Weaving a story: Narrative formation over prolonged time scales engages social cognition and frontoparietal networks

Forming narratives is of key importance to human experience, enabling one to render large amounts of information into relatively compacted stories for future retrieval, giving meaning to otherwise fragmented occurrences. The neural mechanisms that underlie coherent narrative construction of causally connected information over prolonged temporal periods are yet unclear. Participants in this fMRI study observed consecutive scenes from a full‐length movie either in their original order, enabling causal inferences over time, or in reverse order, impeding a key component of coherent narratives—causal inference. In between scenes, we presented short periods of blank screens for examining post‐encoding processing effects. Using multivariate pattern analysis (MVPA) followed by seed‐base correlation analysis, we hypothesized that networks involved in online monitoring of incoming information on the one hand, and offline processing of previous occurrences on the other would differ between the groups. We found that despite the exposure to the same scenes, the chronological‐order condition exhibited enhanced functional connectivity in frontoparietal regions associated with information integration and working memory. The reverse‐order condition yielded offline, post‐scene coactivation of neural networks involved in social cognition and particularly theory of mind and action comprehension. These findings shed light on offline processes of narrative construction efforts, highlighting the role of social cognition networks in seeking for narrative coherence.


| INTRODUCTION
By forming narratives, humans are able to bind large portions of discrete events into meaningful representations. The ability to produce narratives is pivotal to one's sense of personal and cultural identity, as well as to the creation and construction of memories (Hoerl, 2007;Sarbin, 1986). A key defining feature of narrative is causation (Hayles & Richardson, 1999). In 'Poetics', Aristotle emphasized the importance of coherence in narrative, which results from logical, causal connections between occurrences (Barthes, 1982). Accordingly, literary accounts define narrative as a series of actions and events that unfold over time according to causal principles (Mar, 2004). For a coherent narrative to form, events must occur in a logical order, wherein actions and episodes that lead to other events must take temporal precedence given the conflation of logical (if x then y), causal (because x then y) and temporal priority (first x then y) (Graesser et al., 1980;Mar, 2004).
Neural correlates of causal inferences have typically been studied in short time scales. For example, generating inferences during the reading of short texts correlated with activation in superior temporal gyrus (Virtue et al., 2006), and regions in prefrontal cortex ascribed to retrieval and working memory were found to mediate processes essential to causal inferences during narrative comprehension (Kuperberg et al., 2006;Mason & Just, 2004). However, due to the nature of real-life naturalistic events, narrative construction often demands causal inferences in the range of several minutes and beyond. Here, we manipulated causality by changing the order of scenes that collectively construct a prolonged narrative, aiming to delineate neural networks that support narrative construction with causal connections over prolonged time scales.
Besides causal inferences, stories and events are typically formed around a protagonist, who aspires to a particular goal and attempts to realize it (Oatley, 1992). The story line typically contains interactions between two or more figures, requiring one to observe, understand and mentalize others' behaviours and intentions, features that require social skills and particularly theory of mind (ToM) (Mason & Just, 2009). ToM refers to the ability to attribute internal mental states to others, as well as reasoning about one's own mental state (Mason & Just, 2009), to infer and represent their beliefs and desires (Bernhardt & Singer, 2012). Indeed, a protagonist perspective network of brain regions, consisting of temporoparietal junction (TPJ), dorsomedial prefrontal cortex (dmPFC) and superior temporal cortex, has been suggested to underlie narrative comprehension (Mar, 2004).
Narrative formation is essential to the representation of past events, a notion dating back to Pierre Janet , who posited two parallel processes that co-occur during narrative construction. Specifically, ongoing events are perceived and encoded online, and in parallel, the meaning of what had happened-the narrative-is generated (van der Hart et al., 2006). Accordingly, we expect that narrative formation would instigate (1) an online process of perceiving the continuous stream of information and (2) a parallel process of integrating incoming information with recently accumulated information according to causal inferences regarding the characters' aims and the general storyline.
