RESEARCH DOSSIER
Conversational Coordination and Repair
Abstract. Conversational coordination and repair concern the temporally organized, jointly achieved practices through which participants allocate turns, establish sufficient mutual understanding, detect trouble, and restore the progress of dialogue. This fundamental critical narrative monograph synthesizes conversation analysis, experimental psycholinguistics, cross-linguistic corpora, developmental research, neural and respiratory timing measures, collaborative reference paradigms, and recent evaluations of artificial dialogue systems. The historical foundations of turn-taking and self-repair are linked to evidence concerning turn-end projection, overlapping comprehension and response planning, listener feedback, grounding, other-initiated repair, and the specificity of clarification. Cross-language data demonstrate recurrent repair resources with locally variable realizations, while controlled studies distinguish observable coordination from assumptions about its cognitive implementation. The critical analysis examines sampling and transcription, measurement reliability, causal inference, alternative explanations, developmental variability, multimodal affordances, and the generalizability of laboratory findings. The Artificial Era provides an additional experimental context in which grounding failures, third-position correction, and downstream recovery can be evaluated as functional interactional capacities. The synthesis identifies an integrated research program centered on consequential sequences and their resolution. This work presents a source-identified critical narrative synthesis; it does not report original participant data, a registered systematic review, or a meta-analysis.
Keywords: conversational coordination; conversation analysis; turn-taking; conversation repair; other-initiated repair; self-repair; grounding; joint action; common ground; turn-transition timing; turn-end projection; clarification requests; psycholinguistics; interpersonal synergy; cross-linguistic pragmatics; human–AI dialogue; human–artificial interaction
Conversation as a Temporally Organized Joint Activity
Conversational coordination is the continuous regulation of participants' contributions so that speech, silence, listening, bodily conduct, and responses form an intelligible sequence of social actions. Repair is the set of practices by which participants identify and resolve difficulties in speaking, hearing, recognizing a referent, or understanding an action. Their conjunction defines the present research object: how an ongoing interaction maintains progress while permitting participants to interrupt, revise, and reestablish the terms of mutual understanding. Observable units include turn-constructional units, transition-relevance places, adjacency pairs, recipient displays, trouble-source turns, repair initiators, solutions, and subsequent uptake. These units connect a conversation's temporal organization with the participants' practical accountability for understanding Sacks et al. (1974) Schegloff et al. (1977) Clark & Schaefer (1989).
The central problem is that speakers must decide what a prior action means, prepare a relevant contribution, anticipate an opportunity to speak, and monitor the possibility that the shared activity has gone astray. These operations interact at different scales. A pause of a fraction of a second can organize a response; a multi-turn clarification can reopen a reference established several minutes earlier; and an unexpected correction can change the significance of what follows. A robust account therefore integrates microtiming, sequence organization, joint reference, feedback, and conversational history. The relevant outcome is not unbroken fluency. It is the achievement of adequate understanding for a current task, including the capacity to make trouble publicly actionable Clark & Brennan (1991) Kendrick (2015) Levinson & Torreira (2015).
This monograph asks how turn allocation and repair are organized, what information makes next actions predictable, how listeners display understanding, why repair changes interactional timing, what causal and observational evidence supports competing explanations, and how these patterns generalize across languages, ages, communication media, and artificial dialogue systems. It undertakes a critical narrative synthesis of original observational analyses, psycholinguistic experiments, annotated corpora, and comparative studies. There is no new participant dataset, registered comprehensive search protocol, meta-analysis, or claim of original causal discovery. Such a distinction matters because strong ethnographic regularities, experimental manipulations, and computational predictions establish different kinds of scientific knowledge Stivers et al. (2009) Roberts et al. (2015) Fusaroli et al. (2014).
Definitions, Levels of Analysis, and Research Boundaries
Turn-taking describes practices through which conversational participation is distributed over time, including who currently has a turn and who may speak next. Grounding describes the collaborative establishment of evidence that an utterance has been understood sufficiently for present purposes. Alignment describes the emergence of systematic coordination across behavior or representations, whereas repair describes a response to a recognizable problem in producing or interpreting interaction. An acknowledgment may contribute to grounding without indicating that every proposition has been deeply evaluated. A repetition may repair audibility while leaving disagreement intact. A correction may preserve cooperation even as it rejects an earlier formulation. These distinctions prevent the measurement of a social process through an overly convenient but ambiguous surface feature Clark & Brennan (1991) Schegloff et al. (1977) Fusaroli et al. (2014).
Interactional repair is analytically distinct from the self-monitoring processes responsible for detecting a speech-production error before its consequences reach a conversational partner. Both can occur together, but the defining evidence for the present dossier lies in an interactional sequence that permits the other party to observe, request, or ratify a correction. Similarly, shared reference is studied here insofar as it is established, challenged, and reestablished within dialogue. General theories of lexical access, conceptual pacts, and common ground belong to neighboring research domains. This dossier examines their coordinative consequences for the success or failure of an ongoing exchange Clark & Wilkes-Gibbs (1986) Brennan & Clark (1996) Clark & Krych (2004).
The research operates at four explanatory levels. Conversation analysis documents participant-organized sequential practices in naturally occurring situations. Experimental psycholinguistics tests cognitive processes that permit anticipation, planning, and monitoring. Quantitative corpus research estimates temporal and distributional patterns under explicit annotation rules. Comparative and computational work tests how general organizational principles vary with language, medium, task, participant capacities, and agent design. Concordance across these methods strengthens an explanation, but no one level can substitute for the others: an EEG signature does not identify an entire conversational norm, and a recurring sequential practice alone cannot reveal the neural implementation that makes it possible Bögels et al. (2015) Heldner & Edlund (2010) Roberts et al. (2015).
The Historical Establishment of Conversation Analysis
The systematic study of ordinary interaction gained a decisive formulation in the work of Harvey Sacks, Emanuel Schegloff, and Gail Jefferson. Their 1974 article characterized turn-taking through a locally managed system of turn construction and allocation, oriented toward possible completion points and the current configuration of speakers. The achievement was methodological as well as theoretical. Researchers could describe how participants recognize an opportunity for a next contribution and thereby test proposals against naturally occurring sequences. The system also clarified why overlap and silence are relevant events whose interactional significance depends on where and how they arise. Subsequent work has examined the temporal detail and variation that the original analytic program invited Sacks et al. (1974) Heldner & Edlund (2010).
