Conceptual Match as a Determinant of Reference Reuse in Dialogue

Acknowledgement:

Dialogue is a joint activity during which at least two partners collaborate to reach a common goal such as making dinner or working together (Bangerter & Clark, 2003; Clark, 1996). As they interact, speakers produce referring expressions, or references, to designate the objects and entities under discussion (“the living room table,” “the recipe I’m planning on making tonight,” “the boss,” “a timeslot which wouldn’t be very convenient for you,” etc.; e.g., Bard, Hill, Foster, & Arai, 2014; Brennan & Clark, 1996; Clark & Wilkes-Gibbs, 1986; Fukumura, 2015). There are generally several different ways of talking about the same referent. For instance, the same table might be referred to as “the living room table” or “the beautiful wooden table I bought when I first moved into this place.” Speakers thus need to make a number of decisions about how to refer to things throughout the interaction.

In some cases, speakers refer to a referent several times during an interaction. When this happens, speakers reuse the same references, inferring that their partners should be capable of understanding them again (Brennan & Clark, 1996; Metzing & Brennan, 2003). Such inferences require high-level perspective-taking skills. However, reference reuse also depends on lower-level processes such as whether these references are readily accessible in memory to them (e.g., Horton & Gerrig, 2005b; Knutsen & Le Bigot, 2014). Another factor that is likely to play an important role here is the way in which speakers personally view, or conceptualize, the referents under discussion. For instance, a piece of modern art may be viewed as “a weird painting” or as “an enlightened representation of humanity.” How do such different conceptualizations affect reference reuse? The current study seeks to address this question, thus, contributing to a better understanding of referential decisions in dialogue.

According to the collaborative approach to dialogue, decisions about how to refer to things are adaptive: each speaker favors the production of references he or she believes to be easily understandable for his or her current addressee(s) (Brennan & Clark, 1996; Clark & Wilkes-Gibbs, 1986; Fukumura, 2015; Isaacs & Clark, 1987; Nückles, Winter, Wittwer, Herbert, & Hübner, 2006; Yoon & Brown-Schmidt, 2014). For instance, Speaker A and Speaker B are planning an afternoon out. A wants to meet B in front of a shopping center, but B does not know where this center is. In this situation, A could adapt his or her speech by locating the shopping center in relation to another landmark A knows B is capable of locating. For instance, if A and B have already gone to a nearby cinema together, A could refer to the shopping center as “the center which is very close to that cinema we went to.” To determine what their current partners are capable of understanding, speakers rely on their common ground, which includes the knowledge that dialogue partners share and are aware of sharing (Brown-Schmidt, 2012; Clark, 1996; Clark & Marshall, 1978, 1981; Stalnaker, 1978). The common ground includes information relative to the partners’ past shared experiences (e.g., the fact that A and B went to the cinema together in the example above) as well as shared community memberships (e.g., the fact that A and B both belong to a local cinema lovers community).

Past decisions about how to refer are also added to the common ground through a joint contribution process (Clark & Brennan, 1991; Clark & Schaefer, 1989; Clark & Wilkes-Gibbs, 1986). This process is divided into two phases. During the first phase, one of the speakers presents a reference he or she believes to be suited to his or her current addressee’s understanding (e.g., A presents the reference: “the center which is very close to that cinema we went to”). This reference is then accepted by the addressee, with the latter producing some kind of evidence that the reference presented was understood well enough for current purposes (e.g., B accepts this reference by saying: “oh yes, I know where that is”). A reference may be accepted more or less explicitly: the addressee might accept it by nodding his or her head, by saying “okay,” “hm,” by moving on to the next relevant speech turn or by repeating the reference presented. It may also be accepted more or less immediately after the presentation phase, depending on whether the addressee has enough knowledge and information to understand it correctly straight away or whether he or she needs additional information to understand the director’s initial intentions. For instance, rather than accepting A’s reference immediately, B might reply: “I don’t know where that is, I really can’t remember where that cinema was.” In such situations, A might provide more information to B to help him or her to understand the reference initially presented (e.g., “it’s not far away from High Street”). Another possibility is that B might present an alternative reference (e.g., “do you mean the shopping center that’s just in front of the town hall?”). In any event, at some point during this process, A and B will reach mutual understanding: they will converge on a reference that they are both capable of understanding and that B can finally accept.

Once presented and accepted, a reference is deemed part of the partners’ common ground. If either partner needs to refer to the same referent again later, he or she is likely to favor the reuse of the same reference to improve addressee comprehension. For instance, once the reference “the center which is very close to that cinema we went to” (or any alternative reference finally accepted by B) has been added to A and B’s common ground, both A and B can assume that their partner should be capable of understanding this reference successfully if they reuse it again later during the interaction.

The contribution process was illustrated in a study conducted by Clark and Wilkes-Gibbs (1986) in which pairs of participants took part in a matching game (see also Fox Tree & Clark, 2013; Gorman, Gegg-Harrison, Marsh, & Tanenhaus, 2013; Hupet, Seron, & Chantraine, 1991; Krauss & Weinheimer, 1966; Lysander & Horton, 2012; Yoon & Brown-Schmidt, 2014). One of the participants (the director) described abstract pictures (tangram figures; see Figure 1 for an example) to his or her partner (the matcher) to allow the latter to arrange these pictures in a predefined order. The same task was repeated six times. The main result was that the participants consistently used similar references to refer to the same pictures across trials (e.g., “the ice skater” to describe the picture presented in Figure 1), relying on the common ground built during the first trial(s) to infer that the matcher should be capable of understanding these references again in subsequent trials.

Building on these initial findings, Brennan and Clark (1996) suggested that dialogue partners rely on presentation and acceptance to establish conceptual pacts, that is, temporary agreements about how to refer to referents. For instance, in Clark and Wilkes-Gibbs’s (1986) study, the director and the matcher agreed to refer to the picture presented in Figure 1 as “the ice skater”; the director tended to reuse this reference (rather than another, potentially suitable different reference) across trials. Such pacts are defined as temporary because when speakers face new addressees, they must establish new (and potentially different) conceptual pacts. For instance, if the director started a new interaction with a different matcher, he or she might use another reference than “the ice skater” to refer to the picture in Figure 1. Conceptual pacts also affect comprehension, as speakers breaking such pacts (i.e., using a new reference rather than reusing a reference that belongs to the common ground) impairs comprehension for addressees (Metzing & Brennan, 2003; see also Yoon & Brown-Schmidt, 2014).

The findings reviewed in the previous section suggest that speakers tend to reuse the same references when referring repeatedly to the same referent to adapt to their partners. However, there is evidence that not all presented and accepted references are equally likely to be subsequently reused in the remainder of the interaction. To illustrate this point, Knutsen and Le Bigot (2014) ran an experiment in which pairs of participants added references to landmarks featured on a map to their common ground through presentation and acceptance. The main finding was that speakers were more likely to reuse self-presented references than partner-presented ones (i.e., to re-refer to landmarks corresponding to self-presented references more than to landmarks corresponding to partner-presented references). A similar finding was obtained in another study in which the participants had the opportunity to generate references to landmarks rather than simply read them off a map (Knutsen, Ros, & Le Bigot, in press). Based on this finding, the authors have suggested that reference reuse is subject to a self-presentation bias. However, these studies did not focus on decisions about how to refer per se. Indeed, when a speaker reuses a reference, he or she must make two different decisions, the first of which pertaining to whether to mention a referent or not, and the second of which pertaining to how to refer to this referent. To illustrate this distinction, consider the following example. A is discussing a route with B. At some point, A presents the following building to refer to the town hall: “this horrible building in the middle of the town.” B accepts this reference, thus agreeing to establish this conceptual pact and to add this reference to the common ground. Later during the interaction, A has the opportunity to refer to the same building again. At this point, A first needs to decide whether or not to mention this referent again. Such decisions are hereafter referred to as decisions to refer (i.e., decisions about whether to refer or not to a given referent, regardless of how this referent is actually referred to). Actual decisions about how to refer only take place after this, once A has decided to mention the referent. At this point, A must decide whether to maintain the conceptual pact by reusing the reference that was initially agreed upon or to abandon it by using a different reference.

Knutsen and colleagues (Knutsen & Le Bigot, 2014; Knutsen et al., in press) only examined decisions to refer, showing that the self-presentation bias makes speakers more likely to decide to mention a referent again when the reference to this referent was initially self-presented than when it was initially partner-presented. However, a similar self-presentation bias might also constrain actual decisions about how to refer, making speakers more likely to resort to self-initiated conceptual pacts than to partner-initiated conceptual pacts. For instance, if Speaker A refers to a table as “an old piece of furniture” and Speaker B refers to the same table as “a beautiful wooden table,” the self-presentation bias could make A more likely to reuse the reference “an old piece of furniture” and B more likely to reuse the reference “a beautiful wooden table” when referring to this table again. The first purpose of this study is to address this possibility, thereby offering a better understanding of decisions about how to refer in collaborative dialogue.