In recent years, narratives were used as a tool to study various cognitive processes involved in story comprehension, such as event segmentation (Baldassano et al., 2017;Kurby & Zacks, 2008;Richmond & Zacks, 2017;Zacks, 2010;Zacks et al., 2001) and narrative shifts (Whitney et al., 2009). Studies focusing on the understanding of narratives have discerned between brain networks processing short segments of information versus full narratives (Yarkoni et al., 2008;Yeshurun et al., 2017). Particular emphasis was devoted to the default mode network (DMN), and specifically to the precuneus and lateral parietal cortex, which appear to display unique activation patterns that are dependent on the high-level meaning of conveyed scenes, rather than on their physical attributes Simony et al., 2016;Yeshurun et al., 2017). Taken together, these studies identified brain regions that correlate with the outcome of narrative formation. However, the process of narrative formation, that is, how discrete units of information, processed by distributed brain regions, are bound together through causal relationships across prolonged time scales to form a cohesive meaning is yet unclear.
To explore the neural correlates of narrative formation, we presented participants with 22 scenes from the classic film 'Bicycle Thieves' (Vittorio de Sica, 1948) that we edited in a way that preserved the central plot. The scenes were presented either in their chronological order to one group, meeting the requirements of the definition of narrative, or in reverse order to a separate group, breaking temporal causation between events and impeding the formation of a coherent narrative (Graesser et al., 1980), while preserving the same sensory input. We chose to reverse-scene order, as opposed to using other forms of narrative control conditions, such as scene scrambling (Lerner et al., 2011;Simony et al., 2016), to ensure the complete absence of direct causal connections between narrative-forming scenes. We relied on the causality principle, which states that if A is the cause of B, then B cannot precede A (Goldvarg & Johnson-Laird, 2001). This theoretical law of logic is embedded as a cognitive trait that develops already at young age (Burns & McCormack, 2009;Rankin & McCormack, 2013). Scrambling the order of scenes thus poses the possibility that causal inferences may be formed for at least part of the occurrences. Reverse-scene order, on the other hand, precludes this option.
Events that unfold over time and require the binding of information are hypothesized to be processed not only during the event itself but also upon event offset (Ben-Yakov et al., 2013;Cohen et al., 2015). This might be particularly true for processing information that is not easily comprehended online, as the case in scenes that are presented out of their natural order. We therefore introduced short blank epochs between movie scenes, aiming to shed light on processes of narrative formation that may occur during these periods.
Observing a sequence of events over prolonged durations in reverse order may cause alterations in factors such as alertness, attention and motivation. We monitored these factors by recording pupillometry measures throughout the task as an indicator of task engagement (i.e., attending to the presented scenes) and collecting post-task assessments. We also examined the participants' causal understanding by asking them to describe the causal connection between plot-related scenes that appeared in both conditions at roughly the same time after the experiment onset.
Because one of the defining features of an unfolding narrative is the temporal evolution of its plot, we applied an analysis strategy that considers dynamic time courses of blood-oxygenated-level-dependent (BOLD) signal (functional connectivity), rather than univariate approaches of average signal strength, which are blind to temporal changes and rely on repetitive events that are absent in a progressing plot. We employed a functional connectivity multivariate pattern analysis (MVPA) approach for extracting brain regions that differ in functional connectivity between the chronological-and reverse-order conditions, followed by seed-based correlation (SBC) analysis of the regions that were delineated by the MVPA during both scene presentation and interscene ('blank') periods. These analyses were designed to explore our hypotheses that breaking the causal relationship between successive occurrences would be manifested in functional integration of networks that underlie processing of online information, as well as online and/or offline interpretation and mentalization of social interactions.

| Participants
Thirty-eight healthy individuals (mean age 27 ± 3.7 years) participated in the study. Two separate groups of 19 subjects were assigned to either the chronological-order group or to the reverse-order group (eight and nine females, respectively, in the chronological-and reverse-order groups). All participants had normal visual acuity (without glasses or corrective lenses and enabling eye tracking). None of the subjects were familiar with the movie presented during the scan. The experimental procedures were approved by the Institutional Review Board of Tel-Aviv University and the Ethics committee of the Chaim Sheba Medical Center, Tel-Hashmoer, Israel, as required by Israeli law. All subjects provided written informed consent prior to the experiment, and all methods were performed in accordance with the relevant guidelines and regulations.

