<biblStruct> <analytic> <title level="a">A Fairy Tale Gold Standard. Annotation and Analysis of Emotions in the Children’s and Household Tales by the Brothers Grimm

We present a quantitative study of textually encoded emotions in a core set of the Grimms’ Children’s and Household Tales. As a contribution to Computational Literary Studies, we publish (a) a fairy tale corpus (ChildTale-A) with more than 5,000 manually annotated sentences and introduce (b) four aggregated measures for the analysis of textually encoded emotions (Average Valence, Emotional Potential, Emotional Arc, and Emotion Profile), with which we (c) analyze the corpus with regard to the purported cruelty vs. optimism of fairy tales. On average, the fairy tales contain more than 50 % emotional sentences without clear negative sentiment, while emotion trajectory patterns vary. Together, these findings underscore the role of emotions as plot-driving elements in fairy tales as a highly schematized historical genre.

Wir stellen eine quantitative Studie textuell enkodierter Emotionen eines Kernsatzes der Grimmschen Kinder- und Hausmärchen vor. Als Beitrag zu den Computational Literary Studies publizieren wir (a) das ChildTale-A Korpus mit mehr als 5.000 manuell annotierten Sätzen, und stellen (b) vier aggregierte Maße zur Analyse textuell enkodierter Emotionen (durchschnittliche Valenz, Emotionspotenzial, Emotionsverlauf und Emotionsprofil) vor, mit denen wir c) das Korpus bezüglich mutmaßlicher märchentypischer Grausamkeit oder Optimismus analysieren. Die Märchen enthalten im Schnitt mehr als 50 % emotionale Sätze ohne eindeutig negative Tendenz sowie diverse Muster im Emotionsverlauf. Zusammen unterstreichen diese Befunde die Rolle von Emotionen als handlungstreibende Elemente im Märchen als stark schematisiertem historischem Genre.

1. Introduction

In this contribution to Computational Literary Studies (CLS), we take a quantitative approach to investigate textually encoded emotions in the Animal and Magic Tales (German: Tier- und Zaubermärchen, short: AMT), a subset of the Brothers Grimm’s Children’s and Household Tales (German: Kinder- und Hausmärchen, short: CHT). These ›traditional‹ fairy tales that were published as folk tales have been criticized for the cruelty that is applied by and to their characters, Cf. Röhrich 2016; Schenda 1977, p. 359. and, by extension, for the negativity of their content. Yet, at the same time, as suggested for example by Herrmann Bausinger, fairy tales count as an optimistic genre, typically coined by happy endings and essentially fortunate characters. Cf. Bausinger 1983.

This paradox is not easily resolved. First of all, it needs to be taken into account that the Grimms’ fairy tales are handled as literary ›folk tales.‹ To a certain extent, the CHT can indeed be seen in continuation of antique and pre-modern narratives, but they were also re-written for a bourgeois readership of the 19th and early 20th Century. Since the tales were immensely popular among these readers, they apparently satisfied a set of specific needs, one of which is an entertainment function that draws on the repertoire of traditional folk narratives: the stories can be blatantly cruel, but are located in an idealized past; also, they are schematically optimistic. These features can relieve readers of different aspects of the demands of modernity. Cf. Neuhaus’s discussion of Schenda’s socio-historical approach: Neuhaus 2017, p. 30. Secondly, Bausinger argues that fairy-tale happiness does not describe a constant state of bliss, but is an effect that results especially from the downs in the stories’ ups and downs: it is drawn from the overcoming of the hardships, dangers and defeats.[F]airy-tale happiness is a contrastive phenomenon. It comes about because in this happiness we sympathize with the hardships overcome, the dangers and defeats that constitute the actual fairy tale plot. (Bausinger 1983, p. 19). (translation ours: [D]as Märchenglück ist ein Kontrastphänomen. Es kommt zustande, weil wir in diesem Glück die überwundenen Entbehrungen, die Gefahren und Niederlagen mitempfinden, welche die eigentliche Märchenhandlung ausmachen). Along those lines, broad research from anthropology, educational sciences, psychology, and literary studies has posited that the creative grounds for legends and fairy tales relate to motives of fear and need for help, to wishes and hopes, in short: to emotionally charged, and potentially negative, situations.Cf. Horn 2016a.

For our means, it is of high interest that emotions and feelings are a key ingredient of those stories. If emotions are seen as drivers of the narrative plot, Cf. Geck 2014. they should be observable at the textual level as an emotion potential, and not just arise in the reader in response to the text. Adopting a cultural-semiotic perspective, we consider textually encoded emotions as emotion potentials that are textual ›surface phenomena‹ (cf. Hillebrandt 2011, p. 78), and as such amenable to quantitative analysis. There are examples of such an approach, especially in psycholinguistic research, based on the idea that words associated with emotions reflect the actual emotional content (cf. Bestgen 1994, p. 32). So far, only few studies have systematically studied the range and diversity of ›textual‹ emotions, and to our knowledge, still fewer in a clearly delineated corpus of the CHT. For first studies using sentiment analysis tools, cf. Mohammad 2011 and Rotari 2018. It is thus an open question how emotions are distributed in this core corpus — are there overall more negative emotions, or more positive ones? And how do these emotions transition within the plots? What specific discrete emotions are particularly frequent in the fairy tales, and is there much variability across the corpus? Approaching the matter at the interface of cognitive psychology and computational literary studies, we thus ask how a core corpus (ChildTale-A) selected from the Grimms’ CHT textually encodes emotions.

To ask systematic questions about textually encoded emotion, Digital Humanities (DH) scholars can access various resources from computational emotion classification and sentiment analysis (SA). However, research on emotions in historical and literary texts has to face problems such as semantic context sensitivity and domain adaptation when working with texts of a specific genre, a particular time period, and languages other than English.For the most recent comprehensive overview on and critical evaluation of sentiment analysis applied to German literary texts, cf. Kim / Klinger 2019; for a recent application of modern sentiment analysis using transformer models to drama as one domain of German literary texts, cf. Schmidt et al. 2021. Fehle and colleagues run a comparison of 19 German sentiment lexicons and 20 sentiment-annotated corpora and conclude that compared to English lexicon-based resources which can achieve f1 measures above 70 % [...] the German resources perform rather poorly. (Fehle et al. 2021, p. 9). In the present study, we chose not to use ›out-of-the-box‹ sentiment analysis, as for example the syuzhet Package Jockers 2020b. For a review of sentiment analysis applied to story shape research, applications and critical discussion, cf. Elkins 2022. or existing sentiment dictionaries.For a comparison of sentiment dictionaries applied to diverse domains of German texts, cf. Fehle et al. 2021; for a comparison of widely-used sentiment dictionaries applied to a corpus of German literary texts, cf. Grisot / Herrmann, submitted. Instead, to pave the way for a domain-adapted study of the Grimms’ fairy tales and similar texts in German, we apply manual annotation at the sentence level to determine the texts’ emotion potentials in a selected sub-corpus (N = 80) of the Grimms’ CHT. Our annotation procedure implements the specific narrative schema of the fairy-tale genre and two main theoretical approaches from emotion psychology.

In doing so, we aim to contribute to CLS in three ways: Firstly, by introducing a gold standard of textually encoded emotion (the ChildTale-A corpus, see below, Section 3), which in the future may be used to improve supervised machine learning based sentiment and emotion analysis. Secondly, by using the data set for answering research questions about the emotional setup of the fairy tales. Thirdly, as a contribution to CLS methodology, we introduce four measures: Average Valence, Emotion Potential, Emotional Arc, and Emotion Profile which may be useful tools for the in-depth qualitative and quantitative analysis of textual encoded emotion potentials and its visualization.

2. Emotions in Fairy Tales

2.1 Fairy Tales

The Grimm’s Children's and Household Tales feature — next to the Luther Bible — among the most widely known books of German cultural history. As literary folk tales, they were registered by the UNESCO’s Memory of the World Registry in 2005, representing poetry of the human imagination in a universally valid form. The Grimms’ CHT are canonical children’s literature and are regarded, among other things, as an anthropologically relevant genre of evolutionary adaptation via enculturation and socialization, assumed to help fostering children’s emotional development and emotion regulation.Interdisciplinary research between pedagogy, literary studies and psychology suggests that literary reading is a central vehicle for acquiring emotional competence, cf. Petermann / Wiedebusch 2016. This idea has been specifically formulated for fairy tales, c.f. for example Lüthi 2004, pp. 105–114; Zipes 2006. From the vantage point of interdisciplinary affect studies, the CHT are interesting because they both ›represent‹ aesthetic-historical emotions and are likely to shape the understanding of emotions in (young) readers,As D’Arcens formulates for historical studies of emotions in literature: A key challenge [...] lies in developing an approach that identifies how these aesthetic-emotional features are also shaped by, and in turn shape, the emotional discourses and practices particular to their own place and time, yet with an awareness of how they continue to participate in a longer and larger history of human emotion. (D’Arcens 2020, p. 124). a point that, in view of constantly changing reading habits, needs to be regularly re-examined.

Fairy tales as folk talesWith their schematic structure, the fairy tales of the Grimms’ CHT collection are so-called ›folk tales,‹ defined by an idealized orality. As a genre, these are distinguished from ›art fairy tales,‹ which are more complex, clearly attributable to particular authors and a straightforwardly written genre. Cf. for example Neuhaus 2017. are among the ›simple forms‹ cataloged by André Jolles,Jolles 1930. i. e. narrative genres that are morphologically and typologically described as non-complex and that may correspond to anthropologically basic schemas that have evolved over a long history in many cultures. Along these lines, fairy tales have a reduced, simple style, as worldly [welthaltige] adventure narrative[s] of rasping, sublimating stylistic form.Lüthi 2005, p. 77 (translation ours). Formally, they are thus coined by repetition, semantic contrast, and formulaicity. In terms of narrative structure, they are plot-driven and realistically narrated (Max Lüthi’s ›welthaltig‹), despite the abundant fantastic elements. In terms of motives and content, fairy tales are defined as stories of ›lack and remedy,‹ ›problems and solutions,‹ ›seeking and finding,‹ that are outlined schematically and narrated ›flatly,‹ that is, with a single-strand, stereotyped plot, stereotyped settings, one-dimensional, i. e. not psychologized, characters, and simple world views.Cf. Lüthi 2016; Neuhaus 2017. It needs to be noted that the 7th edition of the Grimms’ fairy tales is not as reduced in style as the prior versions, using for example more generic ›folk tale formulas‹ such as ›once upon a time‹ and more detailed descriptions in the typical style of Wilhelm Grimm. They typically end well.The ›happy ending‹ is one of the prototypical features of folk tales. However, this is not uncontested. Even among the Brothers Grimm’s CHTs, there are several tales with no clear happy ending, cf. for example Uther, who lists seventeen CHTs with a ›bad ending’: Uther 2013, p. 612: schlechter Schluß (schlechter Ausgang der Erzählung) 26, 30, 35, 39, 41, 44, 80, 90, 92, 95, 117, 145, 150, 171, 185; 8 (1812), 175 (1840-50) (cf. Röhrich 1958).

With regard to their generic features, it is important to note that the purported ›orality‹ and the ›unadulterated‹ presentation of the tales in the CHT collection has long been demystified: The tales were co-composed by the collectors on the basis of different types of oral and even written sources,Cf. Bluhm 2022; Rölleke 2004; Uther 2013. The CHT were published in seven editions between 1813 and 1857 with subsequent alterations in content, style, and the number of tales included, cf. for example Uther 2013. They were increasingly de-sexualized, while introducing more humorous traits, christianizing the content and stylizing in the sense of a ›housefather literature.‹ which lead to the Grimms’ signature style and structure, which is most prototypical in the Animal and Magic Tales. This set of tales, which deviates from similar genres, such as ›Schwank,‹ legend and fable, but also from the more clearly authored ›art fairy tale,‹ can be retrieved in the catalog of motives first introduced by the Finnish school of systematic-empirical folk tale research.This school takes up the unfinished original research program of the Brothers Grimm, starting with the publication of Antti Aarne’s The Types of the Folktales in 1910, which was subsequently broadened by Stith Thompson in 1927 and 1961, as well as edited and supplemented by Hans-Jörg Uther in 2004 (Aarne 1961; Uther 2004). In the following, we will be focusing on this core set of around 110 tales.

