completeness and forgiveness was mediated by anger reduction (Kirchhoff et al., 2012).

For a severe offence, Study 2 found four elements particularly relevant: conveying emotions, admitting fault, a statement of apology such as “I apologize”, and an attempt at explanation.

The template, corrected#

That last element caught this course out. Earlier versions of this page told you to keep explanation out of an apology. That was too strong, and here is the revision, with what changed marked.

Two rows are new. “I’m sorry” was missing entirely, which is faintly absurd for an apology template, and Study 2 lists a statement of apology among the elements that mattered. The explanation is now in, with two constraints that are the course’s own: it goes after the ownership, never before, and it stays shorter than the apology. An explanation that arrives first, or runs long, can be heard as an excuse.

The one thing the template still refuses is “but”. Everything before it gets deleted by everything after it.

The part that’s easiest to leave out is what it cost them. Without it, an apology is about you; with it, the other person knows you saw what happened from their side. That row is the course’s, not a finding.

Kirchhoff et al. report two vignette studies in Germany. The direction is well supported; the exact four are not a universal parts list.

In the room — same parts, different order (10 min)#

A fictional case: a printer jam meant you arrived without the worksheet. Your partner shared a screen. You can print a copy at the break.

Version A: “The printer jammed. I’m sorry. I didn’t bring the worksheet. That’s my mistake. You had to share a screen. I’ll print it at the break. Would you like me to print your copy too?”

Version B: “I’m sorry. I didn’t bring the worksheet. That’s my mistake. You had to share a screen. The printer jammed. I’ll print it at the break. Would you like me to print your copy too?”

Two minutes, alone: tick the seven template parts in both versions. Three minutes, in pairs: compare the position of the explanation. Which meets the course’s ordering rule? Which do you prefer, and why? Those answers need not be identical. Three minutes: each person writes a one-turn receiver reply: acceptance, a practical request, or a request for time. Two minutes, the room: compare two replies. The checklist records what was said; it does not measure forgiveness. B follows the proposed order. That does not establish that every listener will prefer it.

In the room — write an apology (10 min)#

Four minutes, alone: use the template for a real apology you are willing to work on privately, or the fictional worksheet case. Include all seven parts. Nobody reads the whole apology out.

Four minutes, in pairs: exchange only the line naming the other person’s cost, or use a fictional line if yours would reveal too much. Does it name what the other person had to do? Two minutes: revise that line and note any part you had omitted. Keeping a real apology private does not prevent you practising its structure.

Repair is not a personality#

Fehr, Gelfand and Nag meta-analysed 175 studies and 26,006 participants on whether one person forgives another. Apology came out at r = .42, and state anger at r = -.41, which sits neatly beside Kirchhoff’s anger-reduction route from a completely different direction (Fehr et al., 2010).

Situational constructs accounted for greater variance in forgiveness than victims’ dispositions. That gives us a reason to examine the situation instead of treating forgiveness as a fixed trait.

In the same analysis, the effect of gender was non-significant (r = .01) and the effect of age was negligible (r = .06).

Receiving repair#

Both subroutines have a receiving side, and it is just as undocumented.

• Someone corrects a misunderstanding of yours: update and carry on. “Oh, next Thursday. Great.” No apology needed; nothing went wrong.
• Someone apologises to you: accept in one turn, if you can mean it. “Thanks for saying that. We’re fine.” If you can’t mean it yet, “Thanks. I need a bit of time” is a conforming reply.

If time remains#

What would count as a repair being finished when the person who sent it and the person who received it still disagree about what happened?

Optional discussion. Finish by 15:55.