Tracking Emotional Shifts During Story Reception: The Relationship Between Narrative Structure and Affective Responses

Participants

We used a linear mixed-models analysis with random effects (random intercepts) of participant for analyzing the data (see section on Data Analysis Strategy below) and based the estimation of the required sample size on the method proposed by Westfall, Kenny, and Judd (2014), as implemented in the accompanying web-based app (https://jakewestfall.shinyapps.io/crossedpower/). For the sample size estimation, we assumed that the variance components would distribute equally between the random effects of participants (0.5) and the residual term (0.5). All other variance components were assumed to be 0 since the random intercept of participants was the only random effect in the model. We further assumed the magnitude of the effect size (Cohen’s d) to be 0.6 and specified the Type-I error probability α to be .05 (two-tailed) and the desired power 1-β to be .80 for pairwise comparisons between story parts. Given these specifications, the required sample size was estimated to be 17 participants per story. We chose to collect data by repeated measures, creating a sample that exceeded the required sample size per story by at least 15% because physiological measures often suffer from data loss.

We recruited 32 participants through the online participant platform of the psychology department at the University of Würzburg, Germany. One participant was excluded due to intoxication prior to the experiment, leaving a total sample of 31 participants (14 women) for the analyses (age: M = 24.97 years, SD = 7.72; Min = 18 years, Max = 51 years). Participants were predominantly students (n = 24) and native German speakers (n = 30). Experimental sessions had a mean duration of 1.5 hours and were conducted individually for each participant in two laboratories at the University of Würzburg. They were each paid between €10 and €12.50 depending on overall preparation time.

Text Materials

Three different short stories were selected from searches in German short-story forums on the Web that matched emotional arcs with three distinct emotional shifts: a happy-ending story, a tragedy, and a thriller. We partly changed and shortened the original stories to accommodate the basic genre structures. A comprehensive description per plot unit is provided in Table 1. The happy-ending story was a science-fiction story about the survival of bees in a post-apocalyptic world, narrated in third-person perspective (Frambach, 2018, 1691 words, 12:29 min). The tragedy was a short story about a bullying incident during a school trip, narrated in first person (Praetorius, n.d., 819 words, 5:54 min audio). The thriller story was a story about a sudden attack on the tram, narrated in third person (Eisenmann, 2008, 1641 words, 11:39 min audio). Using the R package syuzhet version 1.0.6 (Jockers, 2017) and the German sentiment lexicon SentiWS version 2.0 (Remus et al., 2010), we conducted sentiment analyses for each story and unit. Figure 1 shows sentiment scores over story units to represent valence trajectories of our stimulus stories.

Figure 1

In preparation for our experiments, we verified that the stories were similarly easy to read by computing Flesch-Reading-Ease scores (adapted scoring for German language according to Amstad, 1978; Flesch, 1948). They corresponded to middle-to-low reading difficulty levels. Every story was audio recorded to ensure timed mapping of online measures during presentations. One neutral text was used to compare with the effects of the emotion-laden stories (Kreibig et al., 2007; Piferi et al., 2000). We used an expository text about apples (Buth, 2014, 500 words, 3:28 min audio).

Design

Participants listened to two of three possible emotional stories in random order after listening to the neutral text, creating six possible story and order combinations that were balanced over the 31 participants. They split into five groups of five and one group of six over these combinations, resulting in a total sample size of 21 participants for the happy-ending story and the tragedy and 20 participants for the thriller.

Procedure

After obtaining written informed consent, participants were seated upright in a chair in front of a PC equipped with headphones. Their left hand, which was their non-dominant hand in 90% of cases, rested on an armrest at the table. Participants were to operate the mouse with their right hand during the questionnaire periods between story presentations. Physiological sensors were attached, and physiological signals were continuously sampled during the experimental session. The experimenter sat in the same room outside the participants’ field of vision but was approachable in case of problems or questions.

We used Inquisit 5 (Millisecond Software, Inc.) as the experimental presentation software. After completing a short questionnaire, including demographics and control questions, participants were instructed to find a comfortable seating position and to avoid any unnecessary movements and speech during the tasks, relaxation periods, and stories. They were instructed to keep their eyes on the screen because their facial expressions were recorded by a webcam.

Before any stimulus presentation, participants completed a paced breathing task similar to the task used by Butler et al. (2006), in which they were asked to look at a sinusoidal wave passing through a vertical line onscreen. Participants were required to breathe in whenever the wave rose and breathe out whenever it fell. A short, repeatable practice trial ensured their understanding of the instructions. The waveform pattern was designed to induce a respiratory frequency of nine cycles per min and was presented for 2:15 min. This measurement period was later used for assessing individual differences in resting RSA (Grossman et al., 1990; Wilhelm et al., 2004).