The second goal of the current study is to offer a better understanding of the processes underlying a potential self-presentation bias in decisions about how to refer. Knutsen and Le Bigot (2014) have suggested that the self-presentation bias in decisions to refer is due to a self-production effect in memory, whereby words produced aloud are remembered better than words read silently, especially when these words are self-produced (as opposed to partner-produced; MacLeod, 2011; MacLeod, Gopie, Hourihan, Neary, & Ozubko, 2010). In cases where speakers generate the references at the time of presentation (Knutsen et al., in press), this effect could also be because of a generation effect in memory (e.g., Burnett & Bodner, 2014; Rosner, Elman, & Shimamura, 2013; Slamecka & Graf, 1978). These suggestions are corroborated by the finding that after the end of an interaction, speakers tend to remember self-generated content better than partner-generated content (Hjelmquist, 1984; Jarvella & Collas, 1974; Knutsen et al., in press). They are also corroborated by the finding that other occurrences of reference production during the interaction (e.g., acceptance through verbatim repetition as well as reuse per se) increase subsequent accessibility in memory (Knutsen & Le Bigot, 2014, 2015). This is in line with the idea that “ordinary” (i.e., nonspecific to dialogue) low-level processes play an important role in reference production in dialogue (Gorman et al., 2013; Horton, 2007, 2008; Horton & Gerrig, 2002, 2005a, 2005b).

Regarding decisions about how to refer, the potential tendency to reuse self-presented references more often could be due not only to memory, but also to additional determinants. In particular, it could be because of the way in which each speaker personally conceptualizes, or “views,” the referents under discussion. For instance, if a speaker refers to the Tangram figure presented in Figure 1 by presenting the reference “the ice skater,” this usually means that the speakers “sees” an ice skater when he or she looks at this figure. More important, dialogue partners might not necessarily conceptualize the same referents in the same way. For instance, when Speaker A refers to the table mentioned previously as “an old piece of furniture” and Speaker B refers to the same table as “a beautiful wooden table,” the two references produced reflect two different conceptualizations of the same referent. It is noteworthy that the notion of “conceptualization” in the current study can be linked with the perspective-taking literature, and more particularly with the notion of Level-2 perspective-taking, which also refers to how people “view” a scene (e.g., Flavell, Everett, Croft, & Flavell, 1981). However, to link the current work to theoretical models of language production (see the General Discussion); the notion of “conceptualization” is preferred here.)

Self-presented references being reused subsequently might, therefore, simply reflect the fact that speakers tend to “view” the same referents in the same way in addition to, or rather than, better memory for self-presented references. One implication would be that in cases where Speaker A presents a reference that happens to match Speaker B’s conceptualization of the referent under discussion, Speaker B could become more likely to reuse a partner-presented reference—a prediction incompatible with the (self-) production/generation effect account (Knutsen & Le Bigot, 2014; Knutsen et al., in press). In other words, there would be no such thing as a single “self-presentation effect” in such cases. Rather, generation and conceptualization would differentially affect reference reuse.

The idea that conceptualization is an important determinant of reference reuse was addressed by Duff, Hengst, Tranel, and Cohen (2006; note that the influence of self- vs. partner-presentation was not assessed in this study). These authors compared decisions about how to refer in participants with amnesia (hippocampal amnesia; see Spiers, Maguire, & Burgess, 2001) and typical participants. The experimental setting was similar to that used by Clark and Wilkes-Gibbs (1986): the participants played a matching game that required referring to abstract Tangram figures. The analysis of the results revealed that just like typical participants, participants with amnesia reused the same references to refer repeatedly to the same referents when they played the role of directors. Such reuse could not be because of the fact that these participants remembered which references had initially been presented, as hippocampal amnesia prevents individuals from encoding new information in episodic memory. This was corroborated by the results of a follow-up test in which tangram figures (that had not been used in Duff et al.’s main experiment) were arbitrarily paired with nouns (e.g., a tangram figure looking like a rabbit was paired with the word “broom”). The participants were asked to learn these pairs. The analysis of their performance on a subsequent memory test revealed that the performance of the participants with amnesia was strongly impaired. On the basis of these results, the authors suggested that reference reuse during the matching task did not reflect the amnesic participants’ memory for references previously presented and accepted, but rather that the participants conceptualized the pictures in the same way across trials, leading them to produce similar descriptions every time they encountered the same referent.

In summary, reference reuse in dialogue could be guided by a production/generation effect (Knutsen & Le Bigot, 2014), a “conceptual match” effect (Duff et al., 2006), or both. In many cases, these two effects are likely to be confounded, as it seems reasonable to assume that speakers often generate references that match their own conceptualizations at the time of initial reference presentation. This makes it particularly difficult to examine the respective contributions of conceptualization and generation to dialogic reference reuse. The current study sought to address this question in a situation where participants made no decisions to refer (i.e., they did not decide which referents to refer to during the task; this was decided in advance by the experimenter), but had the opportunity to make decisions about how to refer as they interacted repeatedly about the same referents.

The general hypothesis tested in this study was that reference reuse in dialogue is guided not only by a generation effect in memory (i.e., speakers reuse self-generated references more often than partner-generated ones), but also by a conceptual match effect (i.e., speakers reuse references that match their conceptualizations more than references that do not match their conceptualizations). In two experiments, pairs of participants added references to abstract Tangram figures to their common ground through presentation and acceptance. In some cases, the participants generated references that matched their own conceptualizations; in other cases, the participants were forced to generate references that did not match their own conceptualizations.

To test the general hypothesis, the study was divided into two steps. The purpose of the first step (Experiment 1) was to confirm that the generation effect affects not only speakers’ memory for which referents were mentioned during an interaction (as shown in previous studies on the self-presentation bias), but also their memory for which references were actually used to refer to these referents—or, in other words, their memory for conceptual pacts (Brennan & Clark, 1996; Metzing & Brennan, 2003). The purpose of the second step (Experiment 2) was to examine how the generation effect and the conceptualization match effect found in Experiment 1 affect reference reuse in dialogue.

Experiment 1 was divided into three phases. During the first phase (Dialogue Phase), pairs of participants were shown various tangram figures one by one. For each figure, each participant was asked to generate a reference and to answer any questions his or her partner might have about his or her reference. Thus, each pair generated two references for each figure discussed.

For each figure, the participants had to generate a different reference; each participant went first on half the trials and second on half the trials. Therefore, in some cases, each participant had the opportunity to present a reference that matched his or her conceptualization of the picture (i.e., when he or she went first or when he or she went second but that the reference presented by his or her partner did not match the participant’s own conceptualization); in other cases, the participant was forced to present a reference that did not match his or her own conceptualization (i.e., when he or she went second and that the reference matching his or her conceptualization had already been presented by his or her partner). In other words, each participant generated references that matched or did not match his or her conceptualizations and heard his or her partner generating references that matched or did not match his or her (i.e., the participant’s) conceptualizations. During the second phase of the experiment (Memory Assessment Phase), the participants were asked to recall (in writing) the two references generated for each tangram figure during the Dialogue Phase. This allowed for examination of whether the participants’ memory for the references generated during the Dialogue Phase was subject to a self-generation bias. Finally, during the third phase of the experiment (Conceptualization Assessment Phase), the participants were asked to write down, for each figure used during the Dialogue Phase, which of the two references generated matched their conceptualization best. Doing so allowed for determination of whether the self-generation bias found in the Memory Assessment Phase held for references that matched the participants’ conceptualizations and also for references that did not.

The main operational hypothesis was that self-generated references are more likely to be recalled than partner-generated ones. The second operational hypothesis was that the self-generation effect affects both references that match the partners’ conceptualization and references that do not match the partners’ conceptualization.

Forty native French speaking students (22 female; mean age 18.75, SD = 1.28) took part in the experiment for course credit. The participants signed an informed consent form before the beginning of the experiment and were fully debriefed after the experiment.

The dialogues between the participants were recorded using a TASCAM DR-40 double-entry digital recorder connected to two lapel microphones (one per participant).

One hundred randomly selected tangram figures were used in this experiment (see Figure 2). These figures came from various tangram sets. Each tangram figure was randomly assigned a number between 1 and 100; the figures were then randomly divided into five groups of 20 figures labeled A, B, C, D, and E (i.e., Group A was made of Figures 1 through 20, Group B was made of Figures 21 through 40, etc.; the 20 figures discussed by a pair of participants during the Dialogue Phase all belonged to the same group). Each figure was printed on a separate A6 sheet of paper for use during the Dialogue Phase.

Paper booklets were then prepared for each participant to use during the Memory Assessment Phase. Each booklet included 20 target tangram figures (i.e., the 20 figures discussed during the Dialogue Phase) as well as 20 distractor tangram figures that belonged to another group of figures (e.g., when the target figures belonged to Group A, the distractors belonged to Group B). Two different versions of these booklets were created (the random order in which the figures were presented in the booklet was different in each booklet). Two lines were printed in front of each figure so that the participant could write down the two references generated during the Dialogue Phase.

Finally, the same pictures as in the Dialogue Phase as well as blank lined A5 sheet of paper were used in the Conceptualization Assessment Phase.

