Dialogue is a collaborative activity during which speakers interact to reach a common goal, such as establishing a route together (Clark, 1996; Clark & Wilkes-Gibbs, 1986). As speakers interact, they increment their common ground, which consists in the information that they are aware of sharing (Clark & Marshall, 1981). The common ground includes the references produced by the speakers earlier in the interaction to refer to objects and entities (e.g., the landmarks to be encountered on a route).

Information is grounded (i.e., added to the common ground) through a joint contribution process (Clark & Schaefer, 1989). One of the speakers starts by presenting a piece of information (e.g., a reference); the latter then accepts this information by signaling that it has been understood well enough for the current purposes (Clark & Brennan, 1991). Reference acceptance can be explicit, with the addressee accepting the reference through verbatim repetition or through anaphoric repetition (i.e., the addressee uses a pronoun to re-refer to the same object or entity). Acceptance can also be implicit, with the addressee initiating the next relevant speech turn. A grounded reference can potentially be reused by either speaker during the interaction (Brennan & Clark, 1996).

Dialogue being a collaborative activity implies that all speakers put efforts into achieving mutual understanding (Clark, 1996). One way of doing so consists of each speaker using the common ground to determine which references his or her partner is capable of understanding, reasoning that he or she should be capable of understanding a reference that had been successfully understood earlier in the interaction (Clark & Wilkes-Gibbs, 1986; Haviland & Clark, 1974; Isaacs & Clark, 1987). However, all of the references from the common ground are not equally likely to be reused, since this depends on their accessibility in memory (e.g., Horton, 2008). This is partly in line with an egocentric approach to dialogue, with reference production depending on the speaker’s state of mind rather than on the addressee’s (e.g., Barr & Keysar, 2002).

In the present study, we sought to investigate reference reuse further. A series of studies conducted in the field of memory research has shown that words produced aloud are remembered better than words read silently (Forrin, Ozubko, & MacLeod, 2012; MacLeod, Gopie, Neary, & Ozubko, 2010; Ozubko, Hourihan, & MacLeod, 2012); this production effect is all the stronger when the production is self-performed (MacLeod, 2011). These findings could have implications for reference reuse in dialogue. First, all speakers involved in dialogue might present references at some point. Self-produced words being more readily accessible suggests that self-presented references should be reused more often than partner-presented ones. Second, acceptance sometimes involves verbatim repetition (Clark & Brennan, 1991). This should cause repeated references to benefit from a self-production effect from the addressee’s perspective, as well. Furthermore, dialogue partners are exposed to references accepted through verbatim repetition twice (once at the time of presentation, and once at the time of acceptance). Because the production effect concerns both self- and partner-produced words, repeated references should benefit from a production effect from each partner’s perspective. An additional question concerns whether such accessibility differences persist after the end of an interaction, since dialogue partners might sometimes need to resort to the common ground established during past interactions. Memory accessibility after the end of the interaction might also depend on how many times a reference had actually been produced during the interaction.

Moreover, multipartite dialogue involves both ratified participants (i.e., participants addressing or being addressed by a partner) and side participants (participants addressing no one and being addressed by no one; Clark, 1996). Side participants gather common ground (Wilkes-Gibbs, 1992), but they do not have the opportunity to actually produce references. Thus, the accessibility in memory of references produced while one was a side participant should be fairly low. This would be consistent with the idea that indirectly established common ground has a weaker influence on reference production (Gorman, Gegg-Harrison, Marsh, & Tanenhaus, 2013), and with the more general idea that active learning is more efficient than passive learning (e.g., Bonwell & Sutherland, 1996).

Two experiments were conducted to address these assumptions. Participants referred to landmarks as they performed a route description task. In Experiment 1, triadic dialogue was investigated to determine whether the conditions in which a reference is initially grounded affect its subsequent reuse. Reference recall was then used to assess reference accessibility in memory after the end of the interaction. In Experiment 2, we sought to replicate the findings obtained in Experiment 1 in dyadic dialogue.

Experiment 1

Participants

A group of 54 native French-speaking students signed an informed consent form before taking part in the experiment for course credit.