The 1977 analysis by Schegloff, Jefferson, and Sacks identified repair as a recurrent conversational organization and explained the preference for self-correction: participants have resources for allowing the speaker of a trouble source to resolve a problem before an interlocutor supplies a correction. Preference here refers to the structural distribution and sequencing of interactional alternatives, not merely a stated individual liking. Repair can be initiated by either speaker and accomplished through repetition, replacement, specification, or confirmation. The historical significance of this study resides in the finding that orderly conversation can absorb difficulties without a separate external supervisor, using opportunities created inside the exchange itself Schegloff et al. (1977) Kendrick (2015).
Experimental approaches extended this program by studying how speakers construct contributions with addressees rather than for passive recipients. Clark and Wilkes-Gibbs used collaborative reference tasks to show how descriptions can be progressively negotiated, and Clark and Schaefer modeled utterances as collective contributions to shared discourse. Clark and Brennan subsequently articulated a grounding framework in which participants seek enough evidence of understanding to proceed. The historical sequence marks a broadening of the explanandum from the allocation of speaking turns to the coordinated acceptance, correction, and reuse of their contents. Each contribution retained its own methodological assumptions, which remain important when assessing present-day evidence Clark & Wilkes-Gibbs (1986) Clark & Schaefer (1989) Clark & Brennan (1991).
The Turn as an Interactional Unit
A speaking turn is not interchangeable with a sentence, an acoustic burst, or an uninterrupted interval of phonation. A participant may construct a single action with several units, while a short expression can perform a complete response. The Sacks–Schegloff–Jefferson framework analyzes recognizable possible completions of turn-constructional units and rules governing self-selection or selection by a current speaker. A turn transition becomes intelligible when participants interpret an ongoing contribution as potentially complete, as inviting a particular answer, or as yielding the floor. This account predicts local flexibility rather than a fixed alternation in which every participant speaks for an equal duration Sacks et al. (1974).
Turn boundaries also participate in a wider sequence. An answer becomes conditionally relevant after a question; an acceptance or refusal acquires significance after a proposal; a request for clarification may suspend the anticipated next action while its object is made recoverable. Sequence organization thus conditions the interpretation of response speed. A delay after a difficult question may reflect planning, reluctance, or the organization of a repair opportunity. The same duration can be produced by distinct mechanisms. Analysts need to establish which action participants understand themselves to be performing before assigning a psychological function to silence or overlap Roberts et al. (2015) Kendrick (2015).
Turn-taking is therefore a form of joint regulation of access to shared communicative space. Speakers provide prospective completion information through syntax, action design, pragmatic relevance, and temporal delivery, while recipients project a next action and position themselves to contribute. This mutual orientation does not presuppose access to one another's internal states. It is manifest in sequentially organized conduct and can be tested through controlled changes in information, timing, and response demands. Treating the turn as a social as well as psycholinguistic unit provides the conceptual bridge to repair, which interrupts ordinary progression in order to restore the conditions for a relevant next move de Ruiter et al. (2006) Levinson & Torreira (2015).
Grounding and the Principle of Sufficient Understanding
Grounding is the cumulative process by which partners make a contribution sufficiently mutually understood for the purposes of their current joint activity. Clark and Brennan formulated this as an achievement whose demands depend on the task and the communicative medium. Ordinary acknowledgment, appropriately timed continuation, a relevant answer, and explicit confirmation can each provide different kinds of evidence that participants are able to proceed. The adequacy threshold is practical: a construction crew coordinating a dangerous movement requires a different assurance than friends identifying a familiar song. The same communicative event can therefore be adequate for one activity and require further clarification for another Clark & Brennan (1991).
The mechanisms of grounding are distributed across contributions. A first speaker proposes information or an action; the recipient shows what was understood; and the first speaker may accept that display or initiate correction. Clark and Schaefer describe such collaborative units as contributions rather than solitary speaker products. A failure to display evidence can motivate further explanation even without explicit words expressing confusion. A response may also disclose an interpretation that is coherent but mistaken, creating a new trouble source. An account of successful grounding must therefore follow what happens next, rather than take an isolated response as direct proof of shared knowledge Clark & Schaefer (1989) Clark & Krych (2004).
Joint effort is an organizing consideration in grounding. Participants can invest additional work now to reduce later ambiguity, and they may rely on established reference or visible objects when conditions permit. In a series of referential experiments, Brennan and Clark found that speakers and addressees formed partner-specific conceptual agreements and continued using those terms even when shorter expressions were available. Such historical dependence reveals how accumulated interaction changes future coordination. Yet repeating the same label cannot independently establish understanding: the test is whether a partner can identify, act on, or appropriately revise the intended reference under controlled demands Brennan & Clark (1996) Clark & Wilkes-Gibbs (1986).
The Timing Problem in Rapid Turn Exchanges
Ordinary conversational turns are exchanged far faster than the time often required to prepare a novel spoken response from an initial intention. This mismatch creates a processing problem: a future speaker must make sense of the incoming utterance while preparing an answer, anticipating its completion, and coordinating the onset of articulation. Levinson and Torreira reviewed timing distributions and processing evidence, proposing substantial overlap between comprehension and production. Their account identifies a plausible organizational solution but also a research challenge: the duration of an interturn gap cannot reveal exactly when a listener began planning, nor which information was available at that moment Levinson & Torreira (2015).
A cross-language corpus study by Stivers and colleagues found that speakers in ten languages generally avoided substantial overlap and minimized silence at transitions, with quantitative differences in the overall tempo of exchanges. These results support a shared interactional infrastructure while allowing linguistic and cultural variation in specific realizations. The dataset examined natural question-answer sequences rather than every possible social action or conversational setting. The most defensible inference is that timing patterns recur across the sampled languages under the study's coding system; different local conversational conventions and activity types require additional observation Stivers et al. (2009).
Naturalistic timing is not a perfectly synchronized relay. Heldner and Edlund analyzed pauses, gaps, and overlaps across three corpora and challenged the assumption that all turn changes display unusually precise boundaries. Their contribution directs attention to the distributional tail, recording thresholds, and the distinction between within-speaker pauses and transitions between speakers. A scientifically adequate measure specifies how silence is segmented, what counts as talk, and whether speech overlapping with another speaker changes floor ownership. Otherwise similar-looking histograms can reflect different annotation conventions rather than distinct interpersonal mechanisms Heldner & Edlund (2010) Roberts et al. (2015).
Projection of Turn Completion and Response Preparation
Projection concerns how listeners use information in an unfolding utterance to anticipate its likely completion. In a controlled experiment, de Ruiter, Mitterer, and Enfield contrasted original utterances, flattened intonation, and speech retaining intonational contours but removing intelligible words. Participants' ability to predict turn endings depended strongly on lexicosyntactic content in that task, challenging accounts that would assign primacy to intonation alone. This establishes an experimentally supported role for linguistic structure in endpoint projection; it does not imply that prosody, bodily cues, and social expectations are irrelevant in other conversational circumstances de Ruiter et al. (2006).