In the plots of the Animal and Magic Tales, textually encoded emotions are particularly clearly identifiable drivers of action.See Geck 2014. An assessment of Uther’s (2013) handbook index rendered more than ninety affective and emotional terms (for example, contentment, rage, anger, just to name the ones towards the end of the German alphabet). The typified characters are distinctly discernible as good and evil, and thus feelings and emotions are distributed according to roles,Cf. Horn 2016a. with a lack of psychologization, while, often, a moral perspective is observable.Cf. Rölleke 2004. Elements of magic (›the marvelous‹), especially in the form of ›supernatural helpers,‹ can be animals, plants, anthropomorphic, as well as magic gifts. Cf. Horn 2016b. These magic components are presented as ›matter of fact‹ within the fictional world; they are key elements of plot development as they typically induce change. Although the fairy tales open up a wide array of potential interpretations, their shortness, stylistic markedness, symbolic underdetermination, dimension of marvelousness, and schematic narration constitute a clearly recognizable and basic structure.

2.2 Theoretical Perspectives in Emotion Research

Emotions have long been a core topic across different disciplines. Within psychological theories, the standard view conceptualizes an emotion observed in humans as a response to a stimulus or a situation initiated by different appraisals that indicate personal significance. Human emotions involve subjective experience, physiology, thoughts, and behavior. Cf. van Berkum 2022, p. 6. The basic emotion perspective suggests that emotion episodes can be sorted into discrete emotion types, i. e., different basic forms that are biological adaptations, characterized by facial and bodily expressions and physiological reactions.Cf. Ekman 1992. Emotional experiences can therefore be categorized using labels like anger, disgust, fear, sadness, surprise, and happiness, whereby there is no agreement about the exact number of categories. Although this approach is close to our everyday experience and underpinned by empirical results from psychological,Cf. for example Lench, et al. 2011. ethnographicCf. for example Ekman 1992. and even neuroscientific research,Cf. for example Pankseep / Watt 2011. there are several shortcomings.Cf. van Berkum 2022, pp. 9–11. For example, even if basic emotions are understood as categories that create a semantic space used to characterize our emotional experiences as well as linguistic descriptions of it, these categories are not clearly demarcated. Alan Cowen and colleagues therefore suggest that emotional experiences are best described by gradients rather than discrete categories.Cf. Cowen et al. 2019, p. 84.

According to another major theoretical perspective in emotion psychology, the dimensional approach, the variety in our emotional experience may be studied along major dimensions. The two most often discussed and explored ones are valence and arousal, the first reaching from positive to negative, the second represents activation and intensity.Cf. van Berkum 2022, p. 14. Following the circumplex model by James A. Russell,Cf. Russell 1980. each specific emotion reflects a particular constellation of a core affect and can be described by a specific location on the dimensions valence and arousal.Cf. Posner et al. 2005. Also words and pictures depicting emotional content can be characterized with these two dimensions.Cf. Bradley / Lang 1999; Lang et al. 2008. Sentiment analysis systems developed to automatically recognize and categorize emotional content in literary textsCf. for example Mohammad 2011; Jacobs / Kinder 2020. implement both the basic emotion approach and the dimensional approach.Cf. also Kim / Klinger 2019.

Also approaches focusing on the development of emotions, including functionalist views such as the hierarchical model,Cf. Fischer et al. 1990. integrate aspects of the two general perspectives described above. For example, the hierarchical model of Fischer and colleagues combines a superordinate level resembling the valence dimension (but with a categorical division into positive versus negative), a level of basic emotion categories, and in addition a level of subordinate categories that are more complex and whose development depends on specific experience in social contexts. A recent study by Gerlind Grosse and colleaguesGrosse et al. 2021. on the development of the emotional vocabulary reveals that German children first use general ›emotion‹ terms like ›good‹ or ›bad.‹ Only later on they acquire labels of basic and even later of complex emotions.

2.3 Emotions in Fictional Narratives and Fairy Tales

Despite the central role that emotions play in the content, structure, and for the effects of literary texts, the study of emotions has long been neglected in literary studies; it picked up only during the affective turn of the 1990s.For recent overviews, cf. the edited volumes by Hogan et al. 2022 and Koppenfels / Zumbusch 2016. Cf. also Flüh 2020. Studies of emotion in literature can normally be distinguished by a predominant focus either on (a) readers, putting a particular emphasis on the emotions potentially evoked in the audience,Cf. Alfes 1995; Anz 2007; Mellmann 2006; Voss 2004; Winko 2022. the (b) cultural-historical context, tracing the different incarnations of specific emotions such as jealousy or joy through genres and literary periods,Such approaches study emotions and language in connection with poetics, social conditions, historical conjunctures and change, including the institutions involved in the political sector, media, the labor market, military, education, religion or family. For joy, for example, Anja Gerigk states that in the German literary discourse around 1800, joy becomes a ›universal idea.‹ She mentions as exemplary texts Klopstock’s odes, some of Hölderlin’s poems, and most prominently Schiller’s hymn of 1785 to the ›beautiful spark of the gods,‹ while in later literary periods, joyfulness becomes a precarious exception (p. 546). As these observations refer to highly canonical texts, they should eventually be tested on broader literary corpora (cf. Gerigk 2016). or on (c) emotions as textual phenomena, based on an understanding of the symbolic, textually encoded, and cultural character of emotions.The different types of text-oriented studies convene in stressing that feelings and emotions in texts are represented, or encoded, and thus not the ›emotions themselves,‹ which emerge in the involved actors (cf. Winko 2003; cf. also Hillebrandt 2011 and (text) linguistic approaches that capture the emotional dimension of texts such as Schwarz-Friesel 2017; cf. also Benthien et al. 2000, p. 7).

Literary studies of emotion have typically focused on ›discrete‹ emotions, such as anger, hatred, jealousy, envy, love, longing, or desire. This trend appears in direct contrast with psychology and even psycholinguistics, where empirical research is just as often based on the dimensional approaches mentioned above.Cf. Lindquist 2021. Literary studies of emotions apply contextual, hermeneutic or semiotic methods, and their analyses deal with specific texts, authors, or – less often – genresCf. Meyer-Sickendiek 2005. or literary periods, such as romanticism.Cf. Schwarz-Friesel 2017; Hogan et al. 2022; Koppenfels / Zumbusch 2016.

In text-oriented studies of emotion, such as the present one, emotional content is viewed as encoded by morphemes, lexemes, sentences, or stylistic units, such as metaphors and allegories. Such studies, as proposed most prominently by Simone Winko,Winko 2003. identify an emotion potential of texts that is constituted by the totality of all emotive and evaluative text elements of a formal and substantive nature.Schwarz-Friesel 2017, p. 355, translation ours. This vantage point is shared in psychological and psycholinguistic research, based on the idea that the affective content of text may be approached or estimated by using the affective meaning of the words composing the text.Cf. Bestgen 1994; Lüdtke / Jacobs 2015. The emotion potential can be determined using text-analytical and linguistic methods, and may even be independent from epochs, genres, and production factors.Schwarz-Friesel 2017, p. 355, translation ours. We are skeptical about the latter, since historical and cultural differences, including historical language change, impose limitations on the ability to identify ›the emotions‹ in a piece of literature.Cf. Lynch 2022, p. 101; D’Arcens 2020. Yet, we fundamentally assume that a precise textual analysis can record and describe the emotive and evaluative aspects of a respective text structure and content, and that accordingly, emotion potentials may be identified for texts.Cf. Schwarz-Friesel 2017, p. 355.

Our semiotically oriented perspectiveCf. Lotman 1993, who emphasizes the role of the poetic function of language. Cf. also Nantke 2017, p. 3, for ensuing thoughts about a digital semiotics of literary texts. approaches literary text without direct statements about reception through readers, and it grasps emotions as a culturally situated code of their own.Cf. Winko 2003, p. 109; Vester 1991. Thus, by contrast to reader-oriented studies, we are focusing on what the text holds. With the import of text-mining practices to the humanities, a large array of computational emotion classification and sentiment analysis resourcesCf. Pang / Lee 2008. has become available. Also in German-language computational literary studies, approaches increasingly apply the above-mentioned psychological dimensions of valence and arousal and the discrete basic emotions to model emotions in literary texts.Cf. Kim / Klinger 2019; Zehe et al. 2017. For a different approach that historically models a set of discrete emotions for drama, cf. Dennerlein et al., accepted. In DH and Natural Language Processing (NLP), fairy tales have been a popular topic,Cf. Alm / Sproat 2005; Finlayson 2012; Geck 2014; Mohammad 2011. which can be explained by their comparatively simple structure that poses less problems to automatic analysis than other genres, as well as the ongoing discussions about their dark and negative content.Cf. Rotari 2018.

One of the robust discoveries of sentiment analysis is that textually encoded emotions can work as a proxy for plot structure, where sentiment trajectories describe emotional states over the course of a narration,Elsner 2015; Jockers 2014; Zehe et al. 2017; Mohammad 2011. or can identify the presence or absence of a happy ending.Zehe et al. 2016. Andrew J. Reagan and colleaguesReagan et al. 2016. proposed six basic shapes of emotional trajectories, while Evgeny Kim and colleaguesKim et al. 2017. analyzed the emotion trajectories in various genres for different discrete emotions.

However, despite dynamic research, the widely applied SA methods are still quite crude, often using word-level sentiment and fixed dictionaries that cannot account for negation, intensification or, for example, figurative language.Cf. Fehle et al. 2021; Kim / Klinger 2019. For the automatic analysis of discrete emotions in German, only few resources are available so far, some of which are still in the prototype stage, and their problems are evident in the assignment of emotion categories (low inter-annotator reliability for core lexica, lack of unique values, emotion lexica translated from English).Cf. Klinger et al. 2016; Stamm 2014. As dictionary-based resources, these still show the above-mentioned weaknesses combined with low semantic context sensitivity and lexical coverage.

From a semiotic perspective, any individual literary text constitutes its particular secondary semiotic system, which makes it highly likely that it encodes emotions in a way that can be resolved eventually only through knowledge of the specific text – and thus not solely by resource to the general language / culture system (the primary semiotic system).Cf. Lotman 1981. This has implications for the automatic or semi-automatic matching of emotion constructs onto linguistic codes, which are typically underspecified, if coming from language-general resources, or overspecified, if coming from specialized resources. In other words, dictionary-based SA is typically based on over- or underspecified models of primary semiotic systems — while what is needed are methods that can model emotions from the secondary semiotic systems.

In the present study, we move towards this goal by starting with annotating the CHT corpus sentence-by-sentence. As a schematic genre, the CHT share emotional codes across the individual texts, something that increases the chances of automatic emotion and sentiment detection, and is implemented in our annotation procedure that considers genre schema as well as the story meaning of the individual texts. By using sentences as basic units, it goes beyond the single-words approach of traditional sentiment dictionaries.Word-embedding based dictionaries are another approach to solve the contextualization problem, cf. for example Jacobs / Kinder 2020. Furthermore, with Gerald Prince, we use sentences as (rough) proxies for narrative events.Prince defines an event in a story as any part of that story which can be expressed by a sentence, where sentence is taken to be the transform of at least one, but less than two, discrete elementary string; (Prince 1973, p. 17). Operationalizing two psychological constructs of emotion, evaluating intercoder reliability, and sharing all research data, our annotation strives to be maximally transparent and to reach a comparatively high degree of validity.