| Stimuli and experimental design
In the MRI scanner, participants watched movie scenes either in their chronological order or in reverse order. The movie consisted of 22 scenes (mean scene length: 55.15 ± 23 s) edited from the movie 'Bicycle Thieves' (Italy, 1948), interleaved by 10 s blank screens. Briefly, the movie portrays a young father, who manages to obtain a pair of bicycles that are required for his employment. On his first day at work, his bicycles are stolen (the plot's turning point), and from that moment on, the main protagonist tries to retrieve them with his son. The unfortunate events that follow feed his desperation, which culminate in a failed attempt to steal someone else's bicycles, just to be caught by the police and shamed in front of his son. The experiment lasted 25 min, wherein the scenes were identical across groups and differed only in their order of presentation, such that in the chronological condition, scenes were presented in their correct order-from first to last, and in the reverse-order condition, the same scenes were presented in reverse order, from last to first. The movie was edited using 'Movie-Maker' software (Microsoft ©), and as presented using Presentation Version 20.1 software (www.neurobs.com).

| Behavioural assessment
Immediately following scanning, participants were asked to answer a questionnaire targeting comprehension and subjective ratings regarding various aspects of the movie (see Appendix S1). Specifically, subjects were required to state whether they understood the story depicted in the movie segments and to state their confidence level. They were also instructed to indicate the roles of the main characters in the movie and to describe the general plot in a few sentences. Participants were asked whether or not they felt they understood the movie's plot and whether they thought the scenes were presented in their original chronological order. Their accounts were subsequently scored by ascribing one point to each character and role they correctly identified (father, kid and bicycle thief), a point for correctly describing the main protagonist's goal (i.e., attaining a bicycle for work purposes) and the main dramatic development (the bicycles being stolen by a thief). In addition, they were asked more specifically about the casual connection between two successive scenes that occurred in the middle of the movie; the first depicted the main protagonist slapping his child, and the second showed a compensation scene wherein the father treats the kid to a meal in a fancy restaurant. This assessment was designed to probe the participants' understanding of the events' causation-a central component in the definition of coherent narrative.

| MRI data acquisition and preprocessing
Whole-brain imaging was performed in a 3T Siemens Magnetom MRI system (Siemens Medical Systems, Erlangen, Germany) using a 16-channel head coil. BOLD-sensitive T2*-weighted functional images were acquired using a single shot gradient-echo EPI pulse sequence (TR = 2000 ms, TE = 30 ms, flip angle = 82 , 64 axial slices, 2 mm 3 , FoV = 192 Â 192 mm, interleaved slice ordering) and corrected online for head motion. Participants completed the task in two fMRI runs, each lasting $12.5 min. The first two volumes were discarded to allow for equilibration effects. Visual stimuli were presented on a screen behind the scanner using Presentation software (www.neurobs.com) and were viewed through a mirror attached to the head coil. Following functional imaging, a high-resolution T1 scan was acquired for anatomic normalization.
Imaging data were slice-time corrected and realigned using the SPM12 package (Wellcome Institute of Cognitive Neurology, London). Functional volumes were coregistered, normalized to the Montreal Neurological Institute (MNI) template brain and smoothed with an 8 mm 3 isotropic Gaussian kernel. We assessed taskrelated functional connectivity using the CONN toolbox (v17) (Whitfield-Gabrieli & Nieto-Castanon, 2012). The implemented CompCor routine was carried out for each participant (Behzadi et al., 2007), aimed at identifying principal components associated with white matter (WM) and cerebrospinal fluid (CSF), which were segmented. These components, as well as realignment correction information, linear BOLD signal trends were entered as covariates in the first-level model. In addition, to reduce the influence of task-induced transient responses in BOLD signal (i.e., scene onset/offset transients), the task segments were convolved with a haemodynamic response function (HRF), defined as noise components and removed separately from each voxel using ordinary least squares (OLS) regression.