Following a 3-min quiet sitting period that served as the resting baseline for autonomic nervous system measures, participants listened to the neutral text that served as a non-emotional, stimulated control period. Next, participants listened to two of three possible short stories in random order, each following a separate unstimulated baseline period of 3 min. Every story was preceded by instructions to listen carefully and to keep their eyes on the screen where a neutral visual animation of the narrator’s voice was shown (for an exemplary screenshot, see online supplement S1). Immediately after each story, participants answered a set of self-report items. They were asked to rate their emotional experience for each story unit with the help of self-assessment manikins (SAM; Bradley & Lang, 1994; Lang, 1980) and a maximum of three out of a list of 15 emotions. A set of nine items targeting the general experience of emotional shifts was sampled for unrelated analyses. Additional SAM ratings of state valence and arousal were obtained after each resting baseline before participants listened to the next story. After the session, participants were debriefed and reimbursed.

Measures

Although every story could be separated into the four structural units of introduction, complication, climax and end, references of story units for the subjective ratings differed between genres. In the thriller story, we deemed it difficult for participants to mentally separate the complication and climax part after listening to the story. Therefore, we obtained ratings for only the introduction, one merged plot unit of complication and climax, and the end part of the story (compared to all four units in the other genres). The wording of each unit reference per story is provided in an online supplement (S2).

Data Scoring

Raw scores of SCL, heart rate, and respiration rate were extracted as the arithmetic mean of the respective story unit, HRV raw scores were obtained by spectral power analysis. Baseline periods were 3:00 min in length, the neutral text had a total duration of 3:30 min, and story units differed between 0:50 and 4:07 min in sequence length over the stories. To make these values comparable despite the underlying differences in recording time, each raw score was corrected for the respective length of each sequence, resulting in parameter values per second. Note that using period-length-corrected mean scores of the physiological data likely leads to conservative estimates of emotional responses.

One data set was excluded from the physiological data analyses because of severe disruptions during baseline recordings. Problems with data collection or non-responding left a total sample of 23 data sets for SCL analyses, 30 sets for heart rate and HRV analyses, and 26 sets for respiratory analyses, with at least one usable story or a minimum of 50% usable story units.

Data Visualization

For the visualization of psychophysiological data (based on different measurement units), we standardized response scores after Kreibig et al. (2007). We subtracted the prior baseline score from the raw score of the respective story unit. The paced breathing task served as baseline for RSA. Individual differences in each person’s overall mean were eliminated by centering around the person’s mean separately for each parameter. RSA parameters were additionally divided by respiration rate. Finally, all scores were subjected to a z-normalization transformation separately for each story (and the preceding neutral text) and mapped onto a T scale with a mean of 50 and a standard deviation of 10.

Data Analysis

Raw scores were analyzed with linear mixed-effects models by using the lmer function of the R package lme4 version 1.12 (Bates et al., 2015). Linear mixed models are the method of choice for analyzing our data due to the cross-dependencies that originate from one person listening to two of three possible stories with multiple measures per story and random missing values (stories nested within participants; Judd et al., 2017; Richter, 2006). For each parameter and story, we used a linear mixed model with random effects (random intercepts) of subjects and fixed effects of baseline level (grand-mean centered) and story part (dummy-coded), comparing every subsequent part to the introduction. The models for RSA differed by including RSA levels of the paced breathing period as baseline and respiration rate as an additional predictor, both grand-mean centered. We further used least-square means (lsmeans package version 2.3; Lenth, 2016) for multiple comparisons with the Kenward-Roger approximation of degrees of freedom and Bonferroni-Holm method for p value adjustment to test differences between the remaining story parts.

Availability of Data, Code, and Materials

Data and R-code for the analyses reported in the present paper are available at the repository of the Open Science Framework (https://osf.io/2wxze/). The experimental materials are available from the authors upon request.

Research Question 1: Narrative Arcs and Subjective Emotional Responses

Subjective emotional responses were acquired after the stories. Participants were given a short description of the particular story unit to guide their recall and asked to rate their emotional experience during listening. They had to choose the three emotions they felt most intensely during the respective story unit and rate their intensities. Intensity scores were computed across all rated emotions per unit. The respective emotions selected serve as descriptive context.