Magyari and de Ruiter later examined whether participants predict a turn's ending partly by anticipating the words that will complete it. Their gating paradigm linked word anticipation with estimates of when an utterance would stop. The important distinction is between estimating duration from acoustic patterns and incrementally constructing a prospective linguistic continuation. If a listener forecasts how a question or sentence will finish, the final boundary becomes easier to anticipate. However, laboratory gating measures explicit predictions under interrupted listening; natural interlocutors have additional obligations, contextual information, and opportunities to prepare actions. Ecological transport must be established, not assumed Magyari & de Ruiter (2012).
The interactional perspective adds another layer to projection. People anticipate a relevant next action because the current turn performs a recognizable social action: requesting, offering, informing, or seeking confirmation. Response planning can begin before full semantic details are available, but premature planning may require revision when an unexpected final clause changes the action. Investigators should separate early recognition of an action, early lexical selection, and the physical execution of a reply. Different experimental tasks identify different components, and none can by itself supply a complete explanation of timing in spontaneous multi-party conversation Levinson & Torreira (2015) Roberts et al. (2015).
Neural and Physiological Evidence for Overlapping Processing
Bögels, Magyari, and Levinson used neurophysiological measurement in an experimental question-answer paradigm to investigate when response preparation emerges. Their findings indicated response-planning signatures during an incoming question, supporting the possibility that listeners initiate preparation before the other speaker has finished. The experimental achievement is the temporal localization of neural activity under a specified task and contrast. It does not demonstrate a single mandatory neural routine for every conversation, and scalp recordings do not identify an entire interactional process independently of behavioral interpretation. The evidence is most informative when considered alongside tasks that manipulate when response-relevant information becomes available Bögels et al. (2015).
Respiration offers a complementary window into production readiness. Torreira, Bögels, and Levinson studied breathing patterns in the course of answering questions, relating respiratory preparation to the timing of upcoming speech. Respiratory changes are observable physiological preparations for an articulatory act, but they also depend on breath cycles and other individual and contextual conditions. They cannot distinguish all possible stages of conceptual or lexical planning. The convergence between neural, breathing, and behavioral evidence supports overlapping preparation more strongly than would any single measurement, while preserving uncertainty about how the components are linked in specific tasks Torreira et al. (2015) Bögels et al. (2015).
A useful integrated model treats incoming speech recognition, action interpretation, content planning, linguistic encoding, articulatory preparation, and turn initiation as processes that may overlap without becoming identical. Such an account can predict slower transitions when the relevant information arrives late, when an answer is complex, or when the upcoming speaker must initiate repair rather than respond directly. To test the model, researchers should manipulate response information separately from utterance predictability, record both internal indices and external performance, and compare the resulting temporal patterns against alternatives. Otherwise a fast response can be mistaken for proof of early comprehension when it may reflect a stereotyped or preconstructed answer Levinson & Torreira (2015) Roberts et al. (2015).
Sequential Organization Beyond Mean Gap Duration
Corpus analyses reveal that timing depends on what a turn does, not solely on how long it is or how difficult its words are. Roberts, Torreira, and Levinson modeled turn transitions in annotated telephone conversation using features of linguistic processing, speaker characteristics, and conversational sequence. Their results motivated a joint explanation integrating cognitive and interactional organization. This matters because a response to a polar question and a response to an assessment differ in what counts as relevant uptake. Pooling them into a single average risks confusing a change in action composition with a change in fundamental response capacity Roberts et al. (2015).
Analysts distinguish between a gap after the prior speaker's voice stops and an overlap in which another speaker starts before that stopping point. Yet these acoustic measures do not settle whether a backchannel, collaborative completion, or interruption has displaced the current speaker. Nor does every longer gap imply confusion: waiting can make an alternative action relevant, permit further self-repair, or manage a sensitive proposal. Turn timing must therefore be related to sequential location and to the embodied behavior of the participants. This motivates combining acoustic segmentation with human coding of turn actions and subsequent evidence of how participants themselves understood the transition Sacks et al. (1974) Heldner & Edlund (2010) Kendrick (2015).
An experimentally useful distinction is whether an intervention alters the information required to produce an answer, the opportunity for taking the floor, or the perceived obligation to respond. The same lengthening of gaps can follow any of these changes. Factorial designs could hold question content constant while manipulating whether an answer is invited, whether an initial misunderstanding has been signaled, and whether the partner remains visibly engaged. Such designs extend descriptive corpus analysis into causal inquiry while maintaining an anchor in naturally occurring conversation. They also provide a basis for assessing the validity of machine-mediated timing systems in actual social exchanges Roberts et al. (2015) Clark & Krych (2004).
Multimodal Participation and Listener Feedback
Conversation is organized through vocal and visible conduct. Eye contact, head movement, facial expression, posture, and manual action can signal attention or comprehension, provide an opportunity for clarification, and contribute to the timing of speaking turns. A listener who nods while a narrator speaks has made an observable contribution even without claiming the next full turn. The meaning of such conduct depends on its timing and sequential location. A nod attached to a recognizable telling point can support continuation, whereas a delayed look or puzzled expression can invite specification. The psychological object is coordination across channels, not a count of gestures in isolation Bavelas et al. (2000) Kendrick (2015).
Bavelas, Coates, and Johnson tested collaborative narration in two experiments involving 63 unacquainted dyads. Listeners' specific responses, which were linked to the unfolding story, differed from generic signs of attention; distractions reduced appropriate listener responses and degraded aspects of the narrators' storytelling. The study supplies experimental evidence that listener participation can causally affect narrative performance in the studied setting. It does not establish a universal effect size for every narrative genre or communication culture. Its distinctive insight is that listeners can participate in making an utterance successful before they supply a substantial verbal reply Bavelas et al. (2000).
Clark and Krych experimentally varied the availability of visual feedback as partners worked through construction instructions. Observing the addressee enabled speakers to monitor understanding and adjust instructions while the activity unfolded. A purely text-based or delayed exchange changes the cues available for that process, the time needed to detect trouble, and the costs of asking for clarification. Consequently, a model learned from face-to-face conversation must be tested against telephone, video, and text contexts with specific operational definitions. Visible signs of difficulty are powerful resources for repair, but their interpretation remains situated within the joint task and the participants' shared history Clark & Krych (2004) Clark & Brennan (1991).