2.4 Operationalisation of Emotions in Fairy Tales

As folk tales, Grimm’s fairy tales do not address the characters’ psychological motivations in a detailed way. Rather, even the main characters’ emotional states and traits are normally rendered in a formulaic and implicit manner.Lüthi 2016. Yet, as mentioned above, it has been shown that that fairy tales appeal to audiences through their emotions,Jones 2002, p. xiv. and that emotions and their symptoms are vital drivers of the narrative. Fear or anger experienced by the main character, for example, are likely to lead to actions, such as the appearance of a supernatural helper / gift, or the initialization of a quest.Horn 2016a.

In the tales’ narratives, emotions are often textually encoded through simple assertions as for example in: Then the girl brought the bowl to her stepmother, was happy and thought she could go to the wedding with her.Cinderella (CHT 21), Sentence 34, emphasis / translation ours. More often, however, they are ›implicitly‹ encoded, that is, not explicitly referenced through lexis, but through the description of bodily actions, including speech, facial expressions, gesture, and other involuntary bodily assertions. The narration of a character’s linguistic or other behavior allows the conclusion that a certain emotion is present without being explicitly named linguistically, relying on a formulaic match between some emotional content and its accompanying symptoms, such as laughter, kissing, beard-pulling, blushing, faintingCf. Horn 2016a. — and tears: Cinderella obeyed, but cried because she would have liked to go to the dance too.Cinderella, Sentence 27, emphasis / translation ours. Among these implicitly encoded emotions also belong more extended physical situations that result from emotions and that produce emotions, for example in: But the dwarf cursed him, and he, like the other, got into a mountain ravine and could not go forward or backward,The Water of Life (CHT 97), Sentence 17, emphasis / translation ours. where the characters are likely to experience anger and fear.In our study, we observe emotions at the primary level of the narrative, at face value of the stories and their characters. Of course, a secondary level of the narrative may be addressed where the story is allegorically read. Here, a situation of ›neither forward nor backward‹ may be matched onto any situation in which one is trapped after having taken a wrong decision (and being punished for it).

3. Data Collection and Method

3.1 Data Collection

To create the ChildTale-A corpus, we started with compiling the titles of the N = 198 Children's and Household Tales of the 7th edition (1857) by the Brothers Grimm in a spreadsheet. This spreadsheet holds the texts’ titles, as well as codes for CHT and the Aarne-Thompson-Uther-Inddex (ATU) (including those for the subset of Animal and Magic Tales) according to Hans-Jörg Uther’s work (based on the Große Auflage).Uther 2004. Two independent experts categorized the preselected tales as either Animal and Magic Tales or ›other‹ excluding such tales that are written in dialect or that have a genre different from our core selection (for example, ›Schwank,‹ legends). This rendered a list of N = 104 tales.

Around ten of these were used for training of annotators. Since texts needed to be assigned to the annotators in a balanced number, we finally selected N = 80 tales for annotation and analysis (see Table S1). We compiled all electronic texts from Wikisource as .txt files, preprocessing the texts by excluding extra white space, comments, and page numbers, fixing special character encoding, and normalizing orthography (the latter for enabling additional automatic sentiment analysis). For the annotation, each plain text was transformed into a single .xlsx table, sentences represented by one line each, with fields for annotation in columns (we created two sets of tables, one for discrete emotions, one for dimensional emotions, see 3.2 and 3.3). Remaining parsing errors were manually corrected.

3.2 Annotation of Emotional Content Expressed in Sentences

Since different perspectives – such as that of the reader or that of the text itself – can influence the type and quality of the annotation,Cf. Buechel / Hahn 2017. annotators were instructed to differentiate between a possible reader’s response and the textual (semiotic) level. They were asked to annotate emotions encoded in the text. Sentences were defined as basic annotation units and treated as rough proxy for narrative events in the sense of Prince.Cf. Prince 1973, p. 17. For a high level of systematicity and precision of annotation, annotators received instructions about genre schema (for example, lack of psychologization, steady perspectivation of heroes, no ›deep interpretation‹) and were prompted to form a story meaning by reading the individual text before annotation.A full definition of what is to be understood as ›text meaning‹ is a desideratum of interdisciplinary theorizing including perspectives from reception studies, literary semiotics, hermeneutics, and psycholinguistics. After attempts by Iser, Eco and others, so far no updated model of meaning representation at the textual level of discourse has been proposed. When we annotate emotions, we are not operating at the level of the situation model that was first proposed by van Dijk and Kintsch 1983, because our model is not a cognitive model. Rather, we address what has been called the text strategy, assessing the encoded signs and inferences (cf. for example Jannidis 2004, p. 28). However, the question of the location of different kinds of inferences, especially when adding genre knowledge, as proposed by Hanauer 1998, is still a topic of unresolved debate even for cognitive models. For our textual modeling of meaning, it is thus an open question how a coherent representation of the narration is built from the text surface, from a network of propositions and from ›probable,‹ ›intersubjective‹ inferences. Future theoretical modeling will want to utilize machine learning with Language Models (LMs) as well as manual annotations.

Following the hierarchical approach of Kurt W. Fischer and colleagues,Cf. Fischer et al. 1990. we used two different annotation schemas (see Table 1) and organized the annotation process in two different runs.

The first schema (›dimensional‹) implements the dimensional approach according to which emotions and the emotional content of language and objects are described on the two dimensions valence and arousal.Cf. Russell 1980. Following the standard research on single word processing,Cf. Bradley / Lang 1999; Võ et al. 2009; Schmidtke et al. 2014. the valence of a sentence was annotated on a seven-point bipolar scale ranging from -3 (labeled as negative) to +3 (labeled as positive), with ›neutral‹ (0) in the center. Arousal was annotated on a 5-point scale labeled with the anchors ›ruhig, entspannt‹ (›calm / relaxed‹) for 1 and ›erregt / aufgeregt‹ (›excited / agitated‹) for 5. Moreover, while annotating the degree of valence and arousal, the annotators also indicated for each sentence to which character and by which narrative instance an emotion is attributed (for example the character itself or another character; see Table 1 for details).

The second schema (›discrete‹) implements basic emotion theory.Cf. Ekman 1992; Levenson 2011. Since there is noresearch consensus about the exact number and nature of the basic emotions,Cf. Tracy / Randels 2011, p. 399. we followed Grosse and colleagues.Grosse et al. 2021. and used the basic emotions anger, disgust, fear, joy, sadness, and surprise. Instead of annotating the degree or intensity of each basic emotion within each sentence, the annotators were instructed to indicate whether one or more of the basic emotions was textually encoded or not. For each sentence, none, one, or any number of the six basic emotions could be annotated. Moreover, if the presence of an emotion was annotated, the annotators also indicated whether the emotion(s) were expressed explicitly, i. e. using emotion words (for example nouns like ›fear,‹ adjectives like ›sadly‹ or verbs like ›to grieve‹) or implicitly, for example a bodily action, metaphor, or paraphrase. If a detected emotional content could not be identified as belonging to one of the six basic emotions, the annotators could add a note describing the identified emotion using their own words.

dimension coding specific instruction anchors / examples Schema 1: ›dimensional‹ valence continuous scale from -3 to 3 Please indicate to what extent the sentence describes something as negative, neutral or positive. negative (-3) over neutral (0) to positive (3) arousal continuous scale from 1 to 5 Please indicate to what extent the sentence describes something as calm / relaxed or something excited / agitated. calm/relaxed (1) to excited/agitated (5) whom? free answer format To ›whom‹ (which character) is the emotion(s) attributed? for example to Cinderella, to the bear who? free answer format ›Who‹ attributes this emotion(s)? for example by the character itself, by another character notes free answer format Schema 2: ›discrete‹ fear presence (categorial): yes or no ›Annotate the emotional content by evaluating which emotions are expressed. You can choose between six different emotions. Here, it is possible that no emotion is present, only one, or several at once.‹ for example, fear like when seeing angry fighting dog w/o a leash anger for example, anger as when s.o. snatches parking space/last cup of coffee from you sadness for example, sadness at death of a beloved pet disgust for example, disgust at the sight of swarming maggots joy for example, joy at a successful birthday party surprise for example, surprise at very unexpected meeting of a person explicit / implicit if presence kind of expression (categorial): explicit or implicit ›Please indicate whether the sentence directly names emotion(s) or emotional action(s) explicitly using emotion words or whether the text merely suggests or paraphrases that an emotion is present. Implicit emotionality can include quite explicit descriptions, but they are still implicit if no emotion is named directly.‹ explicit: for example, he was full of anger, they rejoiced implicit: for example, she wept bitterly notes free answer format Note: As the main focus in this paper is on textually encoded emotions, we here do not further analyze the annotations for the categories whom?, who?, and explicit / implicit. The fairy tales as well as the original instructions are available in our Zenodo repository (Lüdtke / Herrmann 2023). Tab. 1: Annotation schemas and specific instructions for each category.

3.3 Workflow, Data Preparation, and Reliability of the Annotations

3.3.1 Workflow

The annotation process was embedded in an interdisciplinary M. A. seminar (›Emotionsanalyse in der Digitalen Literaturwissenschaft‹) that took place at Freie Universität Berlin in the fall term 2020 / 2021. During the seminar, students were first introduced to folk tales research, specifically, the CHT,Neuhaus 2017. and also learned psychological approaches to the concept of emotions, in particular classical emotion theories (see 2.2). In total, 17 master students participated in the seminar and together with three student assistants, all of them acted as annotators. Annotation was prepared by a training phase, in which all annotators tested the two schemas, annotating respectively one of two fairy tales, followed by a group discussion of disagreements and open questions.

In both annotation runs, each annotator was given eight fairy tales. Each person first annotated all eight tales by schema one, and then the same eight fairy tales by schema two (Table 1). In total, each fairy tale was annotated by two different randomly selected annotators.

Within the two runs, annotation of each fairy tale was conducted in two steps: First and before annotating each single sentence of a fairy tale, an annotator read the tale as a whole, using the .txt file provided. Afterwards, the tale was read again sentence by sentence and annotations were recorded using the excel sheets provided (see 3.1). To prepare the annotated data for further analysis, we collapsed all single excel files from the first and the second annotation into one data table each.

3.3.2 Data Preparation and Reliability

The preprocessing was carried out separately for the two annotation runs. We first checked for missing values in annotations of valence and arousal: In total, out of N = 5579, n = 97 sentences showed only annotations from only one person. For one additional sentence both valence annotations were missing. In a next step we calculated Krippendorff’s alpha to check for inter-rater reliability for valence and arousal using the irr packageGamer et al. 2019. in RR Core Team 2021. (see Table 2 for details). While the overall agreement for valence (meanα-Valence = .64) indicates substantial agreement, the considerably lower overall mean for the arousal (meanα-Arousal = .31) indicates only fair agreement.Cf. Landis / Koch 1977, p. 165. In general, higher agreements for valence compared to arousal are in line with experiences from other rating studies.Cf. Kaakinen et al. 2022, p. 9.; Warriner et al. 2013, p. 1194. The agreement differences between valence and arousal fit the results of Grosse and colleagues,Cf. Grosse et al. 2021, p. 157. who clearly identified the valence dimension in verbal emotion descriptions, but failed to identify the arousal dimension. We therefore decided to concentrate on the analysis of the valence annotation. All further analyses are based on the valence value for each sentence, calculated as the mean over both annotations. If one annotation was missing, the other annotation was used (for the one sentence with no annotation an interpolation was used).