Research Question 2: Narrative Arcs and Psychophysiological Emotional Responses

The subjective experience of arousal is a time-delayed and interpretative component of emotional experience. Psychophysiological activation, on the other hand, represents an immediate stimulus-related response of our body. The activation of the sympathetic and the parasympathetic branch of the autonomic nervous system was measured, each by means of two parameters. The skin conductance level (SCL) and the sympathetic part of heart rate variability (in the following sympathetic power) are measures of sympathetic activation—that is, arousal in the activating sense. The two measures of parasympathetic activation were both obtained from heart rate variability: parasympathetic power and the respiratory sinus arrhythmia (RSA) parameter, both representing attenuating, regulatory activation. The overview across the stories (Figure 4) showed a high degree of consistency between the parameters within a story.

Based on the previous findings, an increase in arousal was to be expected for the happy-ending story well before the end unit, since the ratings of hope and increasing valence already set in earlier than the sentiment analysis suggested. Indeed, we found the most significant activation increase over all physiological measures from the complication to the climax unit (Figure 4a and Table 3). Despite the highest suspense score in the introduction unit, M = 4.80 (SE = 0.36) of 50% of the participants, we found a relative deactivation compared to the neutral story. In the context of subjective ratings, we could attribute this activation pattern to sadness (Kreibig, 2010), but a link to compassion would also be conceivable. Thus, it seemed plausible that sympathetic activation associated with suspense was overshadowed by dampening negative emotions. Accordingly, this would not indicate a negativity bias in the subjective ratings of the introduction unit (if we compared them to the sentiment scores), but rather that the sentiment analysis did not capture the aggregated sentiment of the introduction in a way that was representative of the audience’s feelings. In comparison, we found a high correspondence of the suspense ratings to the physiological activation in the climax unit (Table 3). Here, we notice a clear double activation of both nervous system branches that is frequently found in states of stress or when facing a challenge (Berntson et al., 1991; El-Sheikh et al., 2009).

In the tragedy, a high activation level in the introduction unit was initially noticeable in comparison to the neutral story (Figure 4b). In this story, suspense was also rated highest at the beginning, but here in connection with happiness and amusement, which are physiologically associated with high activation instead of curbed activation (Kreibig, 2010). The following shift to dominantly negative emotions such as sadness and compassion resembled the introduction unit of the happy-ending story (Figure 4; Table 3). Another parallel was the significant double activation towards the climax, which in this story also fits the increased ratings of arousal and emotional intensity. Towards the end, vagal power particularly increased. Although suspense subjectively decreased here, we found a rise in physiological activation in all parameters. The outstanding increase of vagal power could indicate frustration and a heightened control of negative emotions, which seems plausible in terms of the tragic ending.

In the thriller story, subjective ratings suggested an arousal increase in one of the middle story parts, associated with the emergence of fear (Kreibig, 2010). Consistent with the other two stories, we also found a strong significant increase in physiological activation in all parameters toward the climax unit (Figure 4c and Table 3). This corresponded highly with subjective ratings of arousal, emotional intensity, and suspense. The collective drop in physiological activation at the end of the story also clearly corresponded to the subjective ratings of relief and drops below the baseline level of the neutral story.

In sum, we found that physiological activation corresponded more with our structural unit classification and with activation patterns associated with dominant subjective emotions and less with suspense ratings. Across all genres, there was a consistent activation toward the climax unit. The activating dominant emotions of the stories (e.g., happiness, fear) were indeed most often associated with higher physiological activation (mostly of both nervous system branches). The exception is the end of tragedy, where we suspect increased regulatory activation. In terms of excitation transfer, the high correspondence of increased physiological arousal with emotional intensity ratings may suggest that it happened within our narrative units rather than between them.

The Research Questions

We posed two research questions: Research Question 1 addressed the extent of correspondence between each story’s narrative arc (as mapped by the sentiment analysis) and audiences’ self-reported emotional responses (valence, arousal, emotional intensity, distinct emotions). Research Question 2 addressed the extent of correspondence between each story’s narrative arc (as mapped by the previous analyses) and audiences’ psychophysiological emotional responses (in sympathetic and vagal autonomic nervous system activity).

Apart from the limitations of sentiment analysis, which we discuss in detail below, a more differentiated subjective experience of valence was evident in the ratings. Whenever a story section contained within itself a dynamic variation of valence, as in the introduction of the happy-ending story, or the literal content of sections was ambiguously explainable, as in the thriller story, prominent deviations of the rated valence from the sentiment analyses were noticeable. We additionally noticed higher salience of negatively arousing emotions in the subjective ratings than positively arousing emotions. Retrospective emotional responses are constructed by memory and interpretation and therefore commonly affected by a negativity bias, which has been widely demonstrated in previous emotion studies (e.g., Baumeister et al., 2001; Bohanek et al., 2005; Egidi & Gerrig, 2009). This effect is particularly evident in our findings when comparing the emotion ratings with the arousal ratings. Intense and negative arousing emotions were more consistent with the separate arousal ratings than intense and positive arousing emotions.