The experiment was performed by pairs of participants who sat next to each other and facing the experimenter in a quiet experimental room. Before the beginning of the experiment, the participants were informed that the study sought to investigate referential communication and that the experimenter was simply interested in how speakers refer to abstract pictures such as tangram figures. They were also told that the experiment would involve more than one phase but were not informed in advance of how many phases there would be or of the content of the different phases.

The Dialogue Phase was divided into 20 trials (see Table 1 for an example of a dialogue between two participants during a trial; in this example, the participants were describing the figure presented in Figure 3). At the beginning of each trial, the experimenter showed the participants a tangram figure (both participants were shown the same picture; because the participants sat next to each other, they both viewed the figure from the same perspective). Two random presentation orders were used in this phase. One of the participants (hereafter referred to as Participant X) generated a reference to designate the figure. The participants were then free to discuss Participant X’s decision: for instance, Participant Y could ask for additional information or manifest his or her understanding by saying “okay.” For instance, in Table 1, Participant X started by presenting the reference “someone on a boat.” This presentation was followed by a brief discussion during which X provided Y with additional information (“you have the boat here and there you have the guy”) and during which Y manifested his or her understanding (“yes,” “I see”). Y could also have used this discussion to ask X for additional information (e.g., “I don’t understand, if it’s a boat, where is the hull?”) or express a judgment about X’s reference (e.g., “I don’t agree at all”). Once Participants X and Y had finished discussing Participant X’s decision, Participant Y proposed a different reference to the same figure and then briefly discussed his or her decision with Participant X. The first reference was presented by Participant X in 50% of the trials and by Participant Y in the other 50% of trials. At the beginning of each trial, the experimenter prompted the participant whose turn it was to present the first reference to start talking. This participant had as much time as necessary before he or she actually started talking.

Then, during the Memory Assessment Phase, each participant performed an individual memory test. To this end, each participant was given a booklet and a pen. For each tangram figure shown in the booklet, the participant was asked to do two things: (a) to decide whether or not it had been shown during the Dialogue Phase and (b) if this was the case, to attempt to recall all of the words presented by each participant to describe this figure. The experimenter stayed in the room with the participants to make sure that they did not attempt to communicate with each other during this individual phase.

Finally, the participants’ conceptualizations were assessed during the Conceptualization Assessment Phase. During this phase, the participants were shown again the 20 tangram figures they had been shown during the Dialogue Phase. These were presented one by one in the same order as in the Dialogue Phase; this was to make sure that the amount of time elapsed between the presentation of a figure during the Dialogue Phase and the presentation of the same figure during the Conceptualization Assessment Phase was roughly the same for all figures. For each figure, each participant was first asked to say aloud the reference that he or she had initially presented during the Dialogue Phase. If one of the participants could not remember the reference that he or she had presented during the Dialogue Phase, his or her partner was allowed to help him or her by producing the reference him- or herself if he or she could remember it (this happened in less than 2% of trials, suggesting that the participants remembered well the references they had generated during the Dialogue Phase). The participants were allowed to interact at this point. The purpose of this was to make sure that both references were readily accessible in memory to both participants. Each participant’s conceptualization was then assessed by asking him or her to write down which of these two references reflected his or her point of view better (the participants were instructed not to interact as they wrote down their answer and not to tell their partner which reference they had selected). The participants were specifically required to choose between the two references initially generated during the Dialogue Phase: they could not use a new reference during this phase.

The participants performed the three phases of the experiment at their own pace. There was no break between the three phases. The experiment lasted approximately 1 hr. The participants were fully debriefed after the end of the Conceptualization Assessment Phase. The three steps of the experiment are summarized in Table 2.

Coding—generation

The interactions between the participants during the Dialogue Phase were transcribed and the content words (proper and common nouns, adjectives, and verbs) presented to refer to the tangram figures were identified. Content words included proper nouns (e.g., “San Francisco”), common nouns (e.g., “cat”), adjectives (e.g., “tall”), and verbs (e.g., “to eat”). Auxiliary verbs (i.e., “to be” and “to have”), modal verbs (e.g., “can,” “must”), determiners (e.g., “the,” “a,” and “one”), pronouns (“I,” “this”), adverbs (e.g., “often”), prepositions (e.g., “after,” “despite”), coordination conjunctions (e.g., “but,” “and”), disfluencies (e.g., “uh,” “hm”), and interjections (e.g., “phew,” “oh”) are not content words and were therefore not identified in the corpus. In addition, only the content words used to describe the tangram figures were taken into account here: the content words referring to the participants’ perception of the figures (e.g., “I see . . .”) were not taken into account. For instance, in the example in Table 1, the content words “boat,” “guy,” and “hat” were coded as presented, but the words “well” (interjection), “uh” (disfluency), or “the” (determiner) were not coded, as they are not content words. In a similar way, the verb “to see” was not coded as presented, as it referred to Participant X’s perception of the figure. Only the first mention of each content word (for each figure presented) was taken into account.

In some cases, one of the participants would present a reference and the other participant would complete this description. For instance, Participant X would say “I see a man on a boat” and Participant Y would say “yes, he seems very happy.” In such cases, the extra content words presented by the other participant (in this example, “happy”) were not taken into account. Indeed, the participants’ task was to generate two different references for the figures they were shown (and not to complete each other’s descriptions); therefore, the coding only sought to identify which content words were produced by Participant X in his or her descriptions on one hand and which content words were produced by Participant Y in his or her descriptions on the other. This led us to remove 424 content words from the dataset.

Each content word was then coded depending on whether it had been self-generated or partner-generated from each participant’s point of view. For instance, in the example shown in Table 1, the content words “boat” and “guy” were coded as self-generated from X’s point of view and partner-generated from Y’s point of view; as for the content word “hat,” it was coded as self-generated from Y’s point of view and partner-presented from X’s point of view. This level of coding was used as one of the predictors in the main statistical analysis.

Coding—conceptualization

The participants’ responses during the Conceptualization Assessment Phase were examined to determine, for each tangram figure, which reference presented matched each participant’s conceptualization. This allowed for determination of whether the references generated by each participant did or did not match his or her conceptualization. For instance, if both Participant X and Participant Y indicated that they viewed the tangram figure described in Table 1 as “a boat” during the Conceptualization Assessment Phase, this word was coded as matching one’s conceptualization best from both participants’ point of view. All of the other content words presented to describe this picture (i.e., “guy,” which was initially presented by Participant X, and “hat,” which was initially presented by Participant Y) were coded as not matching one’s conceptualization from X’s point of view and from Y’s point of view. This level of coding was used as one of the predictors in the main statistical analysis.

Coding—recall

The participants’ memory for the references presented during the Dialogue Phase was assessed by examining their performance during the Memory Assessment Phase. For each participant and each tangram figure, each content word presented during the dialogue phase was coded either as recalled (Code 1) or nonrecalled (Code 0). This was a binary variable; however, this variable also reflected the proportion of content words recalled. For instance, if Participant X presented two content words and Participant Y presented three content words during the Dialogue Phase and that Participant X recalled four of these content words, this resulted in four out of five content words being coded 1 for recall, which corresponded to an average recall proportion of 0.8. This binary level of coding was used as the dependent variable in the main statistical analysis.

Independent variables

There were two independent variables in this experiment. The first one was Generation. From each participant’s point of view, each content word presented was either self-generated or partner-generated. The second one was Conceptualization. From each participant’s point of view, each content word presented either match or did not match his or her own conceptualization. Both variables were within-participants.

Descriptive statistics—Dialogue Phase

During the Dialogue Phase, the average number of speech turns produced per dyad was 132.70 (SD = 50.81) and the average number of words produced per dyad (regardless of whether these were content or noncontent words) was 1307.65 (SD = 555.23).

The total number of content words presented by the participants to describe the tangram figures was 3,241 (note that this figure does not include the content words presented by one participant to complement a description initiated by the other participant). The average number of content words produced per tangram figure per participant was 4.05 (SD = 2.63).

One of the participants asked his or her partner for more information after a reference had been generated in 18.50% of trials (74 trials out of 20 pairs × 20 figures = 400 trials). Thus, clarification requests were not systematic; nonetheless, this confirms that the participants viewed this phase as interactive and felt that they could ask for additional information when necessary.

Descriptive statistics—Memory Assessment Phase

During the Memory Assessment Phase, fillers were incorrectly identified as having been discussed during the Dialogue Phase in 1.38% of cases (11 incorrect identifications out of 20 fillers × 40 participants = 800 occurrences). Target figures were correctly identified as having been discussed during the Dialogue Phase in 91.88% of cases (735 correct identifications out of 20 target figures × 40 participants = 800 occurrences). This suggests that the participants remembered well which referents had been discussed during the Dialogue Phase.

In total, 1,800 content words that had been discussed during the Dialogue Phase were recalled during the Memory Assessment Phase (the content words corresponding to tangram figures that had not been discussed during the Dialogue Phase and the content words that had not been produced during the Dialogue Phase, although they were recalled in association with a figure that had been discussed during this phase, were discarded from further analysis, as the hypotheses solely concerned previously produced references to known tangram figures). The average number of content words correctly recalled per tangram figure per participant was 2.25 (SD = 1.30).