For the second annotation, a check for missing values was not possible, because annotators only coded the appearance of a basic emotion with1, but not the non-appearance. Missing values were automatically re-coded with 0 indicating non-appearance. Again Krippendorff’s alpha was calculated to check the inter-rater reliability for each basic emotion and each fairy tale separately (see Table 2). Taken together, the reliability measures for the basic emotions were lower compared to the reliability observed for valence: between ›moderate‹ (meanα-anger = .47, meanα-sadness = .48, meanα-joy = .43) and ›fair‹ (meanα-fear = .39, meanα-surprise = .28, meanα-disgust = .17). Again, the relatively low values are in line with the literature on reliability measures for individual emotions.

Cecilia Ovesdotter Alm and Richard Sproat,Alm / Sproat 2005, p. 670. for example, also reported ›moderate‹ and ›fair‹ Kappa values between .2 and .5 for their annotations of specific emotions. We agree with their assumption that one reason may be the difficulty of clearly distinguishing discrete emotions from one another. Moreover, our data show that the rarity of individual emotions, especially in the case of disgust, is problematic, as shown for example by the annotations for the fairy tale Thumbling’s Travels (ID10/CHT45):The original German titles and their English translations (based on D. L. Ashliman’s translation) can be found in Table S1 in the Appendix. Moreover, each fairy tale in ChildTale-A is indicated by an ID. Here, in n = 71 sentences, no presence of disgust was jointly annotated, and only in one sentence one of the annotators indicated disgust, resulting in a Krippendorff’s alpha of α = 0. Further exploration of disgust showed that both annotators indicated that emotional content was represented implicitly. We assume that implicitness is a further reason for low reliability. To handle the shortcomings of Krippendorff’s alpha, we calculated a second alternative measure for reliability, the simple percentages of agreement.See Hallgren 2012, for critical remarks on that measure. As depicted in Table 2, this measure reveals high values between 79 % (joy) and 98 % (disgust), which encouraged further analysis. All further analyses are based on the integration of two annotations available per fairy tale, using a liberal approach. The appearance of an emotion was finally coded with yes or 1 when at least identified by one of the annotators.All data are available in our Zenodo repository (Lüdtke / Herrmann 2023).

Krippendorff’s alpha percentage of agreement4 mean SD min max N 3 mean SD first annotation1 valence .64 .16 .22 .90 79 - - arousal .33 .21 0 .85 79 - - second annotation2 anger .47 .30 0 1 77 93.42 6.42 disgust .17 .30 0 1 38 98.33 2.72 fear .39 .26 0 1 77 89.11 7.47 joy .43 .22 0 1 80 79.31 13.11 sadness .47 .26 0 1 77 88.88 7.85 surprise .28 .24 0 1 80 84.53 8.05 Note: The calculation of the Krippedorff’s alpha permits negative values. We calculated the arithmetic mean by setting all negative values to zeroCf. Wirtz / Caspar 2002. to get reliability estimates that vary between 0 and 1. 1calculated with specification ›interval‹ 2calculated with specification ›nominal‹ 3 Reliability could be calculated only if two annotations were available and/or the emotion at hand (in the case of basic emotions) was indicated at least in one sentence, which lowers the number of fairy tales for which reliability score could be computed. N reports the number of fairy tales for which we could calculate Krippendorff’s alpha. 4 Percentage of agreement could be calculated for the annotations for all N = 80 fairy tales, but only for basic emotions. Tab. 2: Reliability of annotations for valence, arousal, and the six basic emotions calculated separately for each fairy tale as averages of Krippendorff’s alpha and percentage of agreement.

4. Results and Discussion

4.1 Four Measures Representing Different Features of Textually Encoded Emotion

The collected annotations can be used to answer different questions about the characteristics and functions of textual encoded emotions in the specific genre of the CHT in particular, or in literary texts in general. Sentiment analyses like that of Saif Mohammad and Gabriela RotariMohammad 2011; Rotari 2018. asked whether fairy tales are darker and more negative than other literary genres or epochs. Approaches within literary studies assume that communication in literary texts is as a rule an emotionally highly intensive affair,Anz 2007, p. 207 (translation ours). and is thus interested in the extent of a textual emotion potential and whether specific and distinct emotional trajectories may be identified. Moreover, the role of specific emotions for some genres, texts, and epochs are discussed. In order to analyze emotions in literary texts in a comparative way, we suggest four different and relatively simple measures that describe different aspects of textually encoded emotions. In the following, we briefly describe each measure, i. e. how it characterizes the CHT as a specific genre and how it can be used for further in depth quantitative and qualitative analysis regarding different questions with respect to the role of emotions especially in CHT or in literary texts in general.

4.1.1 Average Valence

One often discussed question with respect to CHT is how gloomy or happy fairy tales are.

Testing the widespread idea that Grimms’ fairy tales, even in the ›defused‹ edition from 1857, have a tendency to depict comparably dark and negative content, existing sentiment-analysis studies render conflicting results.Alm / Sproat 2005; Mohammad 2011; Rotari 2018. While Mohammad and Rotari applied dictionary-based sentiment analysis, we use more context-sensitive annotations: The valence annotations (scaled between -3 [›negative‹] and +3 [›positive‹]) for single sentences offer the possibility to characterize the overall emotional tone in each fairy tale as well as in the entire ChildTale-A corpus by calculating the arithmetic means for valence, the Average Valence over all sentences of a text.

The histogram in Figure 1All figures in this article were created using R (R Core Team 2021) and the packages corrplot, cowplot, ggplot2, ggExtra, and ggpubr. shows the relative frequency of the Average Valence (valav) for N = 80 annotated fairy tales. We see that most of the fairy tales of our corpus are ›neutral,‹ which we determine as texts with values in between -0.5 and +0.5 (see below, 4.1.2). Also the overall mean for all Average Valence values is with mean = -0.008 (SDSD = Standard Deviation. = 0.37, median = -.05) neutral. The density curve indicated an approximately normal distribution. Only a few texts are slightly more negative overall, the ›most negative‹ fairy tales being The Godfather (valav = -0.77) and The Goose-Girl (valav = -0.76). However, outliers are observed only in the positive range (indicated by the dots in the boxplot in Figure 1), namely, the fairy tales The Three Brothers, valav = 1.38) and The Elves (valav = 0.83).

Fig. 1: Histogram, density curve and Boxplot for the Average Valence values of all annotated fairy tales. The y-axis shows the relative number of texts per valence segment, the dashed auxiliary line depicts the overall mean = -0.008. [Graphic: Berenike Herrmann / Jana Lüdtke 2023]

In sum, the Average Valence values demonstrate that we do not observe a general tendency of the Animal and Magic Tales to be particularly dark. However, the overall neutral mean for all Average Valence values seems to indicate an absence of the often reported positivity bias in Western literature.Cf. Dodds et al. 2015; Green 2017; Jacobs et al. 2020. As the Average Valence is pitched at the highest level of generalization, a neutral value may however mean two things: either a general tendency towards the absence of emotions, or the presence of variation within each fairy tale. Since the arithmetic means gloss over the text-internal variance, our next analyses will take a closer look at those.

4.1.2 Emotion Potential

To disentangle the possible interpretation of Average Valence values around zero, we focussed on two questions:

To what extent fairy tales are emotionally encoded? What proportion of each tale is in fact ›emotional‹ as opposed to ›neutral‹?

Here, what is under scrutiny is the tendency of fairy tales to encode emotion comparatively sparsely, including typified characters without psychologization,Cf. Lüthi 2016. as well as the presence of emotion variation within each fairy tale.

We classified all sentences according to their valence values as positive (valence ≥ 0.5), negative (valence ≤ -0.5), or neutral (-0.5 < valence < 0.5) and calculated the percentage of positive, negative, and neutral sentences within each fairy tale. As depicted in Figure 2A, there are significantly more neutral sentences (meanneutral = 41.69, SDneutral = 12.57) compared to positive (meanpositive = 29.06, SDpositive = 11.01) and negative sentences (meannegative = 29.25, SDnegative = 11.58) within each fairy tale, while the average percentage of positive and negative sentences is equal.The paired pairwise t-tests for neutral compared to negative and positive sentences are significant (t(79) > 5.16, p < .0001) but not the t-test comparing positive and negative sentences (t < 1). P-values for multiple pairwise paired t-tests are adjusted using the Bonferroni correction. The overall neutrality visible in the Average Valence values relies therefore on both a considerable amount of sentences characterized by an absence of textual encoded emotions and an equal occurrence of sentences with positive and negative emotions. To estimate the amount of emotionality, often discussed under the cultural-semiotic concept of textual Emotion Potential (EP),Cf. Schwarz-Friesel 2017; Hillebrandt 2011; Winko 2022. we operationalized the Emotion Potential by calculating the proportion of sentences with textual encoded emotions, the sum of the percentage of positive and negative sentences. The histogram in Figure 2B depicts the distribution of the Emotion Potential values in our ChildTale-A corpus. More than 50 % of the annotated fairy tales have a proportion of at least 50 % emotion-bearing sentences (a 50 % cut-off is marked with the black dotted line), which is also visible in the overall mean for the Emotion Potential values (meanEP = 58.31, SDEP = 12.57). This indicates a saturation with textual encoded emotion. Only in three fairy tales (including The Strange Musician and The Devil with the Three Golden Hairs) less than one third of the sentences are emotional (see Table S1 for more details).

Fig. 2: A – Relative frequency of the negative, neutral, and positive sentence (in %) in all fairy tales and the results of pairwise comparisons, B – Histogram and boxplot for the Emotion Potential (relative frequency (in %) of emotional sentences in the fairy tales, black dashed line indicates 50 %, red dashed line indicates the overall corpus meanEP = 58.31), C - Scatterplot of the relative frequency of positive and negative sentences (in %) per fairy tale (individual texts coded by color, dashed lines indicate 50 %). [Graphic: Berenike Herrmann / Jana Lüdtke 2023]

To further explore the differences between the fairy tales, we also compared the amount of positive and negative sentences within each fairy tale. A visualization can be found in the scatter plot in Figure 2C, indicating substantial differences. Some fairy tales are predominantly positive, for example The Three Brothers or The Elves (depicted to the right of the black dotted auxiliary line), others are predominantly negative, like The Godfather or Frau Holle (depicted above the black dotted auxiliary line). The majority of texts, however, comprises both positive and negative sentences and can thus explain the Average Valence around zero. This raises the question of the role that textually encoded emotions play in the narrative progression of a single text, which we address in the next section.

4.1.3 Emotional Arcs

In order to address the text-internal variation of emotions, we computed Emotional Arcs that visualize the trajectory of emotion over narrative time. Based on computational analysis of more than 1,000 books, Reagan and colleaguesCf. Reagan et al. 2016, p. 7. proposed six basic shapes of emotional trajectories or emotional arcs. They calculated average happiness scores for text segments with the text-based hedonometer (measuring joy) developed by Peter Sheridan Dodds and colleagues,Cf. Dodds et al. 2011. using different time series decomposition methods to identify core emotional arcs. Similarly, Evgeny Kim and colleagues analyzed the emotion trajectories in various genres for different discrete emotions, including the six basic emotions annotated in this project.Cf. Kim et al. 2017. Unlike Reagean and colleagues, however, their analyses show that trajectories for fear, but not for joy, are the most uniform and informative.Cf. also Archer / Jockers 2016, p. 94, who propose seven basic plots.

To integrate the results of both approaches, we compute Emotional Arcs based on valence annotations, since valence is, in contrast to basic emotions, theoretically considered as bipolar. This is supported by the seminal study of Bestgen,Bestgen 1994, p. 33. who used valence ratings at sentence level to illustrate the emotional curve or profile of a story. How do we calculate the Emotional Arcs? As they approximate the ›emotional development‹ within a story at a macro level, we applied discrete cosine transformation (DCT),Ahmed et al. 1974. a transformation and dimension reduction technique in signal processing, widely used also within natural language processing and sentiment analysis.For example in Jockers’ syuzhet package (Jockers 2020a). DCT was applied by using the dct-function of the emuR package in R.Cf. Winkelmann et al. 2017. Taking into account that the shortest of the annotated fairy tales has six sentences, we used the DCT-function with m = 5 to reconstruct a smoothed version of the valence annotations. Using the DCT-function also allowed a length normalization. Elkins compares the pros and cons of some smoothing algorithms used to model emotional arcs (cf. Elkins 2022, p. 27). Although she criticizes DCT because of the dependence of the final shape of the arc on the selected parameter for m, she used that method together with a length normalization to calculate simplified macro shapes. We therefore transformed the narrative time indicated by the position of a sentence in a text to a uniform window reaching from 1 to 100.The DCT-smoothed data can be found in our Zenodo repository (Lüdtke / Herrmann 2023). Figure 3 depicts the so obtained Emotional Arcs for all fairy tales.