Repair as an Organized Response to Communicative Trouble
A trouble source is an aspect of prior talk or conduct that a participant treats as problematic for hearing, reference, understanding, formulation, or conversational action. Repair involves practices oriented toward dealing with that trouble. It may be initiated by the speaker whose contribution is at issue or by an interlocutor; its solution may likewise be supplied by either party. This yields four broad initiation-and-completion configurations, but their distributions and social meanings are not interchangeable. Schegloff, Jefferson, and Sacks showed that an interactional preference for self-repair structures opportunities for who should address the problem. A sequence can preserve the first speaker's responsibility to clarify what they intended while retaining the recipient's authority to signal what was not understood Schegloff et al. (1977).
Repair differs from a general expression of disagreement or the correction of a false factual proposition. If a listener says that a place name was not heard, an intelligibility problem has been raised; if the listener challenges the truth of the travel story, a distinct conversational action may be underway. Participants can combine these activities, and local practices are needed to identify which one is being treated as relevant. Similarly, an explicit repetition is not automatically successful repair if the partner continues to display uncertainty. The repair sequence must be studied through its initiation, the proposed solution, and subsequent uptake that demonstrates whether progress can resume Kendrick (2015) Clark & Schaefer (1989).
The central explanatory claim is that conversational order includes resources for managing its own disruption. A request such as 'who?' shifts the immediately relevant action away from the ordinary response to the prior contribution toward recovery of a reference. When the original speaker supplies a name, the participants may return to the suspended activity. This detour is structurally productive: the temporary interruption protects the comprehensibility of later actions. The same principle supplies an analytic foundation for studying clarification interfaces and mediated dialogue, where a system must be able to recognize when continuing confidently would perpetuate an earlier mistaken interpretation Schegloff et al. (1977) Dingemanse et al. (2015).
Other-Initiated Repair and the Specificity of Clarification
Other-initiated repair occurs when a recipient signals difficulty with another person's prior contribution. An open request such as 'huh?' or 'what?' leaves the nature of the problem relatively unspecified; a restricted request such as 'which station?' identifies the missing element; a candidate understanding such as 'you mean the northern station?' proposes a potential solution for confirmation. These formats differ in how much work they assign to the original speaker and how much interpretive progress the recipient has already made. Selecting an appropriate level of specificity is therefore a substantive feature of conversational competence, not merely a stylistic preference Dingemanse et al. (2014) Dingemanse et al. (2015).
Dingemanse, Blythe, and Dirksmeyer documented resources for initiating repair across languages and organized them by comparable interactional function. The analysis motivates a pragmatic typology grounded in what participants accomplish in sequences rather than translations of an English expression. Later comparative work by Dingemanse and colleagues examined repair practices across a sample of languages from diverse families, identifying recurrent functional distinctions among repair initiators. The strong result is cross-language similarity in recognizable interactional problems and repair formats. The limitations include the finite sampling of communities, recording contexts, and the analytical decisions required to compare actions across grammatical systems Dingemanse et al. (2014) Dingemanse et al. (2015).
The pragmatics of specificity can be examined experimentally by supplying controlled ambiguity and recording whether a recipient requests repetition, names the uncertain word, offers a candidate interpretation, or proceeds without checking. Appropriate behavior depends on the distribution of uncertainty: a noisy sound may warrant repetition, whereas uncertainty about which object is intended may warrant a targeted question. A study should independently score whether the proposed repair narrows uncertainty, how much burden it imposes on the other participant, and whether subsequent action verifies resolution. A shorter exchange is not always a better exchange when erroneous grounding would create larger downstream costs Clark & Brennan (1991) Dingemanse et al. (2015).
Interactional Priority and the Timing of Repair
The placement of repair in conversation poses a particularly revealing test of the relation between timing and social organization. Kendrick analyzed 169 other-initiations of repair in face-to-face interactions and found that their temporal distribution differed from that of many ordinary next turns: the most frequent cases followed gaps around 700 milliseconds. This finding, interpreted alongside close sequential inspection, challenges any account in which all successful conversational actions should occur with minimal latency. Delayed repair can accomplish social work: a recipient may search for understanding, allow the original speaker to self-correct, or display visible difficulty before asking for assistance Kendrick (2015).
Repair may also take priority when competing for access to the floor. A recipient who signals a serious problem of understanding can interrupt progression because the next activity depends on resolving what the previous action meant. The relation is local and conditional; not every interruption deserves the same priority, and not every prolonged silence indicates repair. The participant's displayed orientation, the prior turn's action, and the response to the initiation establish whether a repair sequence has occurred. Timing is both evidence of processing and part of a normative interactional practice through which speakers make room for one another to remedy trouble Sacks et al. (1974) Schegloff et al. (1977) Kendrick (2015).
An adequate predictive model should thus explain both rapid responses and the distribution of deliberately delayed clarifications. It should specify the moment at which the recipient has enough information to identify trouble, the opportunity for self-repair, and the prospective relevance of the ordinary next action. Experiments that reward only speed could penalize socially effective clarification and promote avoidable misinterpretations. A richer outcome combines the accuracy of problem localization, the eventual success of the joint task, and the temporal cost of restoring understanding. Interturn gaps become meaningful when associated with a sequentially defined communicative objective Roberts et al. (2015) Kendrick (2015).
Repair in Joint Reference and Collaborative Tasks
Collaborative reference studies provide controlled opportunities to observe the negotiation of understanding. In the experiment described by Clark and Wilkes-Gibbs, participants worked together to identify and arrange objects, revising referring expressions until they served their joint purpose. The unit of analysis was a collaborative process: an initial description could invite a query, an addition, or a reformulation, after which later references became more economical. Such evidence shows that meaning-in-use is supported by contingent contributions from both partners. It does not settle whether the participants store a particular formal representation, nor whether the same dynamics occur with equal strength in every communicative context Clark & Wilkes-Gibbs (1986).
Brennan and Clark added a historical dimension through experiments on conceptual pacts. The reuse of a jointly established description can persist when a shorter description would suffice, illustrating the role of partner-specific conversational history. A misunderstanding therefore concerns more than a transient linguistic mismatch. It can force a revision to a previously stabilized way of locating an object and can alter future expectations. Scientific evaluation should compare old and new partners, stable and altered referents, and the availability of previous labels in memory. These contrasts help distinguish simple repetition from interactionally negotiated commitment Brennan & Clark (1996) Clark & Brennan (1991).
Construction tasks in which one person directs another are especially useful because errors can be observed in joint activity. Clark and Krych demonstrated the effects of monitoring an addressee's work while giving instructions. When a speaker sees a partner hesitate or misplace an object, the instruction may be expanded before a conventional verbal clarification request arises. Such behavior shows that anticipatory and responsive repair can operate during the production of an ongoing contribution. A productive research design should track changes in both the verbal record and the physical task state, so that apparent success can be distinguished from a situation where a misunderstanding was never detected Clark & Krych (2004).