Descriptive statistics—Conceptualization Assessment Phase

During this phase, pairs of participants were able to recall the two references initially presented in 98.63% of cases (789 occurrences out of 20 pictures × 40 participants = 800 occurrences). The trials corresponding to the remaining 1.37% (11 trials) were discarded from further analysis, as these corresponded to situations in which the two references were not necessarily readily available for the two participants to choose from. In a similar way, trials in which the participants’ responses included both self- and partner-generated words (8.88% of cases: 71/800 occurrences) were also excluded from further analysis, as these reflected cases in which the participants had no clear preference as to how to conceptualize the referents. The remaining dataset included data from 718 trials, representing a total of 2,895 presented words.

The participants reported that the first reference presented matched their conceptualization best in 60.86% of cases (437/718 occurrences). This confirms that the first reference presented during the Dialogue Phase sometimes matched the participants’ conceptualization even when this reference was not self-generated, as the participants only generated the first reference in 50% of trials. They also reported that the reference that they had generated themselves matched their conceptualization best in 60.31% of cases (433/718 occurrences). This pattern of results is summarized in Table 3.

Main analysis: Influence of generation and conceptual match on reference recall during the Memory Assessment Phase

A logistic mixed model was used to analyze the data (for logistic analysis, see Agresti, 2002; Jaeger, 2008; for mixed models, see Barr, Levy, Scheepers, & Tily, 2013).

Mixed models allow for the inclusion of random intercepts (that are used to account for potential variability across dyads, across participants, and across items) and for the inclusion of random slopes (that are used to account for the fact that different dyads, different participants and different items might differ in their sensitivity to the fixed effects included in the model). Mixed models are also used to account for the nesting of participants in larger groups such as dyads (see McMahon, Pouget, & Tortu, 2006). The current analysis was conducted using SAS 9.4 (GLIMMIX procedure). In line with Barr et al.’s (2013) recommendations, the maximal random structure justified by the experimental design (i.e., all random intercepts and all random slopes corresponding to within-unit IVs) was initially implemented. Three levels of random effects were distinguished in this analysis: dyads, participants, and tangram figures (items). Then, in line with Kiernan, Tao, and Gibbs (2012), the random effects causing G-matrix convergence failure were identified and removed from the model. The identification of problematic random effects is performed automatically in SAS. Removing these random effects from the model has no effect on the outcome of the analysis (i.e., the model parameters remain unchanged; Kiernan et al., 2012; Knutsen & Le Bigot, 2014). The results reported hereafter correspond to the final model (i.e., the model in which the random effects causing convergence problems were not included).

As for logistic mixed models, they are used in situations where the outcome of the analysis is binary (e.g., in this experiment, any content word was either coded as recalled or nonrecalled). One of the parameters returned by logistic models is the odd ratio (OR), which compares the odds associated with two different events (Jaccard, 2001). For instance, the current study compared the odds of recalling self-generated content words and the odds of recalling partner-generated content words during the Memory Assessment Phase. If the latter were used as the baseline in the analysis and that an OR of 1.5 was found, this would mean that 1.5 self-generated content words were recalled for every one partner-generated content word. The corresponding 95% confidence interval (CI) is also usually reported. ORs are also informative with regard to effect size: larger odd ratios reflect larger effects (Agresti, 2002).

The number of content words presented in each cell of the design varied across cells, making it difficult to assess the degrees of freedom should be used to determine whether or not the effects involved in this analysis were statistically significant. In such cases, the Satterthwaite approximation may be used to correct the degrees of freedom, which is what was done in the current study (Keselman, Algina, Kowalchuk, & Wolfinger, 1999; Satterthwaite, 1946).

The model used to analyze the data included Generation (self, other), Conceptualization (match, mismatch) and the interaction between these two factors as fixed effects. The outcome variable was the likelihood of recalling a content word during the Memory Assessment Phase. The random effect structure included (a) by-dyad random intercepts and by-dyad random slopes corresponding to Conceptualization; (b) by-participant random intercepts and by-participant random slopes corresponding to Generation; and (c) by-item random intercepts and by-item random slopes corresponding to Conceptualization. All other random effects (i.e., by-dyad random slopes corresponding to Generation, by-participant random slopes corresponding to Conceptualization and by-item random slopes corresponding to Generation) were removed from the analysis because they caused model convergence failure. Removing these random effects had no influence on the outcome of the analysis.

The results are reported in Figure 4. Generation significantly predicted recall, F(1, 67) = 9.59, p = .003. The participants were more likely to recall self-generated content words than partner-generated ones, OR = 1.43, 95% CI [1.14, 1.80]. Conceptualization also significantly predicted recall, F(1, 22) = 350.09, p < .001. The participants were more likely to recall content words that matched their conceptualization than content words that did not match their conceptualization, OR = 16.34, 95% CI [11.99, 22.28]. Furthermore, the interaction between these two factors was statistically significant, F(1, 5793) = 19.67, p < .001. Additional pairwise comparisons (Sequential Bonferroni) were conducted to offer a better understanding of this pattern of results. These comparisons revealed that the difference between self- and partner-generated references was significant when these references matched the participants’ conceptualization (p < .001) but not when these references did not match the participants’ conceptualizations (p = 1.00). The model parameters are reported in Appendix A (Table A1).

As mentioned previously, generation at the time of presentation is not the sole known linguistic determinant of reference accessibility in memory in dialogue. Reference reuse also depends on how these references were initially accepted: references accepted through verbatim repetition are more likely to be reused in the remainder of the interaction (Knutsen & Le Bigot, 2012, 2014, 2015; Knutsen et al., in press). Furthermore, references reused during an interaction are remembered better after the end of the interaction than references that were not reused (Knutsen & Le Bigot, 2014, 2015; Knutsen et al., in press). In the current study, this kind of reused is labeled “short-term reuse” to distinguish it from longer-term reuse (i.e., reuse occurring in subsequent interactions between the same partners, such as reuse during the Matching Phase in Experiment 2).

These two determinants of reference reuse were not of prime interest in the current study; furthermore, the number of acceptances and short-term reuses in this study was too small to include these as fixed effects in the model. Nonetheless, acceptance and short-term reuse could have affected the participants’ memory performance during the Memory Assessment Phase. To discard this eventuality, the participants’ interactions during the Dialogue Phase were coded for acceptance and short-term reuse (i.e., reuse during the Dialogue Phase); a statistical analysis was then performed to determine whether the effects of Generation and Conceptualization found in the main analysis remained significant when Acceptance and Reuse were controlled for. This analysis is reported in Appendix B: it confirmed that the effects of Generation and Conceptualization on recall remained significant even when these additional determinants were taken into account.

The purpose of Experiment 1 was to determine whether speakers’ memory for decisions about how to refer is subject to a generation effect (Slamecka & Graf, 1978). The results corroborated this idea, as the participants recalled the content words they had generated themselves better than the content words that had been generated by their partner. This confirms and extends previous work on how production and generation affect collaborative dialogue (Knutsen & Le Bigot, 2014; Knutsen et al., in press), shedding light on how low-level processes affect speakers’ memory for past interactions (see also Knutsen & Le Bigot, 2015).

Conceptual match was also taken into account in Experiment 1. The results revealed that content words were more likely to be recalled when they matched the participants’ conceptualizations than when they did not. However, this effect did not necessarily reflect better memory for these words. Indeed, one possibility is that the participants perceived the figures they were shown during this phase in the same way as in the Dialogue Phase, thus causing them to use the same words to describe them again in an individual task (see Duff et al., 2006).

Moreover, the interaction between Generation and Conceptualization revealed that the self-generation effect was only statistically significant when the references recalled matched one’s own conceptualizations. This does not necessarily imply that references that do not match one’s own conceptualizations are not subject to a generation effect as well. Indeed, no conclusions can be drawn from the lack of a significant difference between self- and partner-generated references in this condition. However, this suggests that the generation effect is attenuated for references that do not match one’s own conceptualizations. This could be because of a floor effect in this experiment: the participants might have had difficulty recalling the references that did not match their own conceptualizations, regardless of who had generated them. In any event, the significant interaction also reflects the fact that the participants’ tendency to recall better references that matched their own conceptualizations was stronger when these references had initially been self-generated (rather than partner-generated). Thus, references that benefit from both a generation effect and a conceptual match effect are remembered better than any other references.

In summary, Experiment 1 suggests that past decisions about how to refer are subject to a generation effect—at least in cases where these references match one’s conceptualizations of the referents under discussion. The purpose of Experiment 2 was to build on this initial finding by examining how these two factors affect actual reference reuse in dialogue. To this end, in Experiment 2, the initial Dialogue Phase was followed by a matching game instead of a memory test. This game gave the participants the opportunity to reuse the references presented during the Dialogue Phase, allowing us to examine how generation and conceptualization affect reference reuse.

Just like Experiment 1, Experiment 2 was divided into three phases. The first phase was identical to the Dialogue Phase in Experiment 1. Then, the participants embarked on a Matching Phase during which they took turns describing tangram figures to each other (in target trials, these figures had already been discussed during the Dialogue Phase). Finally, the participants’ conceptualizations were assessed in a third phase, which was identical to the Conceptualization Assessment Phase in Experiment 1.