| Functional connectivity analysis
In order to detect brain regions that differed in their functional connectivity between the two conditions (i.e., the two groups), we carried out the following analysis steps: (1) a functional connectivity MVPA (fc-MVPA) of pairwise connections between all measured voxels separately for film and blank segments (voxel-to-voxel analysis); (2) second-level between-group analyses that tested for differences in connectivity between the groups using F tests; and (3) a seed-based correlation (SBC) analysis (SCA), using the regions that the fc-MVPA analysis yielded as seeds. Correlations were computed between these regions of interest (ROIs) and the rest of the brain (seed-to-voxel analysis) in order to characterize the connectivity patterns of the brain regions associated with the chronological-and reverse-order conditions. These analyses were performed separately for concatenated scene periods and interscene (blank) periods (see detailed description below).

| Functional connectivity MVPA
MVPA provides a whole-brain data-driven approach that provides information as to functional connectivity patterns of each voxel with all other voxels unique to different conditions and/or groups. MVPA was carried out as implemented in the Conn toolbox (Whitfield-Gabrieli & Nieto-Castanon, 2012), similarly to the procedures used in previous publications (Eckstein et al., 2022;Mateu-Estivill et al., 2021;Morris et al., 2021). Pertinent to functional connectivity, MVPA allows for detection of group differences without asserting assumptions concerning a-priori seed regions.
Using the fc-MVPA routine, between-group voxel-tovoxel functional correlation differences were computed across the entire dataset as follows: For each participant, low-dimensionality representations of pairwise correlation matrices were formed using singular value decomposition (SVD) on the preprocessed and de-noised BOLD series, separately for the 'scene' and 'interscene blank' segments, using 64 components (Whitfield-Gabrieli & Nieto-Castanon, 2012). This procedure is implemented in the CONN toolbox 'subject-level dimensionality reduction' procedure (Nieto-Castanon, 2020). These 64 components indicate the number of subject-specific components retained when characterizing the voxel-by-voxel correlation structure for each subject (Kumar et al., 2022). Next, using each voxel as a seed, the functional connectivity with all other voxels in the brain were concatenated into a matrix of M (subjects) Â N (number of voxels) for each condition separately (see Figure 1a). Following this, a second SVD was applied on the connectivity matrix of each voxel, using the equation where R(x) is the connectivity matrix for a particular voxel over subjects, S and P are left and right singular vectors and D is the diagonal matrix of singular values. Jointly across subjects but separately for each voxel, the four strongest eigenvectors were retained from the SVD of the between-subject variability in voxel-to-voxel connectivity, whereby each eigenvector (termed here eigenpattern) represents the explained variance in the connectivity of a particular voxel with all other voxels over subjects, in descending hierarchical order of explained variance (Figure 1b). The eigenpattern scores (separate values for each voxel, subject and component) were then divided to chronological-and reverse-order groups and entered into a second-level GLM for testing for differences in functional connectivity between the groups using multivariate F tests in a voxel-by-voxel manner (Arnold Anteraper et al., 2019;Westfall et al., 2020;Whitfield-Gabrieli & Nieto-Castanon, 2012) (Figure 1c).
One of the main advantages of the SVD approach is that it is useful for summarizing the enormous number of voxel-to-voxel correlations into major directions of variation that capture the greatest amount of functional connectivity (Caffo et al., 2010;Friston, 1994). Upon using multivariate analysis approaches such as PCA/ICA/SVD, the first few principle components are typically extracted and subsequently used as the estimates of connectivity (Worsley et al., 2005). A common approach is to scale the number of dimensions with the dataset size (Nieto-Castanon, 2022), such that relatively small eigenpatternsto-N ratios (e.g., 1:5) are suitable for detecting large effects in small samples, whereas larger rations (e.g., 1:20) are advisable for larger sample sizes. We opted for a 1:10 ratio as a fit to our data size (n = 38), yielding the choice of four dimensions (Argyropoulos et al., 2019;Morris et al., 2021).

