Since Bartlett’s seminal work on constructive memory (Bartlett, 1932), many studies have shown that people’s recall of narrative information is not a literal record, but tends to be reconstructed according to one’s acquired cultural knowledge. Recollections are not simply subject to random forgetting, but are the result of systematic modifications and alterations. The constructive aspect of memory has been explained by processes such as conventionalization, rationalization, simplification, assimilation, and distortion (Wagoner, 2017). What people typically remember is strongly influenced by the categories and schemata they utilize when interpreting their experiences.

Remembering also has a communicative function. We talk about our experiences, and we like to share our memories with others (Hirst and Echterhoff, 2012). Stories are the main carriers of socio-cultural knowledge and serve a social as well as an informative function (Schank and Abelson, 1995). In the process of telling stories to others, people take into account the kind of audience they are addressing; what people tell to others is influenced by their own intentions, for example, wanting to inform or to persuade, but is at the same time influenced by features of the audience they talk to, such as attitudes, expectations, or preferences (Marsh, 2007).

In this study, we apply the constructive memory framework to climate change narratives (Jones and Song, 2014; McBeth et al., 2014; Fløttum and Gjerstad, 2017; Sharman and Howarth, 2017). Concerning the climate change debate, a distinction often made is between people who endorse the assumption that climate change is happening and mainly of anthropogenic origin, and those who question that climate change is real or who consider it to be an entirely natural phenomenon (Poortinga et al., 2011; Hamilton et al., 2015). For the sake of brevity, we call these positions advocates and skeptics, respectively. Advocates and skeptics show substantial differences in their world views and political orientation, and in their beliefs about the causes and consequences of climate change. Between these camps, an increasing polarization has been observed, though the degree of polarization differs considerably between countries (Hoffman, 2011; Dunlap et al., 2016; Ceglarz et al., 2018).

Polarization may partly be attributable to so called ‘echo chambers’ or ‘filter bubbles’; these metaphors mainly refer to social media such as Facebook and Twitter, where like-minded individuals communicate with each other, reinforcing their respective stance and evaluation of climate change (Jasny et al., 2015; Williams et al., 2015). Retelling a story about climate change to people who essentially agree with one’s view may bolster this view in two ways: first, the speaker, holding world view X, adapts her retelling to her own world view; since the speaker and the listener hold the same world view, no conflict arises and the speaker’s statements are likely to be accepted and reinforced by the listener. Second, the speaker tunes her retelling to the listener’s world view X, and thus constructs a version that even more conforms to her own world view X. In a communication between like-minded, adaptation and tuning both enhance the correspondence between the communicators’ world views and the content of the communication. We propose that such processes of constructive memory may partly account for the ‘echo’ in echo chambers, and for the polarization of the climate change debate.

We report a study using methods from computational text analysis (Welbers et al., 2017) to analyze participants’ recollections of a narrative about climate change. This study contributes to a growing line of research which uses natural language texts or open-ended textual responses, that is, ‘text as data’ rather than quantitative survey questions, to investigate how people understand and evaluate climate change (Grimmer and Stewart, 2013; Paschen and Ison, 2014; Tvinnereim and Fløttum, 2015; Fløttum, 2017; Salway, 2017; Tvinnereim et al., 2017). Most research using text analysis employs a bottom-up approach, inducing regularities in massive collections of texts from online sources (Facebook, Twitter, etc.) by automated classification methods, such as clustering or topic modeling (Grimmer and Stewart, 2013; Roberts et al., 2014; Tvinnereim and Fløttum, 2015). We apply computational text analysis to data from an online experiment, analyzing quantitative characteristics of texts as a function of experimental manipulations in order to test hypotheses (Roberts et al., 2016).

Specifically, we test how the speaker’s view and that of the audience shape how a story with a climate change theme is recollected and retold. In addition, we test how these recollections change over time. We hypothesize that story retellings will be adapted to the speaker’s as well as to the listener’s world view, and that the conformity will increase over time.

Based on the theoretical background outlined above, we assume that climate change stories are shaped by the world views of people who tell such stories, as well as by the world views of the audience to whom the stories are told. World views are assumed to work as reconstruction filters, modifying stories when they are repeatedly told and retold. To our knowledge, the role of world views in memory and communicative processes in the context of climate change has not yet been studied. The investigation of audience world views and their interaction with speakers’ world views, in combination with applying computational text analysis, thus contributes to the novelty of this study.

