Ingest: The preference for self-correction in the organization of repair in conversation

Type: types/ingest-report.md

Classification

A theoretical-empirical journal article (Language 53(2), 1977) in conversation analysis (CA). It argues from a large corpus of transcribed natural conversation, analysed through cited fragments. It reports no counts, coding scheme or statistics; its quantitative claims ("vastly more common", "overwhelmingly") are qualitative judgments over the corpus. Author: Emanuel A. Schegloff, Gail Jefferson and Harvey Sacks (UCLA and UC Irvine), the founders of conversation analysis. This paper is one of the field's canonical texts and the standard citation for the organization of repair.

Summary

The paper replaces "correction" with "repair". Repair is not contingent on error: it occurs where nothing is hearably wrong, hearable errors often go unrepaired, and "nothing is, in principle, excludable" from the class of repairables (trouble sources). It separates repair initiation from repair outcome. Either can be done by "self" (the speaker of the trouble source) or by "other", giving six observed initiation-to-outcome routes, including failure. Self- and other-initiation have different positions, techniques and trajectories. Self-initiation occurs in the trouble-source turn (marked by cut-offs, sound stretches and "uh"), in its transition space, or in third turn. Other-initiation occupies one position, next turn, using a graded set of devices: "Huh?"/"What?", question words, partial repeats with or without a question word, and "Y'mean X?". The positions form a three-turn "repair-initiation opportunity space" ordered so that self-initiation opportunities come first. Others withhold initiation until the trouble-source turn completes, and often a little longer, which leaves an extra gap in which the speaker self-repairs. This withholding can override turn-taking's usual minimization of gaps, which the authors take as evidence that repair organization is independent of turn-taking. Self-initiated repair usually succeeds within the same turn. Other-initiation typically only locates the trouble, "even when 'other' clearly 'knows'" the correction, and so yields self-repair in the following turn. Two preferences result: for self-initiation and for self-correction. "Preference" is defined structurally, not as motivation. Other-correction is rare, modulated by uncertainty markers or "Y'mean X?" formats, placed after invitations, and treated as disagreement. Recipients whose understanding suffices to correct should instead produce the next turn the prior turn calls for. Other-correction is more common toward the not-yet-competent, as a device of socialization. The paper closes by calling repair "the self-righting mechanism" of language use in interaction and argues that an adequate theory of natural language needs an account of it.

Contribution assessment (our opinion)

The problem matters: before this paper, linguistics treated correction as error or ignored it, and self- and other-correction were studied by different disciplines. The main contribution is a reconceptualization that has proved durable. Repair becomes a sequential organization with its own positions, devices and preferences, and the observed skew toward self-repair is explained by a case-by-case mechanism (ordered opportunities plus other-initiators that only locate) rather than left as an aggregate fact. That is a strong explanatory move: it predicts specific placements (no other-initiated interruptions, post-completion gaps) that the fragments then display. The evidence is illustrative rather than measured. The fragments show that each type exists and how it is shaped, but the distributional claims are unquantified, the corpus is not described in size or composition, and several mechanisms (the opportunity space, the preference for self-initiation, other-initiator strength) are deferred to promised later papers. The socialization claim rests explicitly on passing observation. My informed guess is that later quantitative CA work and cross-linguistic studies confirmed the core organization while revising details, but this capture cannot establish that.

Quotes

Repair/correction is sometimes found where there is no hearable error, mistake, or fault: --- kb/sources/.snapshots/schegloff-jefferson-sacks-self-correction-repair.md @ sha256:d428ea3cdcc08af0dd4be01aef2c886ca3e98ae18769bc44ad22e9a594e4402f — §2.12

Furthermore, hearable error does not necessarily yield the occurrence of repair/cor-

rection: --- kb/sources/.snapshots/schegloff-jefferson-sacks-self-correction-repair.md @ sha256:d428ea3cdcc08af0dd4be01aef2c886ca3e98ae18769bc44ad22e9a594e4402f — §2.12