Two operational hypotheses about reference reuse during the Matching Phase were formulated. The first hypothesis was that if reuse during the Matching Phase is mainly guided by generation (as suggested by Knutsen & Le Bigot, 2014), the participants will reuse self-generated references more often than partner-generated ones regardless of whether these references match their conceptualization or not. By contrast, the second hypothesis was that if reuse during the Matching Phase depends both on generation and conceptual match, the participants will reuse self-generated references more often than partner-generated references and references that match their conceptualizations more often than references that do not.

Forty participants (35 female; mean age 18.40, SD = 0.93) were recruited under the same conditions as in Experiment 1 to take part in Experiment 2.

The same digital recorders as in Experiment 1 were used to record the interactions between the two participants in the first two phases of the experiment.

The same tangram figures as in Experiment 1 were used in Experiment 2. As in Experiment 1, each figure was printed on a separate A6 sheet of paper for use during the Dialogue Phase.

Paper booklets were prepared for use during the Matching Phase (see Figure 5). Each booklet was 40 pages long. Four different tangram figures were printed on each page. One of these figures was one of the figures discussed during the Dialogue Phase (each of the figures discussed during the Dialogue Phase appeared twice in each booklet) and the other three belonged to a different group (e.g., the booklets prepared for the dyads who initially discussed the figures from Group A also included the figures from Groups B, C, and D). The position of the figure discussed during the Dialogue Phase among the other three figures was randomized. Two random page presentation orders were used to create these booklets. Five booklets were created depending on which group of figures (A, B, C, D, or E) had been discussed during the Dialogue Phase.

Two complementary versions of each of these five booklets were then created (see Figure 5). The two versions of the same booklet included the same tangram figures printed in the same order. In the first version, on even-numbered pages (20 in total), the target figure was encircled by a red square; on odd-numbered pages, the target figure was not encircled (20 in total). In the second version of the same booklet, target figures were encircled on odd-numbered pages. In other words, when the target figure was encircled in one of the booklets, it was not encircled in the other booklet. In 50% of cases, the target figure was the figure discussed during the Dialogue Phase; in the other 50% of cases, the target figure belonged to another group. Pairs of complementary booklets were elaborated so that each figure discussed during the Dialogue Phase served as target once and as distractor once. The purpose of this was to prevent the participants from inferring that when the array of figures included one of the figures discussed during the Dialogue Phase, the figure was necessarily the target during the Matching Phase.

The materials used during the third phase were identical to those used during the Conceptualization Assessment Phase in Experiment 1.

As in Experiment 1, the participants were informed that the experiment would be divided into several phases, but they were not informed in advance of what they would be asked to do during each of these phases. The Dialogue Phase in this experiment was identical to the Dialogue Phase in Experiment 1 (see Table 4 for a dialogue sample in this phase; the figure discussed in this sample can be found in Figure 6).

The Matching Phase was divided into 40 trials, each corresponding to a different page of the participants’ booklets. In each trial, the task of the participant playing the role of director was to describe the tangram figure encircled by a red square to the participant playing the role of matcher so that the latter could find this figure among the four represented on his or her booklet and give the director the corresponding number (1, 2, 3, or 4, starting from the left). The trial ended after the director had confirmed that the matcher’s answer was correct. The participants switched roles (Director and Matcher) after each trial, implying that each participant had the opportunity to describe 10 target figures and 10 filler figures to his or her partner during this phase (see Table 5 for two dialogue samples from this phase).

Finally, the Conceptualization Assessment Phase in this experiment was identical to the Conceptualization Assessment Phase in Experiment 1. The participants performed the three phases of the experiment at their own pace and were fully debriefed after the end of the Conceptualization Assessment Phase. The experiment lasted approximately 1 hr. The three steps of the experiment are summarized in Table 6.

Coding—generation and conceptualization

The data from the Dialogue Phase and the Conceptualization Assessment Phase were coded for generation and conceptualization in the same way as in Experiment 1. There were 532 words removed from the dataset because of having been presented by one of the participants to complement the other participant’s description during the Dialogue Phase.

Coding—reuse during the Matching Phase

The content words produced by the Director during the Matching Phase were identified as in Experiment 1. For instance, in the examples given in Table 5, Director X produced the content words “boar” and “warthog” (Upper Panel) and Director Y produced the content words “UFO,” “diamond,” and “square” (Lower Panel). The list of content words was then used to code each word generated during the Dialogue Phase as reused (Code 1) or nonreused (Code 0). For instance, in the example provided in Tables 4 and 5, the content word “warthog,” which was generated by Participant X during the Dialogue Phase, would have been coded as reused by this participant during the Matching Phase. All other content words initially generated during the Dialogue Phase (i.e., “boat” and “anchor”) would have been coded as nonreused. Just like recall in Experiment 1, reuse was a binary variable. In addition to this, the content words produced in fillers trials were not coded for reuse, as the figures discussed during these trials had not been discussed during the Dialogue Phase. This level of coding was used as the dependent variable in the main statistical analysis.

Independent variables

Just like Experiment 1, Experiment 2 involved two within-participants independent variables, Generation and Conceptualization. These were coded in the same way as in Experiment 1.

Descriptive statistics—Dialogue Phase

During the Dialogue Phase, the average number of speech turns per dyad was 158.60 (SD = 69.11) and the average number of words produced by dyad (regardless of whether these were content or noncontent words) was 1434.95 (SD = 649.84). The total number of content words presented by the participants was 3406 (note that this figure does not include the content words presented by one participant to complement a description initiated by the other participant). The average number of content words produced per tangram figure per participant was 4.26 (SD = 2.85). One of the participants asked his or her partner for more information after a reference had been generated in 14.50% of trials (58 trials out of 20 pairs × 20 figures = 400 trials).

Descriptive statistics—Matching Phase

During the Matching Phase, the average number of speech turns per dyad was 156.10 (SD = 39.90) and the average number of words (including content and noncontent words) produced per dyad was 861.60 (SD = 373.12). In total, 562 (nonnecessarily unique, as part of the words presented within a dyad might have also been presented within one or several other dyads) content words that had been presented during the Dialogue Phase were reused by the Directors on target trials (i.e., trials in which the figure to describe had previously been discussed during the Dialogue phase; the data corresponding to nontarget trials were not analyzed further, as no hypothesis was formulated concerning these trials). The average number of reused content words per tangram figure per director was 1.41 (SD = 0.98). Matchers managed to find the target figure in their first attempt in 96.25% of trials (770 trials out of 20 dyads × 40 trials = 800 trials in total), suggesting that the task was relatively easy for the participants.

Descriptive statistics—Conceptualization Assessment Phase

The trials in which the participants were not able to recall the two references initially presented and the trials in which the participants’ responses included both self- and partner-generated words (15.38% of cases: 123 out of 800 occurrences) were excluded from further analysis. The remaining dataset thus included data from 677 trials, which represented a total of 2,945 presented words.

The participants reported that the first reference presented matched their conceptualization best in 59.68% of cases (404/677 occurrences). They also reported that the reference that they had generated themselves matched their conceptualization best in 63.52% of cases (430/677 occurrences). This pattern of results is summarized in Table 7.

This pattern of results is informative with regard to a potential confound in this study. Indeed, one might suggest that giving the participants the opportunity to reuse the references presented in the Dialogue Phase during the Matching Phase (and to assess the efficiency of such reuse on their partners’ comprehension) might have biased the participants’ responses during the Conceptualization Assessment Phase. For instance, they might have revised their initial preferences by answering that whichever references allowed them to successfully complete trials during the Matching Phase matched their conceptualizations best. The pattern of results reported in Table 2 being so similar to that obtained in Experiment 1 (during which the participants did not perform a matching task) seems to go against this idea. An additional analysis was conducted to confirm that the participants’ responses were not significantly different across both experiments. No differences were found (all Fs < 1). This point is addressed further in the General Discussion.

Main analysis: Influence of generation and conceptual match on reference reuse during the Matching Phase

The data were analyzed following the same rationale as in Experiment 1, except that the dependent variable was reuse during the Matching Phase rather than recall. The model used to analyze the data included Generation (self, other) and Conceptualization (match, mismatch) as fixed effects. The interaction between these two factors was removed from the model, as it failed to reach statistical significance, F(1, 139) = 0.01, p = .914. The outcome variable was the likelihood of reusing a content word during the Matching Phase. The random effect structure included (a) by-dyad random slopes corresponding to Generation and Conceptualization; (b) by-participant random slopes corresponding to Generation and Conceptualization; and (c) by-item random slopes corresponding to Generation and Conceptualization. All other random effects (i.e., by-dyad random intercepts, by-participant random intercepts and by-item random intercepts) were removed from the analysis because they caused convergence failure. Removing these random effects had no influence on the outcome of the analysis.