Three research questions will be examined in more detail:

To address these research questions, we included the speaker’s and the audience’s world view as independent variables. In addition, we had participants retell the story at two points in time. As there exists not the one accepted and valid method to assess story content and structure, we applied three approaches complementing each other: A sorting task of story related concepts (Coxon, 1999), and two methods from computational text analysis: similarity analysis (Kjell et al., 2019) and dictionary analysis (Welbers et al., 2017); for details see the method section.

An online study was conducted with a representative Norwegian sample. All manipulations and measurements were conducted online.

We will first report on the classification of participants into world view clusters; this assignment of participants to world views will then be used as a quasi-experimental factor. Second, the sorting task will be analyzed in order to examine how participants’ and audience world views are reflected in the participants’ cognitive representations of the story. Third, we look at the textual retellings and examine via computational text analysis how the retellings’ content changes as a function of the experimental factors (similarity analysis), and to what extent the retold stories conform to the speakers’ and the audience’s world views (dictionary analysis).

People’s evaluation of contentious social issues such as climate change is often only to a small degree influenced by their factual knowledge about these issues; cultural values and world views often play a stronger role in shaping such evaluations (Kahan et al., 2011). When people communicate about an issue, they tell and retell narratives that embody their beliefs as well as appraisals of the causes and evaluations of the consequences, conferring meaning to the issue (McAdams and McLean, 2013; Brown, 2017). One reason why these narratives and associated evaluations are so persistent and resist change might be due to memory processes which are at work during the communication process. Repeated telling and retelling of narratives can be expected to strengthen those aspects that are told and to weaken those aspects left out. Consequently, narratives may become more and more coherent and compatible with the person’s own core values and beliefs. Especially when conversing with like-minded people, a plausible assumption is that world views mutually reinforce each other and narratives are increasingly adjusted to these world views.

Climate change may serve as a prime example of this. Narratives about climate change provide explanations: whether the phenomenon exists at all, what its causes and its consequences are, what should be done. Narratives appear to be largely immune to scientific facts; given the large scientific consensus on climate change one might otherwise expect that all people would tell the same story. What happens is quite the contrary, narratives of climate change are quite diverse; some tell about villains who destroy the earth, others about conspiracies initiated in order to subdue the free world, and people align the moral of the story with their own basic beliefs and world views (Jones, 2014). Social media in particular may play the role of echo chambers, where communities of like-minded individuals mutually confirm their views about climate change (Jasny et al., 2015; Flaxman et al., 2016).

In this study we examined processes of constructive memory (Bartlett, 1932; Wagoner, 2017) as one possible factor shaping peoples’ climate change narratives. World views, we assume, serve as cultural schemata that operate as filters when people recollect narrative information and share their narratives with others. World views filter meaningful components of a narrative, sifting out what is not compatible with one’s world view. World views also operate when telling stories to others, in the sense that stories are tuned to the world view of the audience.

The focus of this study was on examining actual retellings, that is, texts generated by participants when asked to remember and retell a previously read narrative. We analyzed these natural language data by means of computational text analysis; specifically, we computed similarities between the retellings and the original story, between the retellings and partial aspects (story lines) of the story, and we computed the amount of specific world view-related topics occurring in the retellings via dictionary analysis (Welbers et al., 2017). Since what has been called the narrative turn in the social sciences, narratives have been mostly analyzed by qualitative methods (Riessman, 1993). In contrast, we attempted to quantify the main components of retold stories, and their interrelationships, aiming to capture important aspects of reconstructive processes. Additionally, we assessed the ensuing cognitive representation of the narrative by means of a more traditional sorting task method. Participants sorted the main terms from the narrative into groups, and a cluster analysis was employed to detect the underlying story representation from the derived co-occurrence matrices of terms.

The present study shows that world views exert a small though non-negligible influence on how climate change narratives are remembered and retold. An examination of the mental representation of a climate change story via a sorting task/cluster analysis approach revealed that although the general story structure is very similar across world views, the link between the problem component (a crisis due to climate change) and the proposed problem solutions (strategies to counteract climate change) varies systematically as a function of the audience’s world view and of the speaker’s world view. The audience effect indicates that speakers tune the retelling to the audience’s world view and connect the problem with that solution which is preferred by the audience; for example, retelling the story to an egalitarian who is assumed to prefer a sustainable strategy yields a close association between crisis and sustainable strategy. Also, the egalitarian structure contrasts to both the individualist and the hierarchist structure, the latter two being both more closely associated with a non-sustainable solution.