Material and apparatus

Three identical versions of a map featuring 25 monuments, nine squares, 57 streets names, and three points (A, B, and C; Fig. 1, left panel) and three identical blank versions of this map (Fig. 1, right panel) were printed. The interactions were recorded using two microphones and a digital recorder.

Fig. 1
figure 1

Maps used during Experiment 1. (Left) Map with landmarks used during the dialogue phase. (Right) Blank map used during the drafting phase.

Full size image

Task and procedure

Three participants took part in the experiment. Each sat facing a different wall of the experimental room, so that they could not communicate through nonlinguistic cues. Their task was to establish a tourist route for a person who had no prior knowledge of the town represented on the maps. They knew that they would have to individually write out the entire route after the interaction, but they did not know that they would not have access to the initial map while doing so.

During the first part of the experiment (the dialogue phase), the three participants used their maps to agree on a route running from A to B, from B to C, and from C back to A. Each section was discussed by two ratified participants only, with the third, side participant listening but not being able to intervene. For instance, P01 and P02 would discuss section A–B, P02 and P03 would discuss Section B–C, and P01 and P03 would discuss Section C–A (the three participants were identified depending on their random entry order in the room). This phase lasted for a maximum of 20 min. The time spent on each route section was not determined in advance.

During the second phase of the experiment (the drafting phase), the participants had a maximum of 15 min to individually write out the entire route. They were given a blank version of the map to guide their recall. They could not communicate, nor write anything on their map during this phase.

Experimental design

Three within-participants independent variables were used. The first one was reference status. From each participant’s point of view, a reference had been either self-presented or presented by a partner while one was either a ratified participant or a side participant.

The second independent variable was acceptance type. Within each triad, a reference had been accepted through verbatim repetition, anaphorically, or implicitly.

The third independent variable was the number of times a reference was reused within the triad. This was a standardized continuous independent variable.

Data coding

The interactions between the participants were transcribed and coded for presentation, acceptance, and reuse (see the Appendix for more detail).

When a reference was presented for the first time in a triad, it was coded as being presented for P01, P02, and P03. Who the current speaker and the ratified participant were at the time was used to code for reference status from each participant’s point of view. The evidence produced by the other ratified participant between the moment when the reference was presented and the moment when the initiator of the reference produced another reference was examined to code the reference for acceptance type. All other occurrences of reference production were classified as reuse; the only criterion was that reuse needed to occur in a speech turn preceded by a minimum of two speech turns during which the reference was not produced, which helped distinguish reuse from simple repetition. Two different levels of coding were used: At the participant level, reuse was coded as a dichotomous variable (a reference could be reused or not reused by a participant, regardless of how many times he or she reused it), whereas it was coded as a continuous variable at the triadic level (this coding reflected how many times each reference was reused in the triad, regardless of who had reused it).

Data coding: Drafting phase

The routes written out by the participants were transcribed and coded for reference recall: For each participant, each presented reference could be either recalled or not recalled.

Results and discussion

During the dialogue phase, the average number of words produced per triad was 1,318.94 (SD = 704.18) (A–B: 428.17, SD = 192.40; B–C: 463.28, SD = 333.58; C–A: 427.50, SD = 274.58). A total of 734 references were presented (40.78, SD = 7.76, per triad on average, and 13.59, SD = 5.04, per participant on average). Among these, 148 (20.16 %) were accepted through verbatim repetition, 355 (48.37 %) were accepted anaphorically, and 231 (31.47 %) were accepted implicitly. During the drafting phase, the mean number of words per individual route was 181.24 (SD = 81.94), and the average number of references recalled was 8.61 (SD = 3.77).

The data were analyzed in SPSS 22.0. Multilevel models were used to account for the nesting of the participants within the triads. Such models include random intercepts to account for variability across participants (and, potentially, items) and random slopes to account for participants’ (and items’) different sensitivities to the independent variables. Whenever possible, all random effects justified by the experimental design should be included (Barr, Levy, Scheepers, & Tily, 2013). The analyses reported hereafter included (1) by-triad and by-participant intercepts, to account for variability across triads and across participants; (2) by-participant random slopes corresponding to the independent variables, to account for the participants’ different sensitivities to these variables; and (3) by-participant random slopes corresponding to the landmarks, to account for the participants potentially behaving differently, depending on the landmarks. An additional factor, Section Responsibility, identified which sections (A–B, B–C, and C–A) had been discussed by which participants. By-participant random slopes corresponding to this factor were included to account for its potential influence on reference production. An identity variance–covariance matrix was used. The Satterthwaite correction was applied to estimate the degrees of freedom in the analyses.