| SCA
Because MVPA is an omnibus test that informs on differences between groups but not on the source of the differences (Arnold Anteraper et al., 2019;Morris et al., 2021), we performed a complementary post hoc analysis to further probe the nature of putative group related differences in connectivity patterns of the clusters that originated from the MVPA procedure (Beaty et al., 2015;Eckstein et al., 2022;Mateu-Estivill et al., 2021;Tortora et al., 2019). To characterize the differences between the chronological-and reverse-order conditions, we performed SCAs, using the resulting MVPA regions as seeds. The SCA was performed by computing Fishertransformed correlation coefficients between the nonsmoothed mean seed time courses and all other voxels and subjecting them to between-group comparisons.

| Eye tracking
Eye tracking was performed during scanning using an EyeLink 1000 system (SR Research, Canada) installed at F I G U R E 1 Schematic representation of the functional connectivity multivariate pattern analysis (fc-MVPA). (a) The functional connectivity between each seed voxel and all other voxels is concatenated across subjects. (b) The single-voxel functional connectivity matrix of subjects by voxels is subjected to singular value decomposition (SVD), where each column represents the explained variance in functional connectivity in descending hierarchical order of explained variance. In the current study, the patterns of seed-based correlations across all subjects are represented by the four leading components, which are referred to as eigenpatterns. Each row thus contains a subject-specific set of four eigenpattern sores. (c) Finally, upon completing the SVD procedure for each and every voxel, a second-level multivariate F-test analysis was performed on the eigenpattern scores, under the null hypothesis that the connectivity patterns between the chronological-order and reverse-order groups are equal. This process is then repeated for every source voxel to identify regions that show brain-wide betweengroup differences in functional connectivity. Chron., chronological; Rev., reverse; subs, subjects; SVD, singular value decomposition the rear end of the scanner. Gaze positions and pupil diameter (PD) were sampled with a frequency of 500 Hz. Calibration was performed prior to each scan. Due to technical issues of data quality, eye tracking was successful in 26 participants (13 from each group). For the purpose of this study, PD was analysed by calculating for each participant the normalized PD change (z scores) for each scene and interscene blank period. These values were subsequently averaged across participants of each group and plotted to assess physiological responses during the unfolding of the movie and blank segments.

| RESULTS
Participants were scanned in an fMRI environment while viewing movie scenes that were presented either in their chronological order (chronological-order group) or in reverse order (reverse-order group), thus comprising or omitting a fundamental feature of narrative coherence, that is, the causal connection between events (Figure 2a). To confirm that participants in both groups remained focused in the task and tracked the presented stimuli, we analysed pupillometry measurements during both scene presentation and interleaved blank periods. Across all scenes, both groups showed significantly enhanced pupil size during blank versus scenes periods, as expected from pupil responses to decreased luminance in the blank epochs (repeated-measures ANOVA) using stimuli type (scenes/interscene blanks) and segment number (scenes/ blanks 1-22) as within-subject factors and condition (chronological/reverse order) as between-group factor (F (1) = 147, p < .0001, see Figure 2b,c). Additionally, pupil responses did not differ between the groups both during scene presentation (F(1, 22) = .65, N.S.) and during interscene blanks (F(1, 22) = .12, N.S., see Figures S1 and S2 for mean pupil dilation for each group across the entire timeline and for each scene separately). These results indicate that participants of both the chronological-and reverse-scene groups visually tracked the presented stimuli throughout the experiment.

| Narrative understanding
As indicated by their self-reports in the post-experiment questionnaire (see Appendix S1), all participants in the reverse-order group indeed noticed that the scenes were not presented in their correct order (not necessarily in reverse order, see Table 1). Nevertheless, most of them (14/18) stated that they comprehended the story and with high confidence (see Table 1). Despite their subjective reports, compared with 94.4% (17/18) of participants in the chronological-order condition, only 44.4% (8/18) of the reverse-order condition correctly indicated the roles of the three main characters. It is noteworthy that the questionnaires were answered retrospectively, such that participants may have attempted to create a coherent F I G U R E 2 Experimental design and pupillometry. (a) Twenty-two scenes from the film 'Bicycle Thieves' were presented during the experiment, interleaved by blank screens (10 s). The chronological group were presented with scenes in their natural order (from beginning to end), and the reverse-order group viewed the scenes from last to first. (b) Mean pupil size for an example scene for the chronological (cyan) and reverse-order (grey) conditions. The dashed line indicates the transition from scene to blank periods. (c) Mean pupil size across all scenes (left) and all blank periods (right), shown separately for each condition. Pupil dilation was significantly greater in blank versus scene segments (p < .0001), yet no differences were found between the chronological-and reverse-order groups in neither scenes nor blank periods. narrative only after the experiment had ended. As an indication of causality comprehension, the participants were asked about the causal connection between two consecutive scenes that were presented to both groups at roughly similar latencies from the experiment's onset. In these scenes, the main character treats his child to a fancy meal as a compensation to an aversive event that happened in the previous scene, where the father slapped the child in anger. Only one participant of the reverse group answered correctly, compared with 14/18 participants of the chronological-order group, who correctly understood (and remembered) the causal connection between the two events.