The results are reported in Figure 7. Conceptualization significantly predicted reuse, F(1, 29) = 114.57, p < .001. The participants were more likely to reuse content words that matched their conceptualization than content words that did not, OR = 15.82, 95% CI [9.34, 26.82]. However, the effect of Generation on reuse failed to reach statistical significance, F(1, 30) = 2.90, p = .099. The model parameters are reported in Appendix A (Table A2). An additional statistical analysis confirmed that the same pattern of results was found even when Acceptance and Short-term reuse were taken into account (see Appendix B).

The purpose of Experiment 2 was to examine how generation and conceptualization affect reference reuse to offer a better understanding of dialogic decisions about how to refer (Knutsen & Le Bigot, 2014; Knutsen et al., in press). The results confirmed that reference reuse during the Matching Phase was guided by conceptualization: directors were more likely to reuse a reference when it matched their conceptualization than when it did not. This is in line with the idea that the reuse of references from the common ground might reflect not only their memory of what was said previously, but also the speakers’ tendency to perceive referents in the same way each time they are exposed to them. This replicates Duff et al.’s (2006) work and extends their findings to typical individuals. Unexpectedly, however, the effect of generation on reuse failed to reach statistical significance in this experiment, preventing us from concluding that references that are more readily accessible in memory (as shown in Experiment 1) are also more likely to be reused. This is incompatible with the idea that reference reuse in dialogue is systematically subject to a self-presentation bias, as it implies that the conceptualization match effect sometimes caused participants to favor the reuse of partner-presented references. More important, we cannot discard the possibility that the lack of a significant generation effect could be due at least in part to directors trying to adapt their descriptions by favoring the reuse of references that had originally been generated by their partner (i.e., the matcher). The data do not support the idea that this kind of behavior was systematic across participants and/or across items, but it might have been frequent enough to “cancel” any potential self-generation bias in this study.

The results of Experiment 1 and 2 contribute to a better understanding of the low-level memory and conceptual processes at play during dialogue. First, the results of Experiment 1 suggest that at least some decisions about how to refer (namely, decisions that match one’s conceptualizations) are subject to a generation effect: self-generated content words are remembered better than partner-generated ones (Burnett & Bodner, 2014; Rosner et al., 2013; Slamecka & Graf, 1978). This has important theoretical consequences. Indeed, one of the central ideas developed within the framework of the collaborative approach to dialogue is that all partners involved in an interaction come to share a similar representation of their common ground (Clark, 1996; Clark & Marshall, 1981). The current results nuance this assumption by suggesting that dialogue partners may share a representation of their common ground, but that they are subject to a bias in storing and/or retrieving common ground information. In particular, the generation effect implies that references that are readily accessible to A are not necessarily readily accessible to B, and vice versa.

We have previously suggested that this pattern of results (i.e., each speaker remembering the information they presented better than the information presented by others after the end of an interaction) has direct implications for partner-adaptation in subsequent interactions (Knutsen & Le Bigot, 2015). Indeed, part of the common ground consists in the information mentioned during past interactions between the same dialogue partners (Brennan & Clark, 1996; Clark & Marshall, 1981; Metzing & Brennan, 2003). Such bias in conversational memory could, therefore, lead each speaker to favor the reuse of self-presented references for subsequent adaptation purposes. However, the results of Experiment 2 do not support this suggestion. Although self-generated references were more readily accessible in memory than partner-generated ones, reference reuse in this experiment was mainly guided by conceptualization. Such conceptual match effect affected both self- and partner-generated references. This implies that in a situation where Speaker A hears Speaker B generating a reference that matches his or her (i.e., A’s) own perception of the topic under discussion, A is capable of reusing this information later during the interaction—despite it being less readily accessible in memory than self-generated information.

A potential confound in this study was addressed by comparing the participants’ responses during the Conceptualization Assessment Phase in Experiment 1 and Experiment 2. The main concern here was that having the opportunity to discuss the figures again in Experiment 2 (i.e., during the Matching Phase) might have caused the participants to reconceptualize at least some of the stimuli used. For instance, at the end of the Dialogue Phase, Participant X might have believed that the reference “the cat” was most appropriate to refer to one of the figures, but realizing that Participant Y had difficulty understanding this reference during the Matching Phase or hearing Participant Y using the alternative reference (e.g., “the fox”) during this phase might have altered Participant X’s response during the Conceptualization Assessment Phase. However, the lack of a significant difference in the participants’ responses in Experiment 1 (where the tangram figures were not discussed again after the Dialogue Phase) and Experiment 2 (where they were discussed again) is not in support of this possibility. However, another possibility that cannot be discarded is that the participants’ conceptualizations might have changed during the Dialogue Phase. For instance, when shown a figure during the Dialogue Phase, Participant X might have conceptualized it as a cat, thus leading him or her to generate the reference “the cat,” but subsequently hearing Participant Y generating the reference “the fox” during the same phase might have led Participant X to reconceptualize this referent and to now view it as a fox as well. This would not be a major issue for the current study, as this new conceptualization (that would have potentially affected Participant X’s behavior in Phase 2) would have been captured during the Conceptualization Assessment Phase (i.e., Participant X would have selected “the fox” rather than “the cat” during this phase). Nonetheless, it would be interesting to examine whether initial conceptualizations and reconceptualizations affect reference reuse in the same way. The current study was not designed to address this question, but this point should be addressed in future research.

In summary, the main implication of this study is that the notion of “self-presentation bias” should be used with caution, as it seems that that such bias could reflect conceptual match in addition to, or instead of, self-generation. As suggested in the Introduction, it seems reasonable to assume that these two determinants are often confounded in everyday conversation (and potentially in many laboratory experiments on spontaneous dialogue): speakers tend to generate references that match their conceptualization at the time of initial presentation. However, in cases where a self-presented reference does not match one’s conceptualization of the referent under discussion, reuse depends more on conceptual match than on generation. Thus, the notions of “self-generation bias” and “conceptual match bias” should be preferred to the single notion of “self-presentation bias” to account more precisely for reuse biases in dialogue.

The results from the current study as well as previous findings (Knutsen & Le Bigot, 2014; Knutsen et al., in press) suggest that memory processes and conceptual processes affect dialogic reference reuse. The idea that different kinds of constraints may affect language production and comprehension is not new in dialogue research (see Hanna, Tanenhaus, & Trueswell, 2003; Kronmüller & Barr, 2015). Precisely, many of the studies focusing on this idea have sought to investigate the time course by which these constraints influence language processing. The current study was not designed to answer the question of when accessibility in memory and conceptual match affect reference production (and whether or not they do so simultaneously), but suggestions can nonetheless be formulated at this point.

Most theoretical models of language production involve a conceptual preparation phase (during which the information regarding which ideas a speaker intends to express and which perspective he or she will use to express these is retrieved) followed by a lexical selection phase (during which the corresponding lexical representations are retrieved; e.g., Dell, 1986; Levelt, 1989; Levelt, Roelofs, & Meyer, 1999). Our suggestion is that the conceptual preparation phase is subject to self-generation bias as well as a conceptual match bias. Indeed, decisions to refer and decisions about how to refer are made during this phase. Decisions to refer were not investigated in the current study, but there is evidence that such decisions are subject to a self-generation bias (Knutsen et al., in press; see also Knutsen & Le Bigot, 2014) that makes speakers more likely to re-refer to referents corresponding to self-presented references than to re-refer to referents corresponding to partner-presented references.

As for the conceptual match bias, the “old piece of furniture/beautiful wooden table” example discussed in the Introduction can be used to illustrate our point. In this example, Speaker A refers to a table as “an old piece of furniture” and Speaker B refers to the same table as “a beautiful wooden table” (in this example, these two references reflect each participant’s conceptualization of the referent). Later during the interaction, Speaker A intends to refer to the same table again. To do this, A must first go through the conceptual preparation phase. At this point, A has the choice between at least two different conceptualizations (i.e., the two conceptualizations corresponding to the two references); the conceptual match effect should cause him or her to select the “old piece of furniture” conceptualization. The activation then spreads to the corresponding node at the lexical level, eventually leading Speaker A to reuse the reference “the old piece of furniture” again.

It is noteworthy that in this example, the preferred conceptualization is associated with self-generated words. However, imagine a situation in which this is not the case. For instance, both Speaker A and Speaker B conceptualize the table as an old piece of furniture, and B presents the corresponding reference first. Later during the interaction, A intends to refer to the same table again. In this kind of situation, activation should spread from the “old piece of furniture” conceptualization to the corresponding reference, thus facilitating the production of an initially partner-generated reference. This could also help explain why the generation effect did not significantly affect reuse in Experiment 2: the activation of preferred conceptualizations might have led to the activation of the corresponding references regardless of who had initially generated them.

Although conceptual match plays an important role in dialogic reference production, there are nonetheless a number of situations in which speakers are capable of producing references that do not match their own conceptualizations. This is especially apparent in studies in which experts interact with novices. For instance, computer experts are capable of adapting their speech depending on the level of knowledge of the partners they interact with (Nückles et al., 2006). In such situations, reference production mainly reflects the speakers’ beliefs regarding their partners’ level of knowledge rather than their own conceptualizations.