Because the two variables used as outcomes in the analyses were binary, logistic mixed models were used. One parameter returned by logistic regression models is the odds ratio (OR; Jaccard, 2001), which is informative with regard to the effect size (see Agresti, 2002). Only the significant effects were included in the models. When necessary, additional comparisons were conducted using paired comparisons (sequential Bonferroni, p < .05).

Reference reuse: Dialogue phase

A total of 809 references were reused at least once, but only the data corresponding to the references that had been presented—from each participant’s point of view—while he or she was a ratified participant were considered, since the participants seldom reused the references presented while they were side participants. The 43 cases in which this happened were discarded from further analysis.

The model included reference status and acceptance type as fixed effects, and whether the reference was reused as the outcome variable (Table 1).

Table 1 Experiment 1: Numbers of references reused during the dialogue phase as a function of reference status and acceptance type
Full size table

Reference status significantly predicted reference reuse, F(1, 386) = 5.71, p = .017. Self-presented references were more likely to be reused than partner-presented ones, OR = 1.36, CI .95 = 1.06–1.75. Acceptance type also significantly predicted reference reuse, F(2, 862) = 5.51, p = .004. References accepted anaphorically and implicitly were less likely to be reused than those accepted through verbatim repetition, OR = .67, CI .95 = .49–.93, p = .017, and OR = .55, CI .95 = .39–.79, p = .001, respectively. References accepted anaphorically were more likely to be reused than references accepted implicitly (sequential Bonferroni, p < .05). This confirms that reference reuse depends on how a reference was initially grounded.

Reference recall: Drafting phase

The participants recalled a total of 480 references. Among these, 15 that had not been presented during the dialogue phase were discarded from the analysis. The model included reference status, acceptance type, and the number of times a reference had been reused by a triad as fixed effects, and whether the reference was recalled as the outcome variable (Table 2). Note that although the number of reuses was measured at the triad level, this analysis focused on recall at the participant level.

Table 2 Experiment 1: Numbers of references recalled during the drafting phase as a function of reference status and acceptance type
Full size table

Reference status significantly predicted reference recall, F(2, 1741) = 4.76, p = .009. References presented by a partner while one was a side participant were less likely to be recalled than were self-presented ones, OR = .66, CI .95 = .50–.86, p = .003. However, the difference between the references presented by a partner while one was a ratified participant and self-presented ones failed to reach statistical significance, p = .442. The difference between references presented by a partner while one was a ratified participant and references presented while one was a side participant also failed to reach statistical significance (sequential Bonferroni, p > .05).

Acceptance type also significantly predicted reference recall, F(2, 2196) = 4.03, p = .018. References accepted anaphorically were less likely to be recalled than references accepted through verbatim repetition, OR = .67, CI .95 = .50–.88, p = .005. However, the difference between implicitly accepted references and references accepted through verbatim repetition failed to reach statistical significance, p = .104. The difference between references accepted anaphorically and references accepted implicitly also failed to reach statistical significance (sequential Bonferroni, p > .05).This pattern of results only partly replicates the pattern obtained for reference reuse.

The number of reuses during the dialogue phase also predicted reference recall, F(1, 180) = 79.06, p < .001. The odds of recalling a reference increased with the number of reuses in the triad, OR = 1.72, CI .95 = 1.53–1.95.

To recap, Experiment 1 showed that self-presented references were more likely to be reused during dialogue than were partner-presented references, and that references accepted through verbatim repetition were more likely to be reused than references accepted through other means. Some of these effects had a longer-term influence on reference memorization. Experiment 1 also shed light on the influence of active participation in the dialogue on subsequent reference memory. However, one potential limitation of this experiment was that it involved triads of participants: The results could have been due to the ratified participants knowing that a side participant was listening to them and making extra efforts to repeat the references during the interaction. A second experiment was thus conducted to attempt to replicate the results reported above in a dyadic dialogue situation. Dyads of participants performed the same task as in Experiment 1; the only difference was that both participants acted as ratified participants during the entire dialogue phase.