In view of the point about repair being initiated with no apparent error, it appears that nothing is, in principle, excludable from the class 'repairable'. --- kb/sources/.snapshots/schegloff-jefferson-sacks-self-correction-repair.md @ sha256:d428ea3cdcc08af0dd4be01aef2c886ca3e98ae18769bc44ad22e9a594e4402f — §2.13

they are ordered, alternating turn-by-turn between positions for selfand other-initiation, with positions for self-initiation preceding those for otherinitiation. --- kb/sources/.snapshots/schegloff-jefferson-sacks-self-correction-repair.md @ sha256:d428ea3cdcc08af0dd4be01aef2c886ca3e98ae18769bc44ad22e9a594e4402f — §4.2

The organization of self- and other-initiation is, then, fundamentally located in the organization of repair-initiation opportunity positions; this operates whether or not any repair is initiated, by self or other. --- kb/sources/.snapshots/schegloff-jefferson-sacks-self-correction-repair.md @ sha256:d428ea3cdcc08af0dd4be01aef2c886ca3e98ae18769bc44ad22e9a594e4402f — §4.3

The space is three turns long,22 starting from (i.e. including) the trouble-source turn. --- kb/sources/.snapshots/schegloff-jefferson-sacks-self-correction-repair.md @ sha256:d428ea3cdcc08af0dd4be01aef2c886ca3e98ae18769bc44ad22e9a594e4402f — §4.3; '22' is a footnote marker in the capture

However, the organization of repair is the self-righting mechanism for the organization of language use in social interaction. --- kb/sources/.snapshots/schegloff-jefferson-sacks-self-correction-repair.md @ sha256:d428ea3cdcc08af0dd4be01aef2c886ca3e98ae18769bc44ad22e9a594e4402f — §7

Connections Found

No KB artifact yet cites conversation analysis or this paper. For the occasion, the source is the primary anchor for treating repair as an organized feature of interaction rather than a failure mode. For the wider KB, its role is a human-interaction baseline for where detection and repair are placed between a producer and a recipient. Every comparison below is structural. The paper explains its organization by turn-taking, sequence and social constraints such as disagreement, and an agent-workflow parallel holds only where the mechanism doing the work is identified in that setting.

• On the Means for Clarification in Dialogue -- compares-with: sequential organization of repair versus corpus frequencies and form-by-reading taxonomy of clarification requests; neither separates conventional expression meaning from speaker meaning.
• Acceptance authority does not require a separate repair session -- compares-with: on who locates a defect and who repairs it. Both leave repair with the producer after another party locates the trouble. The paper adds an ordering the note lacks: the producer's own opportunities come first. It is a parallel, not evidence for the note's context-handoff cost argument.
• Silent disambiguation is the semantic analogue of tool fallback -- compares-with: on how a recipient handles trouble in another party's contribution. Conversation provides a visible next-turn channel that hands the repair back to the speaker, whereas the note's agent repairs the gap silently. The paper cuts both ways: under §6.3, a recipient whose understanding is adequate proceeds with the next turn without visible repair. That partly licenses unmarked resolution.
• Error messages that teach are a constraining technique -- compares-with: on whether the detector supplies the fix. Conversational other-initiators withhold the known correction; the note recommends that the error channel supply it. The conversational reasons for withholding (disagreement, the implicated next turn, socialization) do not obviously apply to a tool's error channel, so this marks a difference between settings, not a refutation.
• Revision guided by rationale needs faithfulness, not just legibility -- compares-with: on locating a repairable versus supplying a candidate repair. Same-turn self-repair compacts the two operations into one component; other-initiated repair splits them across turns and parties.
• The Bug That Shipped -- compares-with: on self- versus other-initiation. Coding models diagnose failures when a probe locates them (other-initiated self-repair) but rarely self-initiate. In the paper's human data, self-initiation is the dominant route.
• Re-re-reading Lucy Suchman's Plans and Situated Actions -- compares-with: on interactional repair resources. That ingest's proposed repair-asymmetry synthesis asks for primary grounding. This paper is a primary CA source for the repair mechanisms Suchman's tradition draws on, though it does not ground Suchman's own claims and says nothing about machines.