| Differential functional connectivity between chronological-and reverse-order groups during scene presentation
We performed MVPA to delineate brain areas with differentiable functional connectivity measures during chronological-versus reverse-order conditions. The analysis was performed separately for the 'scene' and 'interscene' (blank) periods and yielded significant differences for both. The per cent of explained variance for the first four MVPA components was 87.54 (see Figure S3 and Table S1 for details). During the scene viewing periods, the between-group MVPA comparison yielded effects (p < .005, cluster-size p-FDR-corrected p < .05) in the supracalcarine (MNI x, y, z peak coordinates 4 À86 10) and in cerebellum crus II (À32 À74 À38) (Figure 3a). To characterize the underlying functional connectivity patterns of the detected MVPA clusters, we computed the functional connectivity profiles that discerned between the two groups with the detected clusters, by using these clusters as seeds in a SCA (Figures 3b and S4). For each seed (defined as a 10 mm radius sphere around the MNI coordinates of the MVPA cluster's peak activation), we computed its functional connectivity with all other voxels in the brain for each subject and compared the mean results between the groups. This procedure provided a F I G U R E 3 Between-group MVPA results and seed-based correlation analysis during scene periods.
(a) Between-group differences in wholebrain functional connectivity MVPA during scene periods, showing effects in the right supracalcarine (top) and cerebellum crus II (bottom). (b) Results of seed-based correlation analysis (SCA), using the regions outlined in (a) as seeds (see also Figure S4). map of regions that displayed differential functional connectivity patterns during scene viewing for chronological versus reverse order (p < .005, cluster level: p < .05, FDR corrected). For the calcarine seed, we found stronger functional connectivity in the chronological group with the posterior precuneus and cuneus (see Figure 3b and Table 2). Performing the same contrast using the cerebellum crus II seed yielded a distributed array of cerebral regions, including bilateral superior parietal lobule, bilateral anterior prefrontal cortices (BA 44 and 46), and bilateral inferior temporal cortices, largely overlapping the frontoparietal network (Figure 3b and Table 2). Applying the same analysis for reverse versus chronological (reverse > chronological) yielded no significant results.
To uncover the underlying patterns of connectivity that differed during the blank periods between the groups, we performed a SCA, using the clusters delineated by the MVPA during the blank periods as seeds. The technique was identical to the one applied for the scene periods (see above). No significant differences emerged from the comparison of chronological-versus reverse-order contrast (chronological > reverse). Conversely, the contrast of reverse versus chronological (reverse > chronological) yielded differential functional connectivity with all the MVPA seeds (Figure 4b and Table 3). In particular, this analysis yielded consistent functional connectivity in regions implicated in the ToM network. These regions, consisting of superior temporal gyrus and sulcus, TPJ and precuneus showed offline heightened correlations with the seed regions. As depicted in Figure 4, portions of the precuneus, and particularly its dorsal-anterior part, were found to be functionally connected with almost all of the MVPA see regions during blank periods (Figure 4b).

| DISCUSSION
Narratives are central to human experience, enabling the representation and communication of complex constructs, ranging from personal memories to cultural identities (Brockmeier, 2002;Hirst & Manier, 2008). Nevertheless, the neural mechanisms that underlie the formation of narratives are still largely unknown. In the current study, we tackled this issue by using an edited version of a full-length movie, which encapsulates core features of narrative characteristics in a cinematic medium, namely, a series of events portraying a protagonist paving his way towards a goal and has disruptions and obstacles in his way towards that goal (Mar, 2004;Tomasulu, 2007).