The effect of the speaker’s world view, in contrast, indicates that speakers tend to connect the aspects of the crisis with that strategy which – from their point of view – is the cause and culprit of the crisis. Thus, unlike the audience effect, which links the crisis with a solution, the effect of the participant’s world view is to link the crisis with the problem. Participants with an egalitarian world view closely associate the crisis with the free market, and with the respective protagonists. An individualistic participant, in contrast, associates the crisis with sustainable strategies. In sum, the findings from the cluster analyses suggest that the reteller’s world view links the story problem with the ‘villain’ (as seen from the reteller’s world view), whereas the audience effect tends to make the retelling compatible with the solution as seen from the audience’s world view.

The computational text analysis examined three hypotheses: first, a time effect was expected, that is, retellings were assumed to become less similar to the original story over time. Second, we expected a schema conformity effect, consisting of an audience effect and a speaker effect: Retellings were expected to become more similar to the speaker’s as well as to the audience’s world view, and consequently become less similar to the original story. Third, we expected that in the control group, retellings would be least affected by reconstruction processes, due to the absence of an audience.

The similarity analyses yielded a time effect, but only in the somewhat trivial sense that people forget when time passes. We found an audience effect, but that was unspecific: in the control condition, when participants simply recalled the original story without telling it to an audience, retellings were most similar to the original story. If an audience was present, retellings decreased in their similarity to the original story, but more or less to the same degree across the various audiences, irrespective of the specific world view held by the listener. In fact, the decay over time was not significant for the control/recall condition. We assume that world views of an audience do in fact operate as filters, but in an unspecific manner, introducing alterations and omissions unrelated to the audience’s specific world view, and distorting the original information in a somewhat random way. This finding suggests that talking to an audience of any kind activates a filter process so that the recollected story bears less resemblance to the original story than under pure recall without an audience. This is counter to our schema conformity hypothesis, which states that if speaker and listener share the same world view the filter effect would be amplified.

We find, however, some suggestive indication of schema conformity, though contrary to our hypothesis. Although the interaction between Audience and World View is not significant, F(6,254) = 1.9, p = 0.08, the interaction pattern suggests that identical world views of speaker and listener may increase story veridicality. When speaker’s and listener’s world views match, less filtering occurs and the narrative is more, rather than less, similar to the original story than when the world views of speaker and listener differ (Figure 4). A tentative explanation might be a common ground hypothesis (Clark and Brennan, 1991; Keysar et al., 1998), assuming that matching world views provide sufficient common ground for mutual understanding, making constructive processes and audience tuning superfluous. The motive to modify and adapt the recollection is low because no explanation or persuasion is needed and common ground is implicitly presumed. Interestingly, there is virtually no decline in similarity over time for the control/recall condition, suggesting that retelling to an audience not only decreases similarity due to audience tuning, but also makes the memory trace more fragile and amenable to deterioration over time.

Results of splitting up the story into its three story lines related to the three protagonists (Greene the egalitarian, Brown the hierarchist, and Wayne the individualist) are in line with the other findings. The control group yields largest similarities to all story characters, whereas the specific audience world views show no effect. However, an interesting interaction between story character and speakers’ world view emerges: those participants who hold an egalitarian world view produce the most similar recollections with respect to all three central characters of the story (Figure 7). Correspondingly, all world view conditions generate retellings that are most similar to the egalitarian character Matt Greene. This pattern might be due to the egalitarian world view being closest to the dominant politically correct view, as presumably seen by many people in Norway. A possible explanation could be that dominant world views operate as weak filters, generating little reconstructive modifications; conversely, minority world views might operate as strong filters and generate more substantial modifications and distortions.

The dictionary analysis largely confirms the findings from the similarity analysis. The pattern of word counts for different term categories does not support the hypothesis that those terms are recollected relatively more often which conform to the participant’s world view or to the audience’s world view. Instead, we find the opposite – no interaction between Term Category and World View, and the interaction between Audience and Term Category yields greater proportions for the control condition with respect to all categories, which is again in line with assuming an unspecific filter mechanism. In addition, the dictionary analysis shows virtually identical and very low proportions for all Audience conditions with respect to the crisis category. A plausible explanation might be that the crisis topic is a story feature that does not discriminate between the story characters and their strategies.

It should be noted that retelling a previously heard story is essentially a memory task and might be expected to substantially depend on age. However, including age as a covariate had virtually no effect on the results; in fact, age per se turned out to be not significantly related to any dependent measure. The task of retelling a semantically rich narrative is quite different from rote learning; although older people usually experience a decline in short term memory, memory loss is less pronounced or absent for personally relevant episodes and similar tasks (Schacter, 1996). Furthermore, since the amount of correctly remembered information from the original story was quite low generally, it could also be a floor effect obliterating age differences.