Cross-Linguistic Infrastructure and Local Variation
Research across languages demonstrates that conversational repair is more than an artifact of English expressions. Dingemanse, Torreira, and Enfield compared forms resembling 'huh?' in ten languages and argued that similar linguistic items could emerge because participants everywhere face recurring demands to request clarification. Their account presents convergent cultural evolution as a proposed mechanism, while the empirical material documents formal and functional similarities in the sampled conversations. The evolutionary interpretation should be assessed independently from the observations: recurrence can be explained by shared communicative constraints while still permitting alternative causal histories Dingemanse et al. (2013).
The broader comparison published by Dingemanse and colleagues in 2015 included twelve languages from eight language families and identified an organized repertoire of open, category-restricted, and candidate-understanding initiators. The authors report an average occurrence of other-initiated repair around once every 1.4 minutes in their corpus sample, showing how common the phenomenon is under their sampling and coding scheme. This number is a descriptive result for the analyzed material, not a global human constant. The important commonality is that interlocutors have means to flag difficulty and return the burden of resolution to the appropriate place in the exchange Dingemanse et al. (2015).
Cross-cultural differences in speech tempo and conventions of engagement remain compatible with shared sequential functions. Stivers and colleagues found quantitative variation around a general turn-taking pattern, while comparative repair research identifies functional similarities implemented by distinct grammars and interactional practices. The scientific challenge is to separate a proposed universal interactional role from the form, frequency, and appropriateness of its realization. Future comparisons require locally informed coding, balanced sampling of interactional situations, and explicit treatment of signed and multimodal languages. A category defined only by English glosses is an inadequate basis for strong universality claims Stivers et al. (2009) Dingemanse et al. (2014) Dingemanse et al. (2015).
Development of Turn-Taking and Repair Competence
Conversational coordination develops through participation in exchanges where a child must recognize a communicative action, prepare a response, and make an opportunity to contribute. Early interaction supplies repeated temporal and social contingencies, while linguistic growth changes the resources available for projection and repair. Casillas, Bobb, and Clark investigated children's responses to questions and found that response delays decrease over development, with adult-like timing emerging later than the ability to produce some relevant replies. The research links changing timing to multiple demands on comprehension and response planning. Because developmental designs often compare age groups rather than manipulating each processing stage, they support a developmental trajectory without identifying one cause of every observed difference Casillas et al. (2016).
The developmental challenge is to distinguish a child's incomplete grasp of a question from limitations in response planning, motor preparation, social expectations, or willingness to answer. Natural interactions provide opportunities to observe what children do when they mishear, misidentify an object, or receive a corrective response; structured tasks offer better control over response type and information availability. The emergence of adult-like gap timing should not be confused with the emergence of the broader ability to sustain joint activity. Children can display responsive, functional coordination with longer pauses and with different resources for clarification. An adequate model therefore tracks competence across several measures rather than selecting adult latency as the only criterion Casillas et al. (2016) Clark & Brennan (1991).
Developmental research also poses a question about the relationship between recurrent practices and linguistic learning. Repair can direct attention to a problematic form or referent and provide an occasion for a more informative example. To establish that repair contributes causally to language learning, however, investigators would need to compare developmental trajectories under controlled differences in opportunities for clarification, adjust for caregiver input and child ability, and measure subsequent retention beyond the immediate interaction. Naturalistic evidence can document plausible learning opportunities but rarely isolates their independent causal contribution. This distinction remains essential when designing educational conversation technologies modeled on adult interactional competence Casillas et al. (2016) Clark & Wilkes-Gibbs (1986).
Communication Medium, Participation Structure, and Interactional Costs
The available channel changes the observable resources for coordinating activity. Face-to-face partners can observe gaze, gesture, body posture, task progress, and the visible consequences of an instruction. Telephone partners largely rely on audible speech and nonlexical vocalization; asynchronous text partners operate with delays, message boundaries, and records that can be reread. Clark and Brennan's grounding account predicts that such affordances alter the costs of presenting an utterance, revising it, and checking whether a partner has understood. The prediction is about relationships between affordances and local collaborative work, rather than an assertion that any one medium is intrinsically superior across all tasks Clark & Brennan (1991).
In a visually grounded construction experiment, Clark and Krych showed that instruction-givers could use ongoing feedback from an addressee's work to monitor and adjust utterances. If a technological medium hides that feedback, a misunderstanding may become visible only after the recipient completes an incorrect action. If it provides persistent access to a previous message, participants may repair a reference by quoting exact wording instead of requesting repetition. These are experimentally tractable alternatives: researchers can manipulate the visibility of the task, the timing of feedback, and the availability of a written record while observing repair onset, type, completion, and downstream error. Successful generalization depends on identifying the active affordance rather than attaching a fixed mechanism to a broad category such as online conversation Clark & Krych (2004) Clark & Brennan (1991).
Participation structure complicates interpretation further. A dyadic exchange offers a relatively straightforward candidate for the next speaker, whereas group talk requires participants to manage multiple recipients, differing knowledge, and overlapping claims to speak. The turn-taking framework specifies how selection can operate locally; it does not guarantee that dyadic timing distributions can simply be transferred to multi-party meetings, interpreting, classroom discussion, or networked agents. Studies should describe who is an eligible next participant, which channels are mutually available, and what counts as a completed joint action. Repair should then be analyzed relative to the participant group for whom the problem is consequential Sacks et al. (1974) Clark & Schaefer (1989).
Interactional Data, Transcription, and Sequential Coding
The most direct evidence for conversational repair comes from recordings of interaction analyzed in their sequential environment. A rigorous corpus records how the data were selected, who participated, what activity was underway, how audio and video were synchronized, and which transcription conventions mark overlap, silence, intonation, or cut-off speech. Identification of a repair initiation requires examining the preceding trouble source and the following response, rather than searching a transcript for isolated tokens. The same interrogative word can introduce a new question or request clarification depending on sequential placement. Annotation therefore depends on explicit criteria and the demonstrable orientation of participants in the unfolding exchange Schegloff et al. (1977) Kendrick (2015).
Quantitative timing requires defensible acoustic boundaries. Heldner and Edlund showed the importance of distinguishing gaps, pauses, and overlap in recorded conversation. Investigators should define voice activity, speech onset and offset, simultaneous phonation, minimal duration thresholds, and the treatment of laughter, breathing, and backchannel vocalizations. For repair studies, delays must be assigned to a specific type of action and a justified reference boundary. A dataset of acoustic intervals without action-level coding can establish turn-transition distributions but cannot reliably answer why a partner waited or what a delayed clarification accomplished Heldner & Edlund (2010) Roberts et al. (2015).