Our results point to four main findings that relate to differences in functional connectivity during chronological-versus reverse-order conditions. During scene viewing, frontoparietal regions showed heightened functional connectivity in the chronological-compared with reverse-order conditions, as well as the posterior precuneus. Two functional networks showed heightened functional connectivity in the reverse-order condition, only during blank periods; the first is a network comprising of STS, STG, TPJ and prefrontal regions, overlapping the 'ToM' network previously described ( Mar, 2011). The second consists of regions along the visual processing hierarchy, including the calcarine sulcus, lingual gyrus and fusiform gyrus. The precuneus, both anterior and ventral, was also prominent during blank periods in the reverse-order condition. The difference between the conditions (chronological/reverse order) was designed to manipulate a core component of narrative, namely, the causal connection between the events that form the narrative's storyline. Reversing scene order was designed to eliminate the temporal precedence that allows for establishing causal connections between unfolding events. In both conditions, chronological and reversed, the scenes were interleaved by short segments of blank screens, which were introduced in order to examine neural processes that may take place 'offline', subsequent to the online presentation of the events themselves. We therefore performed all analyses separately for 'scene' and 'interscene blank' periods, thus shedding light on both online and offline processing of narrative formation. F I G U R E 4 Results of between-group MVPA and reverse versus chronological seed-based correlation analysis during 'blank' periods. (a) Between-group differences in whole-brain voxel-by-voxel functional connectivity during blank periods, showing effects in bilateral calcarine sulcus, ventrolateral prefrontal cortex, supramarginal cortex, fusiform cortex, lingual gyrus, putamen and cingulate gyrus (height threshold: p < .005, cluster threshold: p < .05 FDR corrected). (b) Results of seed-to-voxel correlation analysis for blank periods, using the regions outlined in (a) as seeds overlaid on surface templates in midsagittal and lateral surface planes T A B L E 3 MVPA seeds and reverse versus chronological seed-based correlation analysis results during 'blank' periods As stated at the outset, a primary feature of coherent narrative formation relies on causal inferences among successive units of information (Graesser et al., 1994). Accordingly, interfering with causality by breaking temporal order on a large temporal scale impeded the capacity to construct a coherent narrative. The failure of participants in the reverse-order condition to form a coherent narrative, despite their effort to bind otherwise fragmented events, coincides with the centrality of temporal order in forming narratives of prolonged time scales. Moreover, it stresses the substantial role of temporal succession during the actual (rather than reconstructed) experience in causality inference and coherent narrative formation (Briner et al., 2011;Tylén et al., 2015). Breaking the temporal causation between successive occurrences is bound to lead to differences in the overall experience, presumably in features such as attention allocation, fatigue, cognitive engagement and perhaps frustration. The written descriptions from both groups contained information from the entire scene timeline, and pupil dilation was synchronized with scene onsets throughout the whole experiment, indicating that both groups visually tracked the movie scenes (Hopstaken et al., 2015;Lee & Margolis, 2016), yet only the chronological-order group succeeded in forming a coherent narrative. As previously shown in the comprehension of texts, changing the temporal order of cause and effect does not necessarily prevent their retroactive understanding, yet does affect processing time (Briner et al., 2011) and narrative comprehension (Ohtsuka & Brewer, 1992).
This explorative study focused on delineating functional connectivity profiles between groups based on multivariate analyses, followed by seed-to-voxel functional connectivity, based on the regions yielded by the MVPA. MVPA is an omnibus test that does not provide information regarding directionality or interconnections among the regions it extracts, thus typically requiring complementary investigation. We found that increased functional connectivity in a frontoparietal network in the chronological condition during scene presentation, detected through its connections with the cerebellum crus II. The frontoparietal network was previously shown to coactivate with the cerebellum's crus II (Caulfield et al., 2016), an area considered as a supramodal zone of the cerebellum (O'Reilly et al., 2010). Serving as a key control system, the frontoparietal network supports information integration in diverse cognitive contexts (Naghavi & Nyberg, 2005) and was shown to play a role in context-dependent narrative comprehension (Smirnov et al., 2014). Corroborating our findings, frontoparietal regions were found to correlate with chronological versus inconsistent information in a narrative comprehension task (Ferstl et al., 2005), an effect attributed to working memory demands. We suggest that in the context of narrative construction, the frontoparietal network is engaged in online processing of incoming streams of information, maintaining and integrating it across prolonged temporal scales, allowing causal inferences based on previous events.