In summary, the conceptual preparation phase appears to be subject to two main biases: a self-generation bias that affects which referents speakers decide to talk about, and a conceptual match bias that affects how people talk about these referents. Moreover, as mentioned above, the conceptual preparation phase is followed by lexical selection, raising the question of whether and how these two biases affect this phase as well. One possibility is that in cases where a single conceptualization is associated with both self- and partner-generated words (e.g., A says [about a tangram figure]: “it looks like a guy leaning against a tree” and B replies: “yes, he looks very happy”), a self-generation effect would cause each speaker to favor the reuse of initially self-generated words. The current study was not designed to address this possibility, although self-generated words being more readily accessible in memory in Experiment 1 would be consistent with this account.

One main limitation of the current study is that pairs of participants were explicitly required to generate two different references for each tangram figure discussed. The purpose of this was to distinguish the effects of generation and/or conceptual match on reference reuse; however, it is unreasonable to assume that dialogue partners would spontaneously adopt this kind of behavior in real-life dialogue. Rather, in such situations, one of the partner might generate a reference matching his or her conceptualization of the referent under discussion; this reference would then be accepted by the other participant (Clark & Schaefer, 1989). This raises the question of whether accepting a reference necessarily involves adopting the same conceptualization as the partner who initially presented the reference. Another question raised concerns whether conceptualizations must necessarily belong to the common ground for them to be used as cues to reference production. Indeed, in Experiment 2, the reference that matched the director’s conceptualization best systematically belonged to the participants’ common ground (as it had been discussed during the Dialogue Phase), implying that the matcher was presumably capable of understanding it (regardless of whether it matched his or her own conceptualization as well or not). It was, therefore, not possible to determine whether the directors favored the reuse of such references because they knew that their partners were capable of understanding them, or whether they simply favored the reuse of references that were easy to produce for them. Finally, future studies should seek to generalize the current findings to situations in which dialogue partners refer to stimuli other than abstract tangram figures. One of the main features of these pictures is that they are not usually associated with a predefined label, involving that dialogue partners must explicitly negotiate which references should be used to refer to them (e.g., Clark & Wilkes-Gibbs, 1986; Duff et al., 2006). However, in real-life dialogue, speakers also refer to more consensual referents (e.g., place names, people’s names and object names). The influence of conceptual match on the production of such references might be smaller than in the current study, as there might be less room for speakers to think about how they might conceptualize the referents in this kind of situation.

At least two different constraints affect the reuse of references that belong to the common ground: accessibility in memory (that may depend on initial reference generation) and conceptualization. This finding is compatible with the idea that low-level, “ordinary” processes (i.e., processes that are not specific to dialogue) influence reference production (and comprehension) during dialogue (e.g., Gorman et al., 2013; Horton & Gerrig, 2002, 2005a, 2005b; Horton & Slaten, 2012). This idea was initially developed within the framework on the memory-based approach to dialogue, which emphasizes the role played by automatic memory processes in partner-adaptation. The current findings advocate for a broadening of the scope of this approach, as it seems that reference production depends on conceptual processes as well.

1  This suggestion should nonetheless be qualified in light of the fact that adding references to the common ground through presentation and acceptance is a joint process in which both partners play an active role: the partner performing the acceptance phase may suggest revising the reference initially presented (Clark & Brennan, 1991; Clark & Schaefer, 1989; Clark & Wilkes-Gibbs, 1986). This point will be addressed in the General Discussion.

Agresti, A. (2002). Categorical data analysis (2nd ed.). New York, NY: Wiley. 10.1002/0471249688

Bangerter, A., & Clark, H. H. (2003). Navigating joint projects with dialogue. Cognitive Science, 27, 195–225. 10.1207/s15516709cog2702_3

Bard, E. G., Hill, R. L., Foster, M. E., & Arai, M. (2014). Tuning accessibility of referring expressions in situated dialogue. Language, Cognition and Neuroscience, 29, 928–949.

Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68, 255–278. 10.1016/j.jml.2012.11.001

Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts and lexical choice in conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 1482–1493. 10.1037/0278-7393.22.6.1482

Brown-Schmidt, S. (2012). Beyond common and privileged: Gradient representations of common ground in real-time language use. Language and Cognitive Processes, 27, 62–89. 10.1080/01690965.2010.543363

Burnett, A. N., & Bodner, G. E. (2014). Learnin’ ’bout my generation? Evaluating the effects of generation on encoding, recall, and metamemory across study-test experiences. Journal of Memory and Language, 75, 1–13. 10.1016/j.jml.2014.04.005

Clark, H. H. (1992). Arenas of language use. Chicago, IL: University of Chicago Press.

Clark, H. H. (1996). Using language. Cambridge, MA: Cambridge University Press. 10.1017/CBO9780511620539

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). Washington, DC: American Psychological Association. 10.1037/10096-006

Clark, H. H., & Marshall, C. (1978). Reference diaries. In D. L.Waltz (Ed.), Theoretical issues in natural language processing (Vol. 2, pp. 57–63). New York, NY: Association for Computing Machinery.

Clark, H. H., & Marshall, C. (1981). Definite reference and mutual knowledge. In A. K.Joshi, B. L.Webber, & I. A.Sag (Eds.), Elements of discourse understanding (Vol. 2, pp. 10–63). Cambridge, MA: Cambridge University Press.

Clark, H. H., & Schaefer, E. F. (1989). Contributing to discourse. Cognitive Science, 13, 259–294. 10.1207/s15516709cog1302_7

Clark, H. H., & Wilkes-Gibbs, D. (1986). Referring as a collaborative process. Cognition, 22, 1–39. 10.1016/0010-0277(86)90010-7

Dell, G. S. (1986). A spreading-activation theory of retrieval in sentence production. Psychological Review, 93, 283–321. 10.1037/0033-295X.93.3.283

Duff, M. C., Hengst, J., Tranel, D., & Cohen, N. J. (2006). Development of shared information in communication despite hippocampal amnesia. Nature Neuroscience, 9, 140–146. 10.1038/nn1601

Flavell, J. H., Everett, B. A., Croft, K., & Flavell, E. R. (1981). Young children’s knowledge about visual perception: Further evidence of the Level 1–Level 2 distinction. Developmental Psychology, 17, 99–103. 10.1037//0012-1649.17.1.99

Fox Tree, J. E., & Clark, N. B. (2013). Communicative effectiveness of written versus spoken feedback. Discourse Processes, 50, 339–359. 10.1080/0163853X.2013.797241

Fukumura, K. (2015). Interface of linguistic and visual information during audience design. Cognitive Science, 39, 1419–1433. 10.1111/cogs.12207

Gorman, K. S., Gegg-Harrison, W., Marsh, C. R., & Tanenhaus, M. K. (2013). What’s learned together stays together: Speakers’ choice of referring expression reflects shared experience. Journal of Experimental Psychology: Learning, Memory, and Cognition, 39, 843–853. 10.1037/a0029467

Hanna, J. E., Tanenhaus, M. K., & Trueswell, J. C. (2003). The effects of common ground and perspective on domains of referential interpretation. Journal of Memory and Language, 49, 43–61. 10.1016/S0749-596X(03)00022-6

Hjelmquist, E. (1984). Memory for conversations. Discourse Processes, 7, 321–336. 10.1080/01638538409544595

Horton, W. S. (2007). The influence of partner-specific memory associations on language production: Evidence from picture naming. Language and Cognitive Processes, 22, 1114–1139. 10.1080/01690960701402933

Horton, W. S. (2008). A memory-based approach to common ground and audience design. In I.Kecskes (Ed.), Intention, common ground, and the egocentric speaker-hearer (pp. 189–222). Berlin, Germany: Mouton de Gruyter.

Horton, W. S., & Gerrig, R. J. (2002). Speakers’ experiences and audience design: Knowing when and knowing how to adjust utterances to addressees. Journal of Memory and Language, 47, 589–606. 10.1016/S0749-596X(02)00019-0

Horton, W. S., & Gerrig, R. J. (2005a). Conversational common ground and memory processes in language production. Discourse Processes, 40, 1–35. 10.1207/s15326950dp4001_1

Horton, W. S., & Gerrig, R. J. (2005b). The impact of memory demands on audience design during language production. Cognition, 96, 127–142. 10.1016/j.cognition.2004.07.001

Horton, W. S., & Slaten, D. G. (2012). Anticipating who will say what: The influence of speaker-specific memory associations on reference resolution. Memory & Cognition, 40, 113–126. 10.3758/s13421-011-0135-7

Hupet, M., Seron, X., & Chantraine, Y. (1991). The effects of the codability and discriminability of the referents on the collaborative referring procedure. British Journal of Psychology, 82, 449–462. 10.1111/j.2044-8295.1991.tb02412.x

Isaacs, E. A., & Clark, H. H. (1987). Reference in conversations between experts and novices. Journal of Experimental Psychology: General, 116, 26–37. 10.1037/0096-3445.116.1.26

Jaccard, J. (2001). Interaction effects in logistic regression. Thousand Oaks, CA: Sage. 10.4135/9781412984515

Jaeger, T. F. (2008). Categorical data analysis: Away from ANOVAs (transformation or not) and towards logit mixed models. Journal of Memory and Language, 59, 434–446. 10.1016/j.jml.2007.11.007

Jarvella, R. J., & Collas, J. G. (1974). Memory for the intentions of sentences. Memory & Cognition, 2, 185–188. 10.3758/BF03197513

Keselman, H. J., Algina, J., Kowalchuk, R. K., & Wolfinger, R. D. (1999). The analysis of repeated measurements: A comparison of mixed-model satterthwaite F tests and a nonpooled adjusted degrees of freedom multivariate test. Communications in Statistics Theory and Methods, 28, 2967–2999. 10.1080/03610929908832460

Kiernan, K., Tao, J., & Gibbs, P. (2012, April22–25). Tips and strategies for mixed modelling with SAS/STAT procedures. Presented at the SAS Global Forum, Orlando, FL.