Experiment 2

Participants

A group of 54 students were recruited under the same conditions as in Experiment 1.

Material

The materials were similar to those used in Experiment 1, except that the map only featured two points (A and B; see Fig. 2).

Fig. 2
figure 2

Maps used during Experiment 2. (Left) Map with landmarks used during the dialogue phase. (Right) Blank map used during the drafting phase.

Full size image

Task and procedure

The task and procedure were similar to those of Experiment 1. The participants’ task was to establish a return route between A and B.

Experimental design and data coding

The experimental design and coding were the same as in Experiment 1, except that the reference status independent variable only had two modalities (self- and partner-presented).

Results

During the dialogue phase, the average number of words produced per dyad was 1,378.33 (SD = 756.60). A total of 1,002 references were presented (37.11, SD = 11.45, per dyad on average, and 18.56, SD = 6.70, per participant on average). Among these, 201 (20.06 %) were accepted through verbatim repetition, 506 (50.50 %) were accepted anaphorically, and 295 (29.44 %) were accepted implicitly. During the drafting phase, the average number of words per individual route was 159.70 (SD = 59.38), and the average number of references recalled was 8.61 (SD = 3.77).

The statistical analyses were conducted in the same way as those reported for Experiment 1, except that the random part of the model included no Section Responsibility factor.

Reference reuse: Dialogue phase

A total of 1,032 references were reused at least once (Table 3).

Table 3 Experiment 2: Numbers of references reused during the dialogue phase as a function of reference status and acceptance type
Full size table

Reference status significantly predicted reference reuse, F(1, 245) = 4.32, p = .039; self-presented references were more likely to be reused than partner-produced ones, OR = 1.26, CI .95 = 1.01–1.58. Acceptance type also significantly predicted reference reuse, F(2, 468) = 7.74, p < .001: References accepted implicitly were less likely to be reused than references accepted through verbatim repetition, OR = .58, CI .95 = .43–.79, p < .001. However, the difference between references accepted anaphorically and references accepted through verbatim repetition failed to reach statistical significance. References accepted anaphorically were more likely to be reused than references accepted implicitly (sequential Bonferroni, p < .05). As in Experiment 1, reuse depended on the circumstances in which a reference was initially grounded.

Reference recall: Drafting phase

The participants recalled a total of 460 references (see Table 4). Among these, 17 that had not been presented during the dialogue phase were discarded from the analysis.

Table 4 Experiment 2: Numbers of references recalled during the drafting phase as a function of reference status and acceptance type
Full size table

The influence of reference status on reference recall failed to reach statistical significance, F < 1. As in Experiment 1, the self-presentation benefit was attenuated after the end of the interaction.

Acceptance type still significantly predicted reference recall, F(2, 752) = 3.63, p = .027. References accepted anaphorically and those accepted implicitly were less likely to be recalled than references accepted through verbatim repetition: OR = .72, CI .95 = .52–.99, p = .043, and OR = .62, CI .95 = .43–.88, p = .008, respectively. The difference between references accepted anaphorically versus implicitly failed to reach statistical significance (sequential Bonferroni, p > .05).

The number of reuses during the dialogue phase also predicted reference recall, F(1, 139) = 104.97, p < .001; the odds of recalling a reference increased with the number of reuses in the dyad, OR = 2.26, CI .95 = 1.93–2.64.

The overall pattern of results obtained in Experiment 1 was replicated in Experiment 2. The only differences pertained to acceptance (Table 5). Such differences were probably due to the numbers of references accepted through verbatim repetition, anaphorically, and implicitly varying a great deal, which might have caused the statistical power of the analyses to decrease. Nonetheless, the general pattern that emerges here is that references accepted through verbatim repetition were reused more and recalled better than other references, which is consistent with the production effect hypothesis.