A key finding in the current study is that the reverseorder condition yielded heightened functional connectivity in a large and consistent set of regions during the interscene periods. The recurring regions include the lateral temporal lobes (STG, STS), TPJ, and inferior frontal gyrus (IFG), which have collectivity been reported to underlie social cognition functions (Brunet et al., 2000;Gallagher et al., 2000;Vogeley et al., 2001). In his Poetics, Aristotle stated that narrative is the mimicry of action (mimesis), that is, the ability to simulate someone else's behaviour, and to identify and empathize with them. This early notion regarding the relationship between mimicry and narrative organization resonates with our findings that regions of the mirror system are involved in understanding the goals of others by interpreting their actions (Buccino et al., 2007). This feature may be particularly critical in understanding prolonged sequences of social events, which their understanding relies heavily on Note: Seed regions, derived from the MVPA results, are shown on grey background. Below each seed entry are brain regions that showed increased functional connectivity with each seed region for reverse-versus chronological-order conditions, during 'blank' periods.
tracking the protagonists' behaviours, as in the current experiment. Tracking the occurrences of natural occurrences (such as movies) that focus on human characters and their interactions requires one to track others' behaviours, to interpret their intentions and goals, and often elicit complex emotional responses (Jaaskelainen et al., 2021). These unique processes are fundamental to social cognition and are key to forming narratives that are based on information conveyed by others' behaviours and interactions. ToM abilities, which are a hallmark of social cognition, enable one to attribute internal mental states to others, interpreting their feelings, beliefs and emotions (Baron-Cohen, 2001;Happé, 1993;Saxe et al., 2004). On the neural level, tasks that demand ToM have been shown to recruit the precuneus, medial prefrontal cortex and bilateral temporoparietal cortex (Molenberghs et al., 2016;Schurz et al., 2020). Action observation, a more basic feature of social cognition, is correlated with activations in occipito-temporal regions, as well as temporoparietal and inferior frontal gyrus. Our findings imply that neural networks involved in social cognition may serve to interpret human behaviour, goals and interactions in an offline manner, after the occurrences have taken place. In our study, these networks are particularly apparent when confronted with naturalistic events presented out of order during the interscene buffers, possibly signifying the retroactive effort to make sense of the protagonist's intentions and interactions and to embed them within a broader context or narrative. The precuneus in particular showed heightened functional connectivity in the reverse-order condition during interscene periods. Beyond its role in understanding others' intentions (Mar, 2011), the precuneus is considered to play a pivotal role in other high cognitive functions (Margulies et al., 2009), including retrieval of episodic memories, simulating future scenarios and imagery (Buckner et al., 2008;Stillman et al., 2017). Its activation during the processing of a continuous story has previously been attributed to its capacity to process relatively long and meaningful chunks of information (Lerner et al., 2011), as well as in segmentation and integration of present occurrences with the recent past (Yazar et al., 2014;Zacks, 2010). As part of the DMN, the precuneus is imperative in binding and integrating incoming information with semantic knowledge and episodic memories (Cooper & Ritchey, 2019;Lahnakoski et al., 2017;Uddin et al., 2009;Utevsky et al., 2014). Interestingly, a previous study has demonstrated the involvement of the precuneus, especially when narrative unfolds across prolonged timescales (Tylén et al., 2015). In the context of the current study, the precuneus seems to be particularly active offline, possibly signifying the binding of previously acquired information into a coherent narrative.
In conclusion, narrative construction is a key feature of human experience. The current study provides new insights into the processing of information that unfolds over long durations of time, unearthing wide-range cerebral coactivity that subserves online and offline narrative construction.