Fig. 3: Emotional Arcs of all fairy tales in ChildTale-A based on DCT-smoothed valence annotations. Narrative time is normalized to a time window from 1 to 100. The Emotional Arcs are ordered according to the ascending range of the emotion trajectories, defined as the difference between the highest and lowest smoothed valence value. The assignment of the IDs to the fairy tales can be found in Table S1. Black dashed lines indicate the theoretical mean of the valence scale. [Graphic: Berenike Herrmann / Jana Lüdtke 2023]

The Emotional Arcs in Figure 3 are ordered by the range observed in the DCT-smoothed valence annotations, i. e. the difference between the highest and lowest value in the arc. Besides differences in the patterns of up and downs, the range seems to be an important characteristic of the Emotional Arcs. However, one should bear in mind that the range of the Emotional Arcs is correlated with the length of the tales, a correlation that is an artifact of the used DCT-function: Applying DCT with the fixed parameter m = 5 to all fairy tales means that in short (as compared to long) fairy tales the transformed data are closer to the original raw data and therefore less smoothed. The artificial nature of the smoothing is visible in Figures 4A and 4B: It shows that correlation between range and story length is significantly negative (r = -.49, p < .001, see Figure 4A), while the correlation between story length and the range of the original valence annotations (also called valence span)Cf. Jacobs 2015, p. 5. indicates a positive relationship (r = .42, p = .001, see Figure 4B).

To test whether DCT is nevertheless an appropriate algorithm for creating the Emotional Arcs, we calculated the correlation between the range of the Emotional Arc of each fairy tale and its Emotion Potential (see Figure 4C), which is significantly positive (r = .42, p = .001). For fairy tales with low Emotion Potentials we observe Emotional Arcs with small ranges (that means flat lines), while for fairy tales with high Emotion Potentials we observe Emotional Arcs with high ranges (pronounced ups and downs). This supports DCT as an adequate smoothing method.

Fig. 4: Relationship between fairy tale length and range of DCT smoothed valence annotations (A), fairy tale length and range of the original valence annotations (valence span) (B), and Emotion Potential (percent of positive and negative sentences) and range of the DCT smoothed valence annotations (C). [Graphic: Berenike Herrmann / Jana Lüdtke 2023]

To illustrate the positive relationship between the Emotion Potential and the range of the curvature of the Emotional Arcs, Figure 5 depicts the trajectory of the valence annotations of the sentences and the Emotional Arc for one fairy tale with a low (The Strange Musician, ID34)Each fairy tale in ChildTale-A received its own ID, which we use in the following. For assignment of title to ID, see Tab. S1 in the appendix. and one with a high Emotion Potential (Cinderella, ID02). Also ID02 is considerably longer than ID34, the range of the Emotional Arc is bigger, making an extant curvature and a clear profile visible.

Fig. 5: Trajectory of the changes in the original valence annotation (values between the dashed auxiliary lines correspond to sentences annotated as neutral) and the trajectory of the Emotional Arc, i. e. the DCT-smoothed valence annotations (m = 5) for the fairy tales ID34 and ID02. [Graphic: J. Berenike Herrmann / Jana Lüdtke 2023]

In the following, we address some noticeable characteristics of the Emotional Arcs by focusing on the most extreme cases of Figure 3, the fairy tales The Strange Musician (ID34), characterized by a flattened curve, and Sweet Porridge (ID29) characterized by the most pronounced curve.

As described above, one reason for the pronounced rise-fall-rise pattern in ID29 lies in its brevity, as ID29 is with six sentences the shortest fairy tale in the ChildTale-A corpus.For an overview of main results per text, see Table S1, Appendix. But independently of the shortness, almost all of the sentences (five out of six) bear clear emotions resulting in an Emotion Potential of 83 %. We are thus dealing with a condensation of a prototypical folk narrative, in which a situation of deficiency (an already poor mother-daughter couple has run out of food) first turns to the better (the girl ventures out in the woods and meets an old lady who gives her a magic item, a pot that prepares sweet millet porridge at a command and stops again at another command; from then on, they never have to go hungry again). But then things turn for the worse again (one day the girl is out of the house, and the mother brings the pot to cook, but forgets the second command, so the pot buries the whole town/world under porridge). Eventually, the story turns to the better again, when the child comes home and tells the pot to stop: The town is turned into a kind of Schlaraffenland (›land of plenty‹), where all visitors have to eat themselves through the porridge, which is not a menace anymore, but rather an edible wall that protects the town.

Compared to the six basic shapes proposed by Reagan and colleagues,Reagan et al. 2016, Fig. 4, p. 7. The ›story shape‹ research goes back to anthropological work by — among others — Vladimir Propp and Claude Levi-Strauss. From these symbol- and form-oriented approaches, Kurt Vonnegut distilled his thesis that all stories have simple shapes, something which he explored in a (rejected) Master’s thesis and in (youtube) lectures on the example of Cinderella, for example Comberg 2010; cf. Elkins, 2022. we thus identify the rise-fall-rise pattern, which Reagan and colleagues match onto the iconic fairy tale Cinderella. Since Cinderella is also part of ChildTale-A (ID02, see Figure 5B), we can directly compare the Emotional Arc with the basic shape proposed by Reagan and colleagues. And indeed, the Emotional Arc follows the same basic shape, though in a less prototypical way. ID02 starts with the predicament of the young girl being confronted with evil step-sisters and an uncaring step-mother and her alliance with the birds that help her first with the sorting of the lentils and later by producing beautiful dresses and shoes – all of which allows her to attend the prince’s dance (rise). Then, the drastic actions of her step-family to fit the shoe (cutting toes and heel) present a fall, followed by another rise when the prince understands that Cinderella, not the sisters, is the true dancer. In our annotations the local drop refers to the punishment of the step sisters, who are blinded by the birds.

At the other extreme of Figure 3, we find The Strange Musician (ID34), which shows comparably small oscillations in the Emotional Arc. The difference between the most negative and most positive points in the Emotional Arc is only 0.87 (compared to 2.86 for Cinderella and 6.4 for Sweet Porridge). Tale ID34 includes 68 % sentences annotated as ›neutral,‹ it is therefore the tale with the lowest Emotion Potential in the ChildTale-A corpus (see Table S1). This low value fits well with the trivial plot and the initial problem: A fiddler, who gets bored walking alone, and thus wishes for a companion, which he tries to attract by playing music (I'm growing bored here in the forest, I want to fetch a good journeyman). In three successive narrative events, a wolf, a fox, and a bunny appear, each of whom wishes to learn how to play the fiddle, but each of whom is not the companion the man is looking for. To get rid of them, he misleads and fixates each animal physically, but then gets bored again. Figure 5A depicts the trajectory of the Emotional Arc and the original valence annotations per sentence for ID34. The first small valence dip around sentence no. 9 is related to the encounter with the wolf, the fiercest of the animals, who is tied up with a rather brute action involving a rock, whereas the successive events involve increasingly harmless animals (fox and bunny). The next slightly pronounced dip around sentence 34 and 35 is the wolf freeing first himself and then the other animals, blaming the musician and setting out to hunt their enemy: From the perspective of the main character, this is a moment of danger, corresponding therefore with the clearer dip. Finally, the rise towards the end of the plotline corresponds with the appearance of a new character (a poor woodcutter, sentence 39), and his enchantment by the fiddler’s beautiful music. This character is finally accepted by the fiddler as the right kind of companion. When the animals appear, intending evil, the woodcutter defends the fiddler with his ax. Overall, the plot in ID34 is less drastically emotionally encoded, corresponding with its motive of taking a stroll and looking for leisure. In the story, the realms of animals and mankind are presented as clearly separated, so the threat of the angry animals is not severe: It is not a story of ›life and death,‹ by contrast to Sweet Porridge. This is relevant, as the annotators were instructed to take into account the meaning of the whole text, and not just each sentence alone. As depicted in Figure 5A, the weight of a single clearly emotional sentence is mediated by neutral filler sentences. The overall emotional trajectory, although with comparably small oscillations, is that of fall-rise-fall-rise. This pattern does not directly match onto one of the six basic shapes reported by Andrew Reagan and colleagues.Cf. Reagan et al. 2016, p. 7

In sum, Figure 3 shows that the Emotional Arcs vary considerably. We find different types of trajectories across the ChildTale-A corpus, with variance concerning both sequencing and amplitude of valence. While some of the observed plotlines can be matched onto the six basic shapes identified by Andrew Reagan and colleagues,Cf. Reagan et al. 2016, p. 7. others cannot. To answer questions about the predominant shapes for CHT and their relation to the basic shapes described by Reagan et al., further analysis including different smoothing algorithmsCf. Elkins 2022, p. 35. in combination with hierarchical cluster analyses are necessary, which we will leave for a later publication. In general, future research on narrative patterns using our method based on valence annotations per sentence thus appears promising. So far, we have concentrated on a dimensional model of emotion. In the last analysis, we examine the distribution of discrete emotions.

4.1.4 Emotion Profiles

Hermeneutic approaches as well as computational analysis of fairy tales often focus on discrete emotions, asking about their textual distribution and their role for the narrative plot.Examples are respectively Lange 2016 and Alm / Sproat 2005. Meanwhile, reliable identification of discrete emotions in fictional and artistic texts is a challenge for several reasons, such as the context-dependence of meaning, but also the fuzzy boundaries between single emotions, and finally their partly rare occurrence (see 3.3.2). We used the collected annotations for six discrete basic emotions to assess what we call the Emotion Profile, i. e. the relative frequency of each emotion. The Emotion Potential is calculated by dividing the number of sentences in which a basic emotion is identified by the total number of sentences of the particular fairy tale. Figure 6 depicts the aggregated Emotion Profile for the ChildTale-A corpus as a whole. On average, we observed joy as a predominant basic emotion with a mean = .34 (SD = .17), indicating that joy is encoded in around one third of the sentences of each of the fairy tales. Moreover, joy was annotated in every fairy tale at least in two sentences. All other five basic emotions are observed less often and also do not occur in every fairy tale. By a substantial gap, the second most frequent basic emotion is surprise (mean = 0.21, SD = 0.10), closely followed by sadness (mean = 0.18, SD = 0.10), fear (mean = 0.17, SD = 0.11), and anger (mean = 0.11, SD = 0.09). The by far lowest occurrence was observed for disgust (mean = 0.02, SD = 0.04), which was identified in only n = 38 out of the N = 80 fairy tales (in at least one sentence). The comparison of the mean values for the relative frequencies of the six basic emotions with the overall mean in the ChildTale-A corpus (mean = 0.17, SD = 0.15) underlines that joy and surprise are both significantly overexpressed, while anger and disgust are significantly underexpressed in statistical terms.Comparing the mean values for relative frequency with each other indicated also significant differences (|tall| > 3.8, pall < .001) except for the relative frequencies of fear, sadness and joy (|tall| < 2.5, pall > .09). P-values for multiple pairwise paired t-tests are adjusted using the Bonferroni correction.