In sum, we find mixed evidence on how world views operate on the comprehension, recollection, and communication of climate change narratives. Results from the sorting task suggest that there is a small but systematic effect of world views on how people connect the problem of a story with its possible solutions. Results from the text analyses suggest that world views largely operate in a non-specific manner; irrespective of the audience’s world view, people modify their recollections whenever an audience is present (retelling conditions) in contrast to recalling a story in isolation (Marsh, 2007).

Several limitations of this study should be noted. The validity and reliability of narrative analyses has often been questioned (Riessman, 1993; Grimmer and Stewart, 2013; Brown, 2017). A direct written recording of recollections, as used in our study, yields extremely noisy data. After normalization, the mean number of words of a text unit was about 40, whereas the original story had approximately 600 words; some retellings consisted of just a few words. This constitutes an enormous loss of information that cannot be explained by a systematic influence of schemata on the reconstruction. The similarities between retellings and the original story are generally very low and the signal/noise ratio might be too low to detect substantial effects; especially when splitting the retellings according to the three story lines, this effect might be exacerbated. It may be questioned whether a completely open response format is capable of capturing the conceptual content of peoples’ recollection with sufficiently high precision; it may be that the requirement to articulate and write down one’s recollection simply generates a large amount of noise. Also, measuring similarity via a cosine similarity measure based on common word frequencies has its drawbacks. For example, in some conditions participants might rephrase the original wording using synonyms, which would lead to low similarity just because different words are used for the same concepts.

The sorting task, on the other hand, provides a kind of scaffold that guides recollection and blocks noisy intrusions, yielding a more stable measurement of the story’s representation, albeit in a less direct way. Providing the relevant concepts in the sorting task might artificially increase recollection, independent of experimental condition, and thus yield homogenous representations and obscure real differences. Also, it is somewhat unclear how strongly the sorting process depends on the recollected story; it might mainly be driven by the general world knowledge and understanding of the presented concepts. How to best measure narrative content and the dynamics of narrative change remains an open question and a challenge for future research.

Can memory processes explain echo chamber effects? If we view echo chambers as closed systems, that is, as a collection of like-minded individuals communicating only with each other, we would predict that any story would be distorted and fragmented over time, with little information remaining the same over time. However, the assumption that echo chambers operate as filter bubbles (Flaxman et al., 2016), systematically extracting compatible and distorting incompatible information, might be too strong. Our data indicate that the simple process of retelling, no matter to whom, generates a loss of information that is unsystematic. Furthermore, what we tentatively called a common ground hypothesis suggests that like-minded individuals have no motive to systematically reconstruct the stories that they tell each other, simply because they are like-minded and share a common world view in the first place. In contrast, when confronted with individuals from a different echo chamber, some adjustments in communicated content might well occur; our findings suggest that at least the key components of a story – Who is the villain? Who caused the problem? – might be constructed according to the speaker’s and the listener’s world views. These effects were small, though, in our study and they have also typically been small in other research (Jones and Song, 2014; Williams et al., 2015; Flaxman et al., 2016; Quattrociocchi et al., 2016). The existence of echo chambers appears to be an obvious phenomenon, and there is evidence that people tend to engage in social media activities with like-minded others, albeit not exclusively so (Williams et al., 2015; Flaxman et al., 2016). However, tapping into the specific processes that may lead to increasing polarization between echo chambers through transformation of discourse content, and studying these processes in a systematic scientific manner, may be more challenging than anticipated.

Potential implications of this study might go in two directions. First, we think that it contributes to the understanding of echo chamber effects, shedding some light on the role of reconstructive memory processes. Second, we hope to advance the understanding of computational text analyses in the study of narratives and to contribute to its methodological development.

This empirical study complied with the Norwegian Social Science Data Services (NSD) privacy regulations and the ethical principles of research by the National Committee for Research Ethics in the Social Sciences and the Humanities (NESH). The NSD has approved that the project complies to these rules and regulations (project number 44066).

GB and H-RP contributed conception and design of the study. GB, H-RP, and AS performed the statistical analyses. H-RP wrote the first draft of the manuscript. H-RP and GB wrote sections of the manuscript. All authors contributed ideas to manuscript revision, read and discussed the manuscript, and approved the submission.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.