Table 5 Summary of the results for reference reuse and recall, as a function of acceptance type in Experiments 1 and 2
Full size table

General discussion

This study focused on the influence of grounding on subsequent reference accessibility. In line with MacLeod (2011), the participants in both experiments showed an egocentric bias toward reusing self-presented references more, suggesting that references from the common ground continue to “belong” to their initiator during the interaction. However, this effect was attenuated after the end of the interaction. One possible explanation stems from the fact that the references that had been reused more were also recalled better: Repeated reuse of partner-presented references by their initiator might have caused the accessibility of these initially less accessible references to increase, thus eventually attenuating the difference between self- and partner-presented references. In addition, the production effect grew all the stronger as the number of exposures to a reference at the time of grounding increased, since references accepted through verbatim repetition were reused more by all dialogue partners. These references being reused more could also help explain why they were remembered better. Finally, reference accessibility after the end of the interaction depended on the role played during the interaction (in line with the results of Gorman et al., 2013). This could be due to side participants not having the opportunity to produce the references under discussion themselves; it could also be due to references produced by others constituting weaker episodic traces, and therefore being remembered less well (e.g., Andersson & Rönnberg, 1997). The reuse of such references was not investigated directly in this study, due to the task constraints. However, in a situation in which speakers do have the opportunity to reuse such references, the fact that they are less accessible in memory could have an influence on dialogue management.

These findings help bridge a theoretical gap between research on memory and research on dialogue, showing how “ordinary” memory mechanisms might constrain higher-level processes involved in collaboration (e.g., Horton, 2008). Specifically, these findings build on Clark and Schaefer’s (1989) model by showing how accessibility differences that appear during the first two steps of dialogue management—presentation and acceptance—strongly influence the third step—namely reuse. A number of studies have sought to determine whether speakers are capable of taking common ground into account during dialogue (e.g., Horton & Keysar, 1996), leading to the idea that common ground and other sources of information might both constrain language processing during dialogue (e.g., Hanna, Tanenhaus, & Trueswell, 2003). Our findings contribute to this debate by showing that even when common ground is taken into account, each speaker behaves egocentrically, since what is taken for common ground depends on a reference’s accessibility in memory from each speaker’s point of view. Thus, collaborative behavior is not necessarily nonegocentric.

Moreover, this study shows that reference accessibility varies as the interaction unfolds. We have suggested that the self-presentation egocentric bias could be (partly) compensated for through reuse. Indeed, within a dyad, for instance, Speaker A’s and Speaker B’s egocentric biases are complementary, since Speaker A’s self-presented references correspond to Speaker B’s partner-presented references, and vice versa. Each speaker being more likely to reuse his or her own self-presented references causes all references to be equally likely to be reused at the dyadic level; such repeated exposure could then contribute to decreasing initial individual biases. However, not all differences in accessibility diminish during the interaction. Contrary to the self-presentation benefit, the accessibility of references accepted through verbatim repetition either remains constant or increases as the interaction unfolds, since both speakers are initially equally likely to reuse these.

The self-presentation benefit decreasing as the verbatim repetition benefit increases could help explain why it is generally considered that a reference that belongs to the common ground can be reused by either speaker (e.g., Brennan & Clark, 1996). This study has allowed us to capture a transient egocentric bias whose effects could not be observed in the longer term. Although this effect decreases for each speaker, the accessibility of references accepted through verbatim repetition either remains constant or increases for both speakers, thus causing the same references to become accessible to them. Thus, provided that they are given enough time, the speakers’ representations of the common ground become increasingly similar, so that the references that are most accessible to one speaker also become most accessible to the other. From a broader perspective, these findings are in line with the idea that episodic memory traces can be modified, depending on the reuse that is made of this information (see Marsh, 2007).

Importantly, the rationale concerning bias complementarity holds mainly for symmetric tasks such as the one used in the present study. Not all tasks involve symmetric dialogue. For instance, Clark and Wilkes-Gibbs (1986) asked directors to describe tangram figures to matchers so that the latter could rearrange their figures in a predefined order. In such cases, only one of the participants’ initial knowledge is necessary to perform the task. Thus, the speakers’ biases are not complementary; the speaker who does not have initial access to the relevant information needs to make an extra effort to acquire partner-produced information.

In conclusion, these results contribute to a better understanding of how individual memory processes affect the collaborative processes at play during dyadic and triadic dialogue. They show that the accessibility in memory of grounded references depends on how these were initially grounded, and they also show how accessibility might vary as the interaction unfolds.