Fig. 6: Boxplots of relative frequencies of the six basic emotions in all fairy tales (absolute frequencies divided by number of sentences per fairy tale). Dashed line represents the overall mean. Mean values of each basic emotion are compared to overall mean. [Graphic: Berenike Herrmann / Jana Lüdtke 2023]

While both the annotation-based study by Alm and Sproat Alm / Sproat 2005. and the dictionary-based sentiment analysis by RotariRotari 2018. describe a rare occurrence of disgust, their frequency patterns for the other emotions differ. For example, joy as predominant emotion in ChildTale-A was also observed by Rotari, while Alm and Sproat reported the highest prevalence for anger. Surprise, the second most frequent basic emotion in ChildTale-A, was observed very rarely by Rotari, while Alm and Sproat reported intermediate frequencies.

There are at least four potential reasons for these differences that relate to data and method: First, Alm and Sproat reported results for only a preliminary tie-broken data set of 22 Grimms’ fairy talesAlm / Sproat 2005, p. 5. without information on which fairy tales were annotated, whereas Rotari analyzed N = 145 of the CHTs, including such subgenres as the ›Schwank,‹ which was excluded in the present study because of a different emotional structure. These differences in the setup of the three corpora are thus likely to have caused different overall profiles, something which is supported by our observation of fairy tales varying substantially in the occurrence of single basic emotions (see Figure 8). Second, Rotari used an automated dictionary-based annotation at word level, which could not capture implicitly encoded emotions, something the annotators in our study took into account. Third, while our annotators could indicate more than one basic emotion in a sentence, Alm and Sproat used an exclusive coding. Analyzing our annotations for joint occurrences indicated that in 26 % of all sentences more than one basic emotion was coded (compared to 49 % sentences for which only one basic emotion was annotated). And fourth, there are also differences in the used coding schema. Alm and Sproat, for example, split the basic emotion surprise into positive and negative surprise, while we used only one category, following the definition of surprise as an emotional reaction evoked by unexpected (schema-discrepant) events.Reisenzein et al. 2019, p. 50.

To better understand the Emotional Profile of the ChildTale-A corpus, we calculated correlations between the relative frequencies of all six basic emotions (see Figure 7). All significant relationships are positive, indicating co-occurrences, with the strongest one observed for joy and surprise (r = .49), and slightly weaker ones for joy and sadness (r = .35), and for fear and anger (r = .28). All other relationships are not significant.

Fig. 7: Correlation matrix with spearman correlation coefficients between the relative frequency of the six basic emotions, the Emotion Potential (percent of emotional sentences), Average Valence, range in the Emotional Arc (the DCT-smoothed valence values), and length (number of sentences) for all fairy tales (higher positive and negative correlation coefficient are marked red or blue, respectively; significant values are marked with asterisks: * p < .05, ** p < .01, *** p < .001). [Graphic: J. Berenike Herrmann / Jana Lüdtke 2023]

In addition to the overall occurrence of the six basic emotions, we calculated the Emotional Profiles for each single fairy tale. These profiles (Figure 8) demonstrate some similarities, but also clear differences across the corpus. As described above, joy (red bars) is part of each Emotional Profile. In about half of the profiles joy is the predominant emotion (for example ID2 or ID40), often flanked by surprise (yellow bars), an observation which is also expressed by the significant positive correlation between the relative frequencies of both. There are also some fairy tales in which joy and surprise appear as rarely as the other basic emotions (for example ID35 or ID62). A small group of fairy tales is characterized by fear (light blue bars) being the most prominent emotion (for example ID74 and ID71). In another small group, sadness (dark green bars) is the dominant one (for example ID53 and ID46). There are no fairy tales dominated by anger (dark blue bars) or disgust (light green bars), at least not in terms of frequency of occurrence.

Fig. 8: Emotion Profiles for all fairy tales in ChildTale-A. Fairy tales are ordered as in Figure 3, according to the range of the Emotional Arcs (the difference between the highest and lowest DCT-smoothed valence values). For the legend of the fairy tale IDs, see Table S1. [Graphic: Berenike Herrmann / Jana Lüdtke 2023]

Taken together, the Emotional Profile of the ChildTale-A corpus (see Figure 6), as well as many of the individual profiles of the single fairy tales (see Figure 8) are characterized by an above average occurrence of joy. On the one hand, the dominance of joy may be explained at least in part by the fact that joy is the only clearly positive emotion category, while negative affect is differentiated into four distinct categories. This imbalance is visible in the majority of basic emotion theories and some researchers view this fact as a negativity bias based on a greater cognitive elaboration of negative events.Rozin / Royzman 2001, p. 310. On the other hand, the dominance of joy could also be interpreted as an indication of a positivity bias, often discussed as the Pollyanna effect.Cf. Dodds et al. 2015; Green 2017. According to Jacobs et al., this higher frequency of positive compared to negative words and sentences, is the result of a universal human tendency to use evaluatively positive words more frequently, diversely and facilely than evaluatively negative words.Jacobs et al. 2020, p. 1. Jacobs et al. reported this effect for contemporary German children’s and youth literature. Cf. Jacobs et al. 2020. The results of our annotations of discrete emotions indicate that this universal tendency is also visible at the semiotic level in the Grimms’ CHT, even though it seems to be somewhat weaker: there is no positivity bias in the valence annotations (see Figure 1). Nevertheless, further hermeneutic and narratological analysis should focus on both negative and positive emotions in narrative structure, especially in relation to events. While negative emotions may function as drivers of action,Cf. Horn 2016a. positive emotions appear to correlate with narrative stasis, which might mean that they redress the level of ›eventfulness‹ in the sense of Michael Vauth and colleagues,Vauth et al. 2021, p. 337. ›re-balancing‹ valence to a relatively enjoyable status quo,Alm / Sproat 2005 suggest that a happy emotion expresses contentment with the status quo, whereas negative events force action and keep the narrative plot going (p. 7). that may then be overthrown again as the story unfolds – or lead to a happy ending. So far, much more research has been dedicated to negative emotions. We thus think it is interesting to point to the role of positive emotions, starting with a distributional profile, but extending to their functional roles in counterbalancing the negative emotions in the unfolding of the plot.

As ChildTale-A is to the best of our knowledge the only annotated literary corpus with annotations for both discrete basic emotions as well as the dimension valence, we explore in a last step the relationship between both annotations. Figure 7 depicted the correlation between the relative frequencies of the six basic emotions and the different measures derived from the valence annotations, the Average Valence, the Emotion Potential and the range of the Emotional Arc. In line with the described predominance of joy, the strongest correlation could be observed between Average Valence and the relative frequency of joy (r = .06). Fairy tales with a higher relative frequency of joy are characterized by higher Average Valence values indicating the tendency to be more positive. As one would expect, the correlation between the relative frequency of fear and Average Valence is negative (r = -.34) being the only significant correlation between a distinct negative emotion and Average Valence. The relative frequencies of joy, surprise, disgust and fear correlate significantly positively with the Emotion Potential (rall > .24), the percentage of both positive and negative sentences in a fairy tale. Also for sadness and anger, no significant relationship could be observed, the valence annotation and the annotation for the discrete emotions seem to correspond very well. For more than 85 % of the sentences categorized as positive or negative (done based on the valence annotations), the annotators detected at least one basic emotion (see Table 3). For 10 % or 12 % respectively, no basic emotion were annotated reflecting, at least in part, that not all emotions expressed at the semiotic level are covered by the six basic emotion categories. More interestingly, for only 56 % of the sentences categorized as neutral, none discrete basic emotion was encoded. For the remaining 44 % at least one basic emotion was encoded. Comparable to our interpretation of the overall mean of Average Valence, valence annotations around zero at the sentence level are not in every case a sign of an absence of emotions. Instead, valence annotations around zero could also be an index of emotional ambiguity due to an co-occurence of both positive and negative emotional episodes or events.

sentence discrete basic emotion negative neutral positive none annotated 11.85 % 44.13 % 10.36 % at least one annotated 85.15 % 55.87 % 89.64 % column sum 100 % 100 % 100 % Tab. 3: Relative frequency (in %) of the (non-)occurrence of the six basic emotions for negative, neutral, and positive sentences.

5. Conclusions and Outlook

The goal of the present study was to examine at a semiotic level the emotions encoded in a corpus of eighty German fairy tales constituting the ChildTale-A corpusAll data are available in our Zenodo repository (Lüdtke / Herrmann 2023).. The corpus represents the core genre of 19th Century folk tales, Animal and Magic Tales, a subcollection of the Children and Household Tales published by the Brothers Grimm in the last edition in 1857. To identify emotions, we annotated the strength of valence and arousal as well as the occurrence of the six basic emotions anger, disgust, fear, joy, sadness and surprise for each single sentence per text. The annotations were used to introduce four different measures, characterizing the emotion potential and the trajectories of emotional changes within fairy tales: Average Valence, Emotion Potential, Emotional Arc, and Emotion Profile.

In sum, our results indicate that the Grimmsʼ fairy tales are not particularly ›dark,‹ as assumed by some. Our results rather support the assumption that fairy tales are an optimistic genre in the sense of Bausinger, as the corpus mean for Average Valence is in a neutral range. This neutrality at the macro level is not a result of missing emotional content. It rather reflects an often equal amount of positive and negative sentences visible in the proportion of positive and negative sentences constituting the Emotion Potential. The overall neutrality, however, may indicate that a great part of the sentences annotated in our corpus are comparatively less highly emotionalized, as supported by folklore research that finds the prototypical fairy tales ›less clearly emotional‹ than legends, fables, or artistic fairy tales. Whether the amount of neutrality observed in our corpus indeed is a unique feature of that highly schematic historical genre awaits further quantitative comparison with contemporary children’s literature.

This comparison should also focus on the co-representation of different emotions at the sentence level: in ChildTales-A, almost half of all emotional sentences encode just one discrete emotion. This again appears to indicate the simple schematic structure of the folk tale and perhaps might constitute a measurable feature when compared to other genres. The Emotional Arcs calculated on the basis of the valence annotations per sentence indicate substantial variation of trajectory patterns across the corpus: While some fairy tales are characterized by flat trajectories and minor changes on the valence dimension, others show substantial oscillations, especially tales with a high Emotion Potential. Some, but not all, of the oscillation patterns are in line with the prototypical patterns described by Reagan and colleagues.Cf. Reagan et al. 2016, p.7. The Emotion Profiles based on the occurrences of six basic emotions indicate that joy is the most dominant basic emotion, which gives support for the Pollyanna effect documented by Jacobs and colleagues.Cf. Jacobs et al. 2020. However, based on the emotion profiles of individual fairy tales in ChildTales-A, it is noteworthy that in only 50 % of the annotated tales joy is quantitatively predominant, while the rest of the tales is coined by predominant fear or anger, or by absence of a prevailing emotion. Together, our data suggest the hypothesis that emotionally encoded sentences may correlate with a degree of ›eventfulness.‹Cf. Vauth et al. recently proposed a continuous metric to operationalize eventfulness ranging from non-event to high narrativity in the sense of changing states (Vauth et al. 2021, p. 337). Further in-depth analysis should prove whether negative emotions (such as fear) are indeed drivers of action, while positive emotions may cater to a ›re-balancing‹ of valence to a relatively enjoyable (transitory) status quo, Alm / Sproat 2005 suggest that a happy emotion expresses contentment with the status quo, whereas negative events force action and keep the narrative plot going (p. 7). often to be overthrown in the following succession of events.

Taken together, ChildTale-A and the four introduced measures represent a rich data set for more fine-grained analyses. The annotations for more than 5,000 sentences are a comfortable starting point for training supervised machine learning algorithms (classifiers) that can predict sentiment in further fairy tales and potentially other texts. Our data are suitable to test the applicability of existing sentiment analysis tools on literary texts that were first developed in other domains. Moreover, and to the best of our knowledge, ours is the first data set providing both annotations for the valence dimension and annotations for basic emotions. The dataset therefore opens the field to further exploration of the relationship between two theoretical approaches of discrete versus dimensional models of emotions.