Coding reliability is an empirical property of a study, not a guarantee attached to a transcription system. Independent annotators can be asked to identify trouble sources, repair initiators, proposed solutions, uptake, action types, and the moment of reentry into the suspended activity. Reports should document training materials, uncertainty, disagreements, and an appropriate agreement statistic when the coding task permits one. Qualitative examination remains necessary for borderline sequences where the participants' next conduct determines which interpretation is warranted. A publishable evidence chain preserves both excerpts that demonstrate how a practice works and an accountable sampling frame for claims about how frequently it occurs Dingemanse et al. (2014) Kendrick (2015).
Experimental Paradigms and Causal Identification
Collaborative reference games permit experimental manipulation of what partners know, which descriptions have already been negotiated, and whether a previous misunderstanding receives a correction. Clark and Wilkes-Gibbs's referential work and Brennan and Clark's conceptual-pact experiments exemplify how an interactional phenomenon can be rendered experimentally observable through later reference choices. A stronger intervention systematically distinguishes effects of prior exposure from partner-specific coordination by crossing repeated objects with partner identity, shared history, and visible contextual information. The dependent measures should include recognition, selection errors, clarification requests, and delayed transfer, since fluency alone may conceal uncertainty Clark & Wilkes-Gibbs (1986) Brennan & Clark (1996).
Turn-completion experiments create a different inferential structure. De Ruiter, Mitterer, and Enfield manipulated lexical and intonational information available during the prediction of an utterance boundary; Magyari and de Ruiter used lexical anticipation in a gating paradigm; Bögels, Magyari, and Levinson studied planning-related neural timing. The convergence of methods supports an account in which linguistic prediction and early response preparation contribute to rapid exchange. Yet manipulated, isolated utterances may underrepresent the social pressures and participant expectations of a naturally embedded sequence. Strong evaluation would therefore combine controlled predictions with a follow-up interaction in which participants can actually take the floor and negotiate repairs de Ruiter et al. (2006) Magyari & de Ruiter (2012) Bögels et al. (2015).
Experimental repair tasks should create authentic information asymmetries rather than an artificial rule to issue a clarification request. A speaker might be given an ambiguous target, while a listener has access to several plausible referents; experimental factors could vary the cost of a wrong choice and the availability of a targeted question. Investigators can then estimate whether clarification increases final success, how often an uncertain participant incorrectly commits, and whether different initiation formats impose predictable time and cognitive costs. Random assignment, preregistered coding decisions, manipulation checks, and task-specific baselines are necessary before describing differences as causal improvements in conversational competence Clark & Krych (2004) Dingemanse et al. (2015).
Theoretical Accounts of Alignment and Interpersonal Synergy
A theoretical synthesis must explain how local conversational actions become coordinated without treating identical behavior as the sole objective. Alignment accounts emphasize how partners develop comparable linguistic expressions or representations through interaction, while a synergistic approach treats dialogue as an organized system of complementary actions. Fusaroli, Rączaszek-Leonardi, and Tylén developed the notion of interpersonal synergy, highlighting context-sensitive alignment, complementary dynamics, and interactional routines. Its relevance is that a successful conversation may require one person to ask and another to answer, rather than both to produce matching utterances. The important test is whether proposed forms of coordination predict performance under changes to task demands and partner knowledge Fusaroli et al. (2014).
Grounding accounts instead emphasize the public evidence required to accept an utterance into a common activity. The advantage of this framework is that it connects linguistic choice with practical criteria for continuation, including correction. The turn-taking framework specifies the opportunities through which those contributions can occur, and repair analysis explains how a breakdown is handled in the sequence. These approaches describe mutually relevant structures at distinct levels. Integrating them requires explicit mappings: a projection of possible turn completion creates an opportunity for a contribution; that contribution may display understanding; a trouble signal initiates repair; and the resolution alters subsequent reference practices Clark & Brennan (1991) Sacks et al. (1974) Schegloff et al. (1977).
No single indicator exhausts these theoretical mechanisms. Speech-rate similarity may be a form of coordination, but it can also arise from stable personal habits. Shared lexical choices may reflect a conceptual pact or repeated exposure to the same visual context. A rapid next turn can be correctly responsive or prematurely committed to an erroneous interpretation. Model comparison should favor theories that predict when people will align, when complementary behavior is preferable, when clarification is requested, and how earlier repair shapes later actions. Such predictions can be adjudicated through partner-switch manipulations, sequential analysis, and independent tests on conversations involving new tasks and participants Brennan & Clark (1996) Fusaroli et al. (2014) Roberts et al. (2015).
Conversational Coordination with Artificial Participants
Dialogue systems introduce a setting in which the outward organization of turns can be tested independently of assumptions about biological language production. An artificial participant can be evaluated for how it handles the current question, whether it maintains a previously established referent, whether it detects an underspecified instruction, and whether a correction changes later behavior. The relevant performance unit is a sequence of jointly consequential actions, because an isolated grammatical response can be produced without an adequate appreciation of its place in the activity. Conversation analysis and grounding theory supply useful functional categories, while controlled human–computer experiments determine which of those categories a particular system reliably instantiates Sacks et al. (1974) Clark & Brennan (1991) Shaikh et al. (2024).
Shaikh and colleagues compared language-model generations with human conversational data through an explicit taxonomy of grounding acts. Their 2024 study reported that the tested models produced fewer acts oriented to establishing common ground, often proceeding as though understanding had already been secured. The result is a comparative behavioral finding under the study's datasets, sampling, and measurement system. It does not demonstrate that all artificial systems lack the capacity for clarification, nor does the frequency of a particular question establish genuine shared understanding. A system could mechanically overproduce clarifications, generating avoidable delays while still making erroneous commitments Shaikh et al. (2024).
These observations motivate a functional account of artificial dialogue performance. A successful agent should estimate when more information is needed, ask a locally relevant question, incorporate the partner's answer, and preserve the revised state in later decisions. It should also recognize cases in which an ordinary response is sufficiently supported and further questioning would impose needless effort. The empirical question is whether these capabilities survive changes in partner, task, phrasing, error severity, and history length. Observable repair competence concerns successful behavior under specified tests; stronger claims about artificial subjectivity or lived understanding require a separate evidential framework Shaikh et al. (2024) Clark & Krych (2004) Fusaroli et al. (2014).
Human–Artificial Grounding Failures and Repair Benchmarks
Chiyah-Garcia, Suglia, and Eshghi constructed BlockWorld-Repairs to examine third-position repair in a visually grounded instruction-following environment. In such a sequence, an artificial addressee responds to an instruction, its response displays an interpretation problem, and the original speaker corrects the mistaken interpretation. The benchmark evaluates whether vision-and-language models can use that correction to recover the intended task state. The researchers found that evaluated systems underperformed human participants in the tested repair conditions. This is evidence about a specific multimodal task, not a comprehensive estimate of a model's capacity in unrestricted social conversation Chiyah-Garcia et al. (2024).