"Self-repair" here carries the CA sense (repair by the speaker of the trouble source). Several KB ingests use the term for LLM code regeneration under an oracle; the two senses are unrelated.

Extractable Value

1. Repair is organized independently of error -- The paper shows repair is not keyed to error (§2.12: repair occurs without hearable error; errors pass unrepaired; nothing is excludable from the repairables). It also shows that a repair-initiation opportunity space attends every turn whether or not anyone uses it (§4.3). Repair organization can supersede turn-taking (fn. 20), and the paper calls it the "self-righting mechanism" of language use (§7). Together these support "a structural feature of conversation with its own organization, not a failure mode" at the strength of a foundational, illustratively argued CA account. [quick-win]
2. The organization the occasion asked for -- Initiation and outcome are separate, and self/other applies to each independently (§2.2). Positions: self-initiation in same turn, transition space or third turn; other-initiation in next turn, withheld past completion (§3.1, §4.2). Preferences: self-initiation over other-initiation, and self-correction over other-correction, defined structurally (fn. 4, §4.3, §5). Other-initiators locate but do not correct (§5.22), and other-correction is modulated and heard as disagreement (§6). [quick-win]
3. The paper does not keep apart conventional meaning, speaker intention and what to do next -- Its unit is the trouble source as participants treat it, across "problems in speaking, hearing, and understanding". It classifies trouble sources by kind (word replacement, person reference, next-speaker selection), not by meaning layer. Some fragments implicitly concern what the speaker meant to say versus said (selling/buying cigarettes, ex. 43), and "Y'mean X?" checks the recipient's understanding of the turn (§6.2). The paper never theorizes a distinction between expression meaning and speaker meaning. It does tie repair to the next action: other-initiations treat trouble as an "obstacle" to producing the next turn that the prior turn calls for, and adequate understanding is shown by producing that turn (§6.3). So it supports "repair serves continuing the sequence" more than any three-way separation of meanings. A claim needing that separation requires evidence beyond this paper. The Purver–Ginzburg–Healey clarification ingest offers a taxonomy of readings, but it too does not separate conventional expression meaning from speaker meaning. [quick-win]
4. Producer-first repair as a design reference for agent workflows -- Ordered opportunities, others who locate but leave the fix to the producer, and visible initiation form a coherent human baseline. Agent workflows depart from it in both directions: weak self-initiation (The Bug That Shipped) and silent repair of others' gaps (silent disambiguation). A theory note could use the baseline to argue where detection and repair belong, but it must first establish which of the paper's mechanisms (sequential position, disagreement, competence norms) operate for agents. [deep-dive]
5. Other-correction as a socialization device -- Other-correction appears more freely toward the not-yet-competent, as a transitional usage awaiting self-correction (§6.5). This is a possible frame for operator-to-agent correction, but it rests on passing observation. [just-a-reference]

Limitations (our opinion)

The central claims are argued from selected fragments, not measured. The corpus is never characterized in size, setting or speaker population. "Vastly more common" and "overwhelmingly" are unquantified, and the fragments are chosen to show that each type exists and how it is shaped, not how often it occurs. The organization therefore rests on the analysts' judgment that the fragments are representative. All data are mid-1970s American English conversation, mostly among peers, so cross-linguistic and cross-setting reach (institutional talk, asymmetric expertise, text chat) is not established here. Several load-bearing pieces are deferred to later papers: the opportunity-space turn metric, the preference for self-initiation, other-initiator strength and the adjacency-pair structure of repair sequences. The socialization claim is explicitly based on "passing observation". "Preference" is a technical structural notion: reading it as speaker motivation, or transferring it to an agent setting as a normative rule, goes beyond what the paper claims. Finally, the paper's explanation of the skew is internal to conversation's sequential organization. Using it as evidence about agent detection and repair is analogy unless the operative mechanism is shown to exist there.

Recommended Next Action

Add this paper's organized-repair account as a human-interaction comparison in Natural-language meaning is coordination among learned approximations, alongside the clarification-request corpus study. Keep the limit explicit: the paper ties repair to the next action but does not separate conventional meaning from speaker intention.