Limitations of our study include a lower than ideal inter-rater agreement especially for discrete emotions, a problem which is shared by the field, as well as the fundamental problem of how to reliably model literary expert knowledge into a specific set of instructions. Our approach was to monitor representation of genre knowledge (by giving prompts as to the lack of psychologization, schematic meaning and focalization) and story meaning (by reading the whole text) through the instructions. The general challenge of reliably identifying text meaning points to a full(er) representation of the text surface, the propositional text base, and different types of inferences which necessarily presuppose genre knowledge (see Section 3.2). However, such representations appear still beyond the control of study design even for a simple genre such as the Grimms’ tales that does not involve much interpretative inferences, both for manual annotation and for deep machine learning. Yet, our study takes a step ahead, as the provided annotations promise to be useful for more in-depth narratological analysis as well as for empirical studies of reader response. Future studies about the role of emotions in literary texts would benefit from focusing not only on the forms, functions, and representations of negative, but also on those of positive emotions.

Acknowledgements

This research was supported by the project ›Advanced sentiment analysis for understanding affective-aesthetic responses to literary texts: A computational and experimental psychology approach to children’s literature‹ (424250469) funded by the DFG grant ›SPP 2207: Computational Literary Studies‹ (402743989). The research was also funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy in the context of the Cluster of Excellence Temporal Communities: Doing Literature in a Global Perspective – EXC 2020 – Project ID 390608380. Annotations were carried out by students as part of a MA seminar at Freie Universität Berlin (winter term 2019/20) as well as by student assistants from the University of Basel, Bielefeld University, and Freie Universität Berlin. We thank Dr. Giulia Grisot for her thorough reading and very helpful comments.

Appendix and Supplementary Material

All data, including the scripts for preparing annotations and preprocessing, as well as original instructions for annotation of ChildTale-A are available in our Zenodo repository (Lüdtke / Herrmann 2023).

ID German title English title1 CHTIndex text length Average Valence (scaled from -3 until 3) Percent positive sent. Percent negative sent. Percent neutral sent. Kripp. α (valence) 1 Allerleirauh All-Kinds-Of-Fur 65 85 -0.14 22.35 28.24 49.41 .40 2 Aschenputtel Cinderella 21 103 0.26 43.69 26.21 30.1 .56 3 Brüderchen und Schwesterchen Little Brother and Little Sister 11 95 -0.16 28.42 35.79 35.79 .59 4 Das blaue Licht The Blue Light 116 69 -0.49 14.49 39.13 46.38 .52 5 Das Mädchen ohne Hände The Girl without Hands 31 94 -0.18 30.85 35.11 34.04 .68 6 Das Meerhäschen The Rabbit 191 57 -0.13 28.07 26.32 45.61 .45 7 Das singende springende Löweneckerchen2 The Singing, Springing Lark 88 73 0.08 35.62 32.88 31.51 - 8 Das Waldhaus The Hut in the Woods 169 70 0.06 30 24.29 45.71 .58 9 Das Wasser des Lebens The Water of Life 97 83 -0.34 14.46 31.33 54.22 .70 10 Daumerlings Wanderschaft Thumbling's Travels 45 72 -0.06 25 27.78 47.22 .60 11 Daumesdick Thumbthick 37 100 -0.11 22 27 51 .58 12 Der Baerenhäuter Bearskin 101 71 -0.01 33.8 36.62 29.58 .76 13 Der Eisenofen The Iron Stove 127 84 0.01 28.57 27.38 44.05 .54 14 Der Froschkönig oder der eiserne Heinrich The Frog King. or Iron Heinrich 1 52 -0.42 19.23 38.46 42.31 .75 15 Der Fuchs und die Frau Gevatterin The Fox and His Cousin 74 13 0.15 30.77 15.38 53.85 .24 16 Der Geist im Glas The Spirit in the Glass Bottle 99 71 -0.27 19.72 40.85 39.44 .69 17 Der gelernte Jaeger The Trained Huntsman 111 80 -0.08 27.5 36.25 36.25 .88 18 Der Gevatter Tod Godfather Death 44 51 -0.27 19.61 41.18 39.22 .73 19 Der gläserne Sarg The Glass Coffin 163 87 -0.32 27.59 41.38 31.03 .72 20 Der goldene Vogel The Golden Bird 57 123 0.26 39.02 19.51 41.46 .68 21 Der Herr Gevatter The Godfather 42 32 -0.77 15.63 59.38 25 .77 22 Der Hund und der Sperling The Dog and the Sparrow 58 59 0.13 37.29 20.34 42.37 .22 23 Der König vom goldenen Berg The King of the Golden Mountain 92 88 -0.23 28.41 39.77 31.82 .81 24 Der Königssohn der sich vor nichts fürchtet The King's Son Who Is Afraid of Nothing 121 81 0.12 37.04 33.33 29.63 .73 25 Der Krautesel The Cabbage-Donkey 122 100 0.51 38 11 51 .37 26 Der Liebste Roland Sweetheart Roland 56 57 -0.25 12.28 24.56 63.16 .55 27 Der Ranzen, das Hütlein und das Hörnlein The Knapsack, the Hat, and the Horn 54 81 0.54 58.02 20.99 20.99 .38 28 Der Sperling und seine vier Kinder The Sparrow and His Four Children 157 34 0.56 35.29 8.82 55.88 .39 29 Der süsse Brei Sweet Porridge 103 6 0.17 50 33.33 16.67 .86 30 Der Teufel mit den drei goldenen Haaren The Devil with the Three Golden Hairs 29 117 0.08 14.53 17.95 67.52 .44 31 Der treue Johannes Faithful Johannes 6 111 0.01 20.72 22.52 56.76 .74 32 Der Trommler The Drummer 193 174 -0.06 11.49 21.26 67.24 .51 33 Der Wolf und die sieben jungen Geislein The Wolf and the Seven Young Kids 5 48 -0.09 25 31.25 43.75 .52 34 Der wunderliche Spielmann The Strange Musician 8 44 -0.14 11.36 20.45 68.18 .64 35 Der Zaunkönig und der Baer The Wren and the Bear 102 34 -0.56 23.53 47.06 29.41 .69 36 Des Teufels russiger Bruder The Devil's Sooty Brother 100 44 0.34 45.45 29.55 25 .74 37 Die beiden Wanderer The Two Travelers 107 174 0.13 29.89 27.01 43.1 .52 38 Die Bienenkönigin The Queen Bee 62 27 -0.33 22.22 40.74 37.04 .75 39 Die Bremer Stadtmusikanten The Bremen Town Musicians 27 45 0.09 24.44 20 55.56 .46 40 Die drei Brüder The Three Brothers 124 21 1.38 61.9 4.76 33.33 .68 41 Die drei Federn The Three Feathers 63 44 0.43 27.27 18.18 54.55 .68 42 Die drei Männlein im Walde The Three Little Men in the Woods 13 82 0.06 28.05 24.39 47.56 .35 43 Die drei Schlangenblätter The Three Snake-Leaves 16 53 -0.25 20.75 30.19 49.06 .61 44 Die drei Spinnerinnen The Three Spinning Women 14 35 0.34 42.86 28.57 28.57 .67 45 Die drei Sprachen The Three Languages 33 38 0.13 34.21 26.32 39.47 .73 46 Die Gaensemagd The Goose-Girl 89 71 -0.76 16.9 56.34 26.76 .79 47 Die Geschenke des kleinen Volkes The Gifts of the Little People 182 30 0.28 53.33 26.67 20 .84 48 Die goldene Gans The Golden Goose 64 54 0.19 24.07 12.96 62.96 .51 49 Die Goldkinder The Gold-Children 85 79 0.22 32.91 22.78 44.3 .66 50 Die Hochzeit der Frau Füchsin Mrs. Fox's Wedding 38 46 0.12 23.91 23.91 52.17 .78 51 Die Krystallkugel The Crystal Ball 197 36 -0.13 30.56 36.11 33.33 .67 52 Die Nelke The Carnation 76 64 0.06 28.13 29.69 42.19 .57 53 Die Nixe im Teich The Nixie in the Pond 181 96 -0.11 22.92 29.17 47.92 .65 54 Die Rabe The Raven 93 88 -0.2 25 39.77 35.23 .74 55 Die sechs Diener The Six Servants 134 100 0.6 46 13 41 .52 56 Die sechs Schwäne The Six Swans 49 78 -0.25 29.49 44.87 25.64 .88 57 Die sieben Raben The Seven Ravens 25 37 -0.27 18.92 43.24 37.84 .63 58 Die Wassernixe The Water Nixie 79 10 -0.05 20 20 60 .68 59 Die weisse Schlange The White Snake 17 58 0.15 29.31 25.86 44.83 .47 60 Die Wichtelmänner The Elves 39 40 0.83 55 12.5 32.5 .80 61 Die zwei Brueder The Two Brothers 60 307 0.15 28.34 21.17 50.49 .41 62 Die zwölf Brueder The Twelve Brothers 9 74 -0.28 27.03 40.54 32.43 .80 63 Dornröschen Little Brier-Rose 50 42 0.49 40.48 16.67 42.86 .75 64 Frau Holle Frau Holle 24 42 -0.55 33.33 57.14 9.52 .84 65 Fundevogel Foundling-Bird 51 42 0.27 35.71 19.05 45.24 .55 66 Hänsel und Grethel Hansel and Gretel 15 125 -0.38 20.8 40 39.2 .66 67 Hans mein Igel Hans-My-Hedgehog 108 67 -0.02 25.37 29.85 44.78 .76 68 Jorinde und Joringel Jorinde and Joringel 69 44 0.02 25 27.27 47.73 .62 69 Katze und Maus in Gesellschaft Cat and Mouse in Partnership 2 42 -0.3 16.67 26.19 57.14 .27 70 Märchen von der Unke Tales of the Toad 105 20 -0.05 25 20 55 .88 71 Marienkind Mary's Child 3 65 -0.55 24.62 50.77 24.62 .86 72 Rapunzel Rapunzel 12 51 -0.41 27.45 45.1 27.45 .90 73 Rothkäppchen Little Red Cap 26 59 -0.08 13.56 22.03 64.41 .65 74 Rumpelstilzchen Rumpelstiltskin 55 37 -0.34 18.92 29.73 51.35 .45 75 Schneeweisschen und Rosenroth Snow-White and Rose-Red 161 87 -0.22 34.48 42.53 22.99 .83 76 Sechse kommen durch die ganze Welt How Six Men Got On in the World 71 72 0.52 44.44 18.06 37.5 .71 77 Simeliberg Simeli Mountain 142 31 -0.26 22.58 29.03 48.39 .88 78 Sneewittchen Little Snow-White 53 131 -0.61 19.85 50.38 29.77 .65 79 Spindel, Weberschiffchen und Nadel Spindle, Shuttle, and Needle 188 41 0.51 39.02 9.76 51.22 .68 80 Tischchen deck dich, Goldesel, und Knüppel aus dem Sack Table-Be-Set, Gold-Donkey, and Cudgel-out-of-the-Sack 36 151 0,6 49.01 15.23 35.76 .59

sum

5579

mean

-0.008

mean

29.06

mean

29.26

mean

41.69

mean

.64

1 Titles are translated according to https://sites.pitt.edu/~dash/grimmtales.html

2 For technical reasons this fairy tale was annotated by only one person.

Tab. S1: Overview of all annotated fairy tales in ChildTale-A (N = 80): ID (used in Figures 3 and 5), German and English title, CHT-Index, text length (number of sentences), Average Valence (mean of valence annotations of all sentences), percentages of negative, neutral, and positive sentences (sent.), and Krippendorff’s alpha (Kripp. α) for valence annotations of the sentences.