Shaikh and colleagues subsequently analyzed grounding behavior in three human–assistant datasets and developed the Rifts benchmark. Their reported comparison found that language models in the sampled settings were three times less likely to initiate clarification and sixteen times less likely to make follow-up requests than humans. They also reported that early grounding failures were predictive of subsequent interaction breakdowns. These numerical ratios are conditional on the paper's operational categories and comparison design; they cannot be transferred as fixed population rates to arbitrary deployed assistants. The study nonetheless makes a substantial contribution by connecting missed opportunities for clarification with downstream consequences in interaction Shaikh et al. (2025).
The contrast between behavioral benchmarks and natural human conversation suggests a stronger evaluation framework. Tests should distinguish first-position trouble recognition, second-position clarification, third-position correction, and sustained repair after the context changes. They should compare ambiguous and unambiguous tasks, measure costs of unnecessary questions, examine actual recovery rather than an acknowledgment of correction, and require successful performance on later turns. Such benchmarks would treat conversational coordination as a longitudinal property of an interactional process. A high score on one-shot question answering cannot substitute for evidence that earlier misunderstandings were located and resolved without compounding them Chiyah-Garcia et al. (2024) Shaikh et al. (2025) Schegloff et al. (1977).
Evidence Quality and Alternative Scientific Explanations
The evidence base combines strong but methodologically different traditions. Conversation analysis demonstrates recurring practices through fine-grained recordings; its generalizations depend on the diversity of recorded settings and the credibility of sequential interpretation. Cross-language corpora give quantitative evidence for recurrent repair formats, but their rates depend on sampling, transcription, and coding. Laboratory studies isolate variables such as available lexical information, partner history, and visual feedback; they sacrifice some of the consequences that make a naturally occurring interaction consequential. Neural and respiratory measures constrain temporal models while requiring assumptions about what physiological signatures mean. These limits identify an integrative research problem rather than invalidating any of the methods Schegloff et al. (1977) Dingemanse et al. (2015) de Ruiter et al. (2006) Bögels et al. (2015).
Competing explanations deserve explicit comparison. Short interturn gaps may reflect early linguistic prediction, familiar action sequences, simple responses, or anticipatory readiness; the observed latency alone cannot isolate them. Lexical repetition may indicate coordination, a stable personal preference, or the constraints of the available vocabulary. An apparent repair can signal a genuine correction, a polite display of attention, or a new misunderstanding disguised by an affirmative response. Researchers should formulate mutually discriminating predictions and manipulate the information that different theories require. A useful synthesis preserves the explanatory commitments of each model while specifying where its predictions diverge from alternatives Roberts et al. (2015) Fusaroli et al. (2014) Brennan & Clark (1996).
Generalizability also depends on population and setting. Published corpora contain finite languages, recording situations, and interactional roles; many studies prioritize adults in familiar institutional or laboratory contexts. Some participant groups communicate through signed, augmentative, or assisted modalities that require distinct methods for identifying completion, feedback, and repair. Technology-mediated interactions introduce new timing and memory affordances. Claims that a conversational practice is universal should distinguish its recurring functional role from the specific realization documented in a particular corpus. Replication across independent laboratories and communities, with comparable definitions and locally informed coding, remains a central scientific requirement Stivers et al. (2009) Dingemanse et al. (2014) Dingemanse et al. (2015).
A Cumulative Research Program for Coordination and Repair
A cumulative program should begin with a common description of observable interactional phenomena and preserve the contextual information necessary to interpret them. Corpus releases can provide deidentified audio or video when ethically permissible, time-aligned transcripts, definitions of relevant boundaries, and examples of ambiguous classifications. Researchers should report whether observations derive from task-oriented laboratory exchanges, informal multiparty conversation, remote communication, or training dialogues. Shared annotation does not eliminate differences among these situations; it makes their effects measurable. The central objective is to connect replicable distributions to the sequential practices that produced them Heldner & Edlund (2010) Roberts et al. (2015) Dingemanse et al. (2015).
The next level is experimental triangulation. Turn-end prediction paradigms should be paired with settings where a participant actually contributes and receives feedback; grounding paradigms should be paired with delayed tests of whether negotiated descriptions remain usable; repair benchmarks should track downstream effects after a correction. Developmental comparisons should separate changes in linguistic knowledge from planning demands and communicative norms. A design that includes several kinds of outcome can reveal whether an intervention improves comprehension, turn timing, accuracy of clarification, and recovery together or trades one goal against another de Ruiter et al. (2006) Casillas et al. (2016) Clark & Wilkes-Gibbs (1986) Chiyah-Garcia et al. (2024).
Research on artificial participants offers an especially clear opportunity for controlled perturbations, since an evaluator can deliberately vary task history, conflicting messages, available referents, and the cost of incorrect action. However, synthetic scenarios need independent validation against real human interaction. Model outputs should be assessed under multiple prompts and across versions, with explicit attention to whether a system has access to the requisite history. A reliable artificial-side repair system should demonstrably improve the partner's subsequent capacity to act, not simply increase the number of apologetic or clarifying phrases. This program connects interactional science with the Artificial Era through measurable, partner-sensitive performance rather than through assumptions imported from technological identity Shaikh et al. (2024) Shaikh et al. (2025) Clark & Brennan (1991).
Conclusion
Conversational coordination and repair form an integrated organization of human interaction. Turn taking creates recurrent opportunities to contribute; projection and response planning enable timely participation; grounding establishes workable evidence of understanding; and repair permits speakers and listeners to reorient their joint activity when something goes wrong. These processes vary in temporal scale but converge on a common practical achievement: participants can make the course of a conversation intelligible, accountable, and open to revision. The foundational studies of turn allocation, self-repair, collaborative reference, and grounding remain central because they each identify a distinct but interacting component of this achievement Sacks et al. (1974) Schegloff et al. (1977) Clark & Wilkes-Gibbs (1986) Clark & Brennan (1991).
The strongest scientific picture combines the observational detail of conversation analysis, the comparative reach of cross-language corpora, the causal leverage of experiments, and the temporal constraints furnished by physiological measurement. Cross-linguistic work documents recurring means for identifying and resolving trouble, while developmental and cognitive studies help explain the coordination of comprehension, planning, and speech onset. No single number for response latency, count of clarification words, or model alignment measure provides a comprehensive account. The strongest inferences follow from convergent evidence concerning what a participant actually accomplished, how the partner reacted, and whether the resumed activity proceeded on a sound interpretation Stivers et al. (2009) Dingemanse et al. (2015) Bögels et al. (2015) Casillas et al. (2016).