Generally, we found a higher correspondence of subjective valence trajectories and sentiment arcs in the genres that implied shifts between happiness and sadness, the happy-ending story and the tragedy compared to the thriller story. Subjective valence ratings corresponded much more consistently with structural elements such as the story climax. This is in line with preceding research on suspense trajectories over structural story patterns (Boyd et al., 2020). The immediate, physiological responses also confirm this: Consistently across all stories, we found sympathovagal double activation in the climax unit.

Within each story, we found very homogenous trajectories of the psychophysiological parameters, which speaks for the reliability of our method. In relation to the preceding analyses, we found that physiological activation does not correspond strongly with ratings of suspense at any given narrative unit. Although the data was highly consistent in the climax units between genres, we found high variance at other points that were rated as particularly suspenseful, namely the introductions of the happy-ending story and the tragedy. There, it seems, the sympathetic activation that is associated with an intense feeling of suspense is overlayed by activation patterns of other dominant narrative emotions, such as happiness and sadness. In terms of excitation transfer, our findings support the idea that cumulation of arousal happens in successions of arousing states, but not from affective states of low arousal. Thus, future studies on excitation transfer in narrative reception should pay attention to emotional successions and the excitation potentials of dominant emotions apart from suspense. Furthermore, we sought to connect occurrences of high physiological activation with higher ratings of emotional intensity, as we would expect based on excitation transfer theory. We found rather little support for this between story units. No particularly high emotional intensity was rated in the end units after evidently arousing climax units. However, this does not necessarily mean that excitation transfer had not occurred within the respective climax unit. Our method is based on measures of aggregated physiological activation within story units. Here, we found high correspondence of increased physiological activation with increased emotional intensity. Therefore, we cannot draw a firm conclusion with the unit division that we used.

Limitations

Despite our substantial findings, at least four limitations of the present study must be noted. First, we were able to use only one story for each of the three narrative genres. Thus, whereas our assessment regarding emotional responses was rather broad and multifaceted, our stimulus set was limited. Future researchers are encouraged to include multiple stories per genre and further controlled (or experimentally manipulated) properties, such as narrator perspective.

Second, although the sentiment analysis we used offers an easily implementable approach to comparing textual and individual valence trajectories, it is also limited by its word-for-word evaluation. In sections of possibly mixed or rapidly changing valence, the reduction of complexity becomes particularly apparent in aggregations over several sentences. However, it is also striking that the valence ratings of the audience can lead to markedly different conclusions regarding the overall valence rating of a story unit than the mere summation of the sentiments would lead one to expect. In our opinion, this shows that valence experience is more than the sum of its parts, and that research on emotional flow could profit from more elaborate text-analytic methods. We encourage future research to test different text-analytic approaches.

Third, we examined self-reported emotional responses on an aggregated level of prototypical plot units and forewent the investigation of moment-to-moment responses. Examining emotional responses as they occur is challenging, alternative approaches such as real-time-response measurement (Winkler et al., 2022) or self-probed emotional retrospections (Winkler et al., 2023) have their own limitations. That said, an investigation of smaller story segments or even moment-to-moment analyses would further clarify our findings. It should be noted, however, that reducing the size of the story segments will be especially difficult in terms of comparability of physiological and subjective data. Participants need mental landmarks for their retrospective judgements. With reduced segment sizes comes the obstacle of reduced possibilities for guidance into comparable mental units. In our study, for example, the subjective measurements for the thriller story did not differentiate between the complication and climax unit for this reason. This, in turn, made it more difficult to determine where exactly the overall subjective valence shifted in this story. In this case, our determination was based on substitutional information of arousal profiles of the dominant underlying emotions.

Fourth, our methodological approach did not allow us to clearly differentiate between narrative and aesthetic emotions. The emotion word list included concepts that are more or less likely to be fiction- rather than aesthetics-related, and per instruction it explicitly referred to the story content. We therefore assume the retrospective measures to be mostly fiction-related but cannot rule out an unwanted influence of artefact-related emotions as a byproduct of aesthetic appraisals of the story (Menninghaus, Wagner et al., 2019; Menninghaus, Schindler et al., 2020; Tan, 2013; Wassiliwizky et al., 2017). Future research could profit from a more detailed assessment of comprehension-related affect, such as suspense, curiosity, surprise (as suggested by Brewer & Liechtenstein, 1982), or confusion (as suggested, e.g., by D’Mello & Graesser, 2014). Clarifying the role of narrative comprehension for emotional flow could broaden our understanding of the relationship between narrative structure and emotional responses considerably.