Knutsen, D., & Le Bigot, L. (2012). Managing dialogue: How information availability affects collaborative reference production. Journal of Memory and Language, 67, 326–341. 10.1016/j.jml.2012.06.001

Knutsen, D., & Le Bigot, L. (2014). Capturing egocentric biases in reference reuse during collaborative dialogue. Psychonomic Bulletin & Review, 21, 1590–1599. 10.3758/s13423-014-0620-7

Knutsen, D., & Le Bigot, L. (2015). The influence of reference acceptance and reuse on conversational memory traces. Journal of Experimental Psychology: Learning, Memory, and Cognition, 41, 574–585. 10.1037/xlm0000036

Knutsen, D., Ros, C., & Le Bigot, L. (in press). Generating references in naturalistic face-to-face and phone mediated dialogue settings. Topics in Cognitive Science.

Krauss, R. M., & Weinheimer, S. (1966). Concurrent feedback, confirmation, and the encoding of referents in verbal communication. Journal of Personality and Social Psychology, 4, 343–346. 10.1037/h0023705

Kronmüller, E., & Barr, D. J. (2015). Referential precedents in spoken language comprehension: A review and meta-analysis. Journal of Memory and Language, 83, 1–19. 10.1016/j.jml.2015.03.008

Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press.

Levelt, W. J. M., Roelofs, A., & Meyer, A. S. (1999). A theory of lexical access in speech production. Behavioral and Brain Sciences, 22, 1–38. 10.1017/S0140525X99001776

Lysander, K., & Horton, W. S. (2012). Conversational grounding in younger and older adults: The effect of partner visibility and referent abstractness in task-oriented dialogue. Discourse Processes, 49, 29–60. 10.1080/0163853X.2011.625547

MacLeod, C. M. (2011). I said, you said: The production effect gets personal. Psychonomic Bulletin & Review, 18, 1197–1202. 10.3758/s13423-011-0168-8

MacLeod, C. M., Gopie, N., Hourihan, K. L., Neary, K. R., & Ozubko, J. D. (2010). The production effect: Delineation of a phenomenon. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36, 671–685. 10.1037/a0018785

McMahon, J. M., Pouget, E. R., & Tortu, S. (2006). A guide for multilevel modeling of dyadic data with binary outcomes using SAS PROC NLMIXED. Computational Statistics & Data Analysis, 50, 3663–3680. 10.1016/j.csda.2005.08.008

Metzing, C., & Brennan, S. E. (2003). When conceptual pacts are broken: Partner-specific effects on the comprehension of referring expressions. Journal of Memory and Language, 49, 201–213. 10.1016/S0749-596X(03)00028-7

Nückles, M., Winter, A., Wittwer, J., Herbert, M., & Hübner, S. (2006). How do experts adapt their explanations to a layperson’s knowledge in asynchronous communication? An experimental study. User Modeling and User-Adapted Interaction, 16, 87–127. 10.1007/s11257-006-9000-y

Rosner, Z. A., Elman, J. A., & Shimamura, A. P. (2013). The generation effect: Activating broad neural circuits during memory encoding. Cortex, 49, 1901–1909. 10.1016/j.cortex.2012.09.009

Satterthwaite, F. E. (1946). An approximate distribution of estimates of variance components. Biometrics, 2, 110–114. 10.2307/3002019

Slamecka, N., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4, 592–604. 10.1037/0278-7393.4.6.592

Spiers, H. J., Maguire, E. A., & Burgess, N. (2001). Hippocampal amnesia. Neurocase, 7, 357–382. 10.1076/neur.7.5.357.16245

Stalnaker, R. (1978). Assertion. In P.Cole (Ed.), Syntax and semantics (Vol. 9, pp. 315–332). New York, NY: Academic Press.

Yoon, S. O., & Brown-Schmidt, S. (2014). Adjusting conceptual pacts in three-party conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 40, 919–937. 10.1037/a0036161

To examine whether the participants’ performance during the Memory Assessment Phase depended on other determinants than Generation and Conceptualization, the data were recoded for Acceptance and Short-term reuse.

A presented content word was coded as accepted through verbatim repetition (Code 1) when the participant who did not perform the presentation repeated it between the moment when its initiator presented it and the moment when he or she produced another content word. All other presented content words were coded as accepted through another mean (Code 0).

All occurrences of content word production that did not count as presentations or acceptances through verbatim repetition were coded as reuses. Following previous work on the self-presentation bias, a reference was counted as reused only if the speech turn in which it occurred was preceded by a minimum of two speech turns during which it did not occur (Knutsen & Le Bigot, 2012). Therefore, each reference presented was either coded as reused (Code 1; note that this code was used regardless of how many times the reference had actually been reused) or nonreused (Code 0). Note that these two variables (Acceptance and Short-term reuse) were coded at the dyadic level: An accepted reference was coded as accepted from both participants’ point of view and a reused reference was coded as reused from both participants’ point of view. In the final dataset (i.e., the dataset used in the main analysis), the total number of content words coded as accepted through verbatim repetition was 183 (4.77% of all content words presented) and the total number of content words coded as reused was 162 (5.60% of all content words presented).

A logistic mixed model was used to analyze the data. Following the same rationale as in the main analysis, this model included Generation and Conceptualization as fixed effects and recall as the outcome variable. The initial model also included all random intercepts and random slopes justified by the design. It also included by-dyad, by-participants, and by-item random slopes corresponding to Acceptance and Reuse. This allowed us to remove from the model the variability associated with these two factors.

The random effects causing model convergence failure were then removed from the model, which did not affect the outcome of the analysis (removing the random slopes corresponding to Acceptance and Short-term reuse during this process would imply that the variability associated with these random effects was equal to zero).

The random effects structure of the final model included (a) by-dyad random slopes corresponding to Conceptualization, Acceptance and Short-term reuse; (b) by-participant random intercepts and by-participant random slopes corresponding to Generation; and (c) by-item random slopes corresponding to Conceptualization, Acceptance and Short-term reuse.

Generation significantly predicted recall, F(1, 69) = 10.99, p = .002. The participants were more likely to recall self-generated content words than partner-generated ones, OR = 1.45, 95% CI [1.16, 1.82]. Conceptualization also significantly predicted recall, F(1, 25) = 408.01, p < .001. The participants were more likely to recall content words that matched their conceptualizations than content words that did not match their conceptualizations, OR = 16.43, 95% CI [12.35, 21.86].

Furthermore, the interaction between these two factors was statistically significant, F(1, 5793) = 18.84, p < .001. Additional pairwise comparisons (Sequential Bonferroni) were conducted to offer a better understanding of this pattern of results. These comparisons revealed that the difference between self- and partner-generated references was significant when these references matched the participants’ conceptualization (p < .001) but not when these references did not match the participants’ conceptualizations (p = 1.00). The model parameters are reported in Table B1.

This pattern of results is identical to the one reported in the main analysis. These results confirm that the effects of Generation and Conceptualization remained statistically significant even when Acceptance and Short-term reuse were taken into account in the analysis.

The dialogues between the participants during the Dialogue Phase were coded for Acceptance and Short-term reuse. In the final dataset, the total number of content words coded as accepted through verbatim repetition was 102 (3.46% of all content words presented) and the total number of content words coded as reused was 95 (3.23% of all content words presented).

This additional analysis was conducted following the same rationale as the additional analysis in Experiment 1. The random effects structure of the final model included the same random intercepts and slopes as the main analysis (i.e., by-dyad random slopes corresponding to Generation and Conceptualization, by-participant random slopes corresponding to Generation and Conceptualization and by-item random slopes corresponding to Generation and Conceptualization). In other words, random slopes corresponding to Acceptance and Short-term reuse were not included in the final model because the variability associated with these random effects was equal to zero, implying that it was not necessary to control for these variables in this analysis (the results were identical regardless of whether or not these slopes were included in the model). Because the random structure was identical to that used in the main analysis, the results were also identical to those of the main analysis (i.e., only a significant effect of Conceptualization was found, F(1, 29) = 114.57, p < .001). The model parameters were also identical to those found in the main analysis (see Table A2; because the parameters were identical to those reported previously, they are not reported again in this appendix). This confirms that Conceptualization significantly affected reuse in this experiment, regardless of potential acceptance and/or short-term reuse.

Submitted: January 5, 2016 Revised: June 7, 2016 Accepted: June 9, 2016