The research horizon extends into mediated and human–artificial dialogue, where systems must handle uncertainty, respond to correction, and sustain changes in conversational state. Contemporary benchmarks have identified consequential weaknesses in artificial grounding and repair, while demonstrating ways to construct operational tests. The outstanding scientific task is a cumulative account connecting sequential organization, cognitive prediction, the pragmatics of clarification, and the measured consequences of successful or failed recovery. Conversational competence is best evaluated through the quality of sustained joint coordination: the capacity to advance a shared activity while maintaining the means to question, correct, and reestablish what its participants have understood Chiyah-Garcia et al. (2024) Shaikh et al. (2025) Fusaroli et al. (2014).
References
- Bavelas, J. B., Coates, L., & Johnson, T. (2000). Listeners as co-narrators. Journal of Personality and Social Psychology, 79(6), 941–952. https://doi.org/10.1037/0022-3514.79.6.941
- Bögels, S., Magyari, L., & Levinson, S. C. (2015). Neural signatures of response planning occur midway through an incoming question in conversation. Scientific Reports, 5, Article 12881. https://doi.org/10.1038/srep12881
- Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts and lexical choice in conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(6), 1482–1493. https://doi.org/10.1037/0278-7393.22.6.1482
- Casillas, M., Bobb, S. C., & Clark, E. V. (2016). Turn-taking, timing, and planning in early language acquisition. Journal of Child Language, 43(6), 1310–1337. https://doi.org/10.1017/S0305000915000689
- Chiyah-Garcia, J., Suglia, A., & Eshghi, A. (2024). Repairs in a block world: A new benchmark for handling user corrections with multi-modal language models. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 11523–11542. https://doi.org/10.18653/v1/2024.emnlp-main.643
- Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), Perspectives on socially shared cognition (pp. 127–149). American Psychological Association. https://web.stanford.edu/~clark/1990s/Clark%2C%20H.H.%20_%20Brennan%2C%20S.E.%20_Grounding%20in%20communication_%201991.pdf
- Clark, H. H., & Krych, M. A. (2004). Speaking while monitoring addressees for understanding. Journal of Memory and Language, 50(1), 62–81. https://doi.org/10.1016/j.jml.2003.08.004
- Clark, H. H., & Schaefer, E. F. (1989). Contributing to discourse. Cognitive Science, 13(2), 259–294. https://doi.org/10.1207/s15516709cog1302_7
- Clark, H. H., & Wilkes-Gibbs, D. (1986). Referring as a collaborative process. Cognition, 22(1), 1–39. https://doi.org/10.1016/0010-0277(86)90010-7
- de Ruiter, J. P., Mitterer, H., & Enfield, N. J. (2006). Projecting the end of a speaker's turn: A cognitive cornerstone of conversation. Language, 82(3), 515–535. https://doi.org/10.1353/lan.2006.0130
- Dingemanse, M., Blythe, J., & Dirksmeyer, T. (2014). Formats for other-initiation of repair across languages: An exercise in pragmatic typology. Studies in Language, 38(1), 5–43. https://doi.org/10.1075/sl.38.1.01din
- Dingemanse, M., Roberts, S. G., Baranova, J., Blythe, J., Drew, P., Floyd, S., Gísladóttir, R. S., Kendrick, K. H., Levinson, S. C., Manrique, E., Rossi, G., & Enfield, N. J. (2015). Universal principles in the repair of communication problems. PLOS ONE, 10(9), Article e0136100. https://doi.org/10.1371/journal.pone.0136100
- Dingemanse, M., Torreira, F., & Enfield, N. J. (2013). Is “Huh?” a universal word? Conversational infrastructure and the convergent evolution of linguistic items. PLOS ONE, 8(11), Article e78273. https://doi.org/10.1371/journal.pone.0078273
- Fusaroli, R., Rączaszek-Leonardi, J., & Tylén, K. (2014). Dialog as interpersonal synergy. New Ideas in Psychology, 32, 147–157. https://doi.org/10.1016/j.newideapsych.2013.03.005
- Heldner, M., & Edlund, J. (2010). Pauses, gaps and overlaps in conversations. Journal of Phonetics, 38(4), 555–568. https://doi.org/10.1016/j.wocn.2010.08.002
- Kendrick, K. H. (2015). The intersection of turn-taking and repair: The timing of other-initiations of repair in conversation. Frontiers in Psychology, 6, Article 250. https://doi.org/10.3389/fpsyg.2015.00250
- Levinson, S. C., & Torreira, F. (2015). Timing in turn-taking and its implications for processing models of language. Frontiers in Psychology, 6, Article 731. https://doi.org/10.3389/fpsyg.2015.00731
- Magyari, L., & de Ruiter, J. P. (2012). Prediction of turn-ends based on anticipation of upcoming words. Frontiers in Psychology, 3, Article 376. https://doi.org/10.3389/fpsyg.2012.00376
- Roberts, S. G., Torreira, F., & Levinson, S. C. (2015). The effects of processing and sequence organization on the timing of turn taking: A corpus study. Frontiers in Psychology, 6, Article 509. https://doi.org/10.3389/fpsyg.2015.00509
- Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696–735. https://doi.org/10.2307/412243
- Schegloff, E. A., Jefferson, G., & Sacks, H. (1977). The preference for self-correction in the organization of repair in conversation. Language, 53(2), 361–382. https://doi.org/10.2307/413107
- Shaikh, O., Gligorić, K., Khetan, A., Gerstgrasser, M., Yang, D., & Jurafsky, D. (2024). Grounding gaps in language model generations. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 6279–6296. https://doi.org/10.18653/v1/2024.naacl-long.348
- Shaikh, O., Mozannar, H., Bansal, G., Fourney, A., & Horvitz, E. (2025). Navigating rifts in human-LLM grounding: Study and benchmark. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 20832–20847. https://doi.org/10.18653/v1/2025.acl-long.1016
- Stivers, T., Enfield, N. J., Brown, P., Englert, C., Hayashi, M., Heinemann, T., Hoymann, G., Rossano, F., de Ruiter, J. P., Yoon, K.-E., & Levinson, S. C. (2009). Universals and cultural variation in turn-taking in conversation. Proceedings of the National Academy of Sciences, 106(26), 10587–10592. https://doi.org/10.1073/pnas.0903616106
- Torreira, F., Bögels, S., & Levinson, S. C. (2015). Breathing for answering: The time course of response planning in conversation. Frontiers in Psychology, 6, Article 284. https://doi.org/10.3389/fpsyg.2015.00284