References Antti Aarne: The Types of the Folk-tale. A Classification and Bibliography. Translated and enlarged by Stith Thompson. Helsinki 1961. Nasir Ahmed / T. Natarajan / Kamisetty Ramamohan Rao: Discrete Cosine Transform. In: IEEE Transactions on Computers 23 (1974), i. 1, pp. 90–93. DOI: 10.1109/T-C.1974.223784 Henrike F. Alfes: Literatur und Gefühl: Emotionale Aspekte literarischen Schreibens und Lesens. Wiesbaden 1995. Cecilia Ovesdotter Alm / Richard Sproat: Emotional Sequencing and Development in Fairy Tales. In: International Conference on Affective Computing and Intelligent Interaction. (ACII 2005: Beijing, 22.10.–24.10.2005) Beijing 2005, pp. 668–674. DOI: 10.1007/11573548_86 Thomas Anz: Kulturtechniken der Emotionalisierung: Beobachtungen, Reflexionen und Vorschläge zur literaturwissenschaftlichen Gefühlsforschung. In: Im Rücken der Kulturen. Ed. by Karl Eibl / Katja Mellmann / Rüdiger Zymner. Paderborn 2007, pp. 207–239. Jodie Archer / Matthew Lee Jockers: The Bestseller Code: Anatomy of the Blockbuster Novel. New York 2016. Hermann Bausinger: Märchenglück »Happiness in the fairy-tale«. In: Zeitschrift für Literaturwissenschaft und Linguistik 50 (1983), i. 13, pp. 17–27. Emotionalität. Zur Geschichte der Gefühle. Ed. by Claudia Benthien / Anne Fleig / Ingrid Kasten. Köln 2000. Yves Bestgen: Can Emotional Valence in Stories be Determined from Words? In: Cognition & Emotion 8 (1994), i. 1, pp. 21–36. DOI: 10.1080/02699939408408926 Lothar Bluhm: Märchen als Literatur aus Literatur: Die »Kinder- und Hausmärchen« der Brüder Grimm. Abhandlungen zur Literaturwissenschaft. Berlin 2022. DOI: 10.1007/978-3-662-64644-1 Margaret M. Bradley / Peter J. Lang: Affective Norms for English Words (ANEW): Instruction Manual and Affective Ratings. Gainesville 1999. PDF. [online] Sven Buechel / Udo Hahn: Readers vs. Writers vs. Texts: Coping with Different Perspectives of Text Understanding in Emotion Annotation. In: Proceedings of the 11th Linguistic Annotation Workshop. (LAW: Valencia, 03.04.2017) Valencia 2017, pp. 1–12. DOI: 10.18653/v1/W17-0801 David Comberg: Kurt Vonngeut on the Shapes of Stories. YouTube. 30.10.2010. [online] Alan Cowen / Jessica L. Tracy Disa Sauter / Dacher Keltner: Mapping the Passions: Toward a High-Dimensional Taxonomy of Emotional Experience and Expression. In: Psychological Science in the Public Interest 20 (2019), i. 1, pp. 69–90. DOI: 10.1177/1529100619850176 Louise D’Arcens: Emotions in Fiction. In: Sources for the History of Emotions. Ed. by Katie Barclay / Sharon Crozier-De Rosa / Peter N. Stearns. Abingdon 2020, pp. 114–126. Katrin Dennerlein / Thomas Schmidt / Christian Wolff: Computational Emotion Classification for Genrecorpora of German Plays from 17th to 19th Century. In: Digital Scholarship in the Humanities (DSH). [In appearance] Peter Sheridan Dodds / Kameron Decker Harris / Isabel M. Kloumann / Catherine A. Bliss / Christopher M. Danforth: Temporal Patterns of Happiness and Information in a Global Social Network: Hedonometrics and Twitter. In: PLOS ONE 6 (2011), i. 12, pp. 1–26. DOI: 10.1371/journal.pone.0026752 Peter Sheridan Dodds / Eric M. Clark / Suma Desu / Morgan R. Frank / Andrew J. Reagan / Jake Ryland Williams / Lewis Mitchell / Kameron Decker Harris / Isabel M. Kloumann / James P. Bagrow / Karine Megerdoomian / Matthew T. McMahon / Brian F. Tivnan / Christopher M. Danforth: Human Language Reveals a Universal Positivity Bias. In: Proceedings of the National Academy of Sciences 112 (2015), i. 8, pp. 2389–2394. DOI: 10.1073/pnas.1411678112 Paul Ekman: An Argument for Basic Emotions. In: Cognition & Emotion 6 (1992), i. 3–4, pp. 169–200. DOI: 10.1080/02699939208411068 Katherine L. Elkins: The Shapes of Stories: Sentiment Analysis for Narrative. Cambridge, UK et al. 2022. DOI: 10.1017/9781009270403 Micha Elsner: Abstract Representations of Plot Structure. In: Linguistic Issues in Language Technology 12 (2015), pp. 1–29. DOI: 10.33011/lilt.v12i.1381 Jakob Fehle / Thomas Schmidt / Christian Wolff: Lexicon-based Sentiment Analysis in German: Systematic Evaluation of Resources and Preprocessing Techniques. In: Proceedings of the 17th Conference on Natural Language Processing. (KONVENS 2021: Düsseldorf, 06.09.–09.09.2021) Düsseldorf 2021, pp. 86–103. PDF. [online] Mark Finlayson: Learning Narrative Structure from Annotated Folktales. Cambridge, MA 2012. [online] Kurt W. Fischer / Philip R. Shaver / Peter Carnochan: How Emotions Develop and How They Organise Development. In: Cognition and Emotion 4 (1990), i. 2, pp. 81–127. DOI: 10.1080/02699939008407142 Marie Flüh: Emotionsanalyse. In: forTEXT. Literatur digital erforschen. 2020. HTML. [online] Matthias Gamer / Jim Lemon / Ian Fellows Puspendra Singh: Package ›irr‹: Various Coefficients of Interrater Reliability and Agreement. 2019. HTML. [online] Sabine Geck: »Du dauerst mich!« La expresión de las emociones en los cuentos de los hermanos Grimm. In: Revista de Filología Alemana 22 (2014), pp. 169–188. DOI: 10.5209/rev_RFAL.2014.v22.45315 Anja Gerigk: Freude. In: Handbuch Literatur & Emotionen. Ed. by Martin von Koppenfels / Cornelia Zumbusch. Berlin 2016, p. 546. Clarence Green: Introducing the Corpus of the Canon of Western Literature: A Corpus for Culturomics and Stylistics. In: Language and Literature 26 (2017) pp. 282–99. DOI: 10.1177/0963947017718996 Giulia Grisot / Berenike Herrmann: Examining the Representation of Landscape and its Emotional Value in German-Swiss Fiction Around 1900. Journal of Cultural Analytics (2022). [Submitted] Gerlind Grosse / Berit Streubel / Catherine Gunzenhauser / Henrik Saalbach: Let’s Talk About Emotions: the Development of Children’s Emotion Vocabulary from 4 to 11 Years of Age. In: Affective Science 2 (2021), i. 2, pp. 150–162. DOI: 10.1007/s42761-021-00040-2 Kevin A. Hallgren: Computing Inter-Rater Reliability for Observational Data: An Overview and Tutorial. In: Tutorials in Quantitative Methods for Psychology 8 (2012), i. 1, pp. 23–34. DOI: 10.20982/tqmp.08.1.p023 David Hanauer: The Genre-specific Hypothesis of Reading: Reading Poetry and Encyclopedic Items. In: Poetics 26 (1998), 2, pp. 63–80. DOI: 10.1016/S0304-422X(98)00011-4 Claudia Hillebrandt: Das emotionale Wirkungspotenzial von Erzähltexten: mit Fallstudien zu Kafka, Perutz und Werfel. Berlin 2011.DOI: 10.1524/9783050057163 The Routledge Companion to Literature and Emotion. Ed. by Patrick Colm Hogan / Bradley J. Irish / Lalita Pandit Hogan. (= Routledge Literature Companions). London et al. 2022. Katalin Horn (2016a): Gefühle. In: Enzyklopädie des Märchens Online. Ed. by Rolf Wilhelm Brednich / Heidrun Alzheimer / Hermann Bausinger / Wolfgang Brückner / Daniel Drascek / Helge Gerndt / Ines Köhler-Zülch / Klaus Roth / Hans-Jörg Uther. Berlin 2016. DOI: 10.1515/emo.5.126 Katalin Horn (2016b): Helfer. In: Enzyklopädie des Märchens Online. Ed. by Rolf Wilhelm Brednich / Heidrun Alzheimer / Hermann Bausinger / Wolfgang Brückner / Daniel Drascek / Helge Gerndt / Ines Köhler-Zülch / Klaus Roth / Hans-Jörg Uther. Berlin 2016. DOI: 10.1515/emo.6.157 Arthur M. Jacobs: Neurocognitive Poetics: Methods and Models for Investigating the Neuronal and Cognitive-affective Bases of Literature Reception. In: Frontiers in Human Neuroscience 9 (2015), pp. 1–22. DOI: 10.3389/fnhum.2015.00186 Arthur M. Jacobs / Annette Kinder: Computing the Affective-Aesthetic Potential of Literary Text. In: AI 1 (2020), i. 1, pp. 11–27. DOI: 10.3390/ai1010002 Arthur M. Jacobs / Berenike Herrmann / Gerhard Lauer / Jana Lüdtke / Sascha Schroeder: Sentiment Analysis of Children and Youth Literature: Is There a Pollyanna Effect? In: Frontiers in Psychology 11 (2020), pp. 1–8. DOI: 10.3389/fpsyg.2020.574746 Fotis Jannidis: Figur und Person: Beitrag zu einer historischen Narratologie. Berlin et al. 2004. DOI: 10.1515/9783110201697 Matthew Jockers: A Novel Method for Detecting Plot. 2014. HTML. [online] Matthew Jockers: Revealing Sentiment and Plot Arcs with the Syuzhet Package. 2015. HTML. [online] Matthew Jockers (2020a): Introduction to the Syuzhet Package. 2020. HTML. [online] Matthew Jockers (2020b): Syuzhet. An R package for the extraction of sentiment and sentiment-based plot arcs from text. Version 1.0.6. 2020. [online] André Jolles: Einfache Formen. Legende, Sage, Mythe, Rätsel, Spruch, Kasus, Memorabile, Märchen, Witz. Halle / Saale 1930. Steven Swann Jones: The Fairy Tale: The Magic Mirror of the Imagination. New York et al. 2002. Johanna K. Kaakinen / Egon Werlen / Yvonne Kammerer / Cengiz Acartürk / Xavier Aparicio / Thierry Baccino / Ugo Ballenghein / Per Bergamin / Núria Castells / Armanda Costa / Isabel Falé / Olga Mégalakaki / Susana Fernández: IDEST: International Database of Emotional Short Texts. In: PLOS ONE 17 (2022), i. 10, pp. 1–21. DOI: 10.1371/journal.pone.0274480 Evgeny Kim / Sebastian Padó / Roman Klinger: Investigating the Relationship between Literary Genres and Emotional Plot Development. In: Proceedings of the Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature. (LaTeCH-CLfL 2017: Vancouver, 03.08.–04.08.2017) Vancouver 2017, pp. 17–26. DOI: 10.18653/v1/W17-2203 Evgeny Kim / Roman Klinger: A Survey on Sentiment and Emotion Analysis for Computational Literary Studies. In: Zeitschrift für digitale Geisteswissenschaften 4 (2019). HTML. DOI: 10.17175/2019_008 Roman Klinger / Surayya Samat Suliya / Nils Reiter: Automatic Emotion Detection for Quantitative Literary Studies. A Case Study Based on Franz Kafka’s »Das Schloss« and »Amerika«. In: Digital Humanities 2016: Conference Abstracts. (DH2016: Kraków, 11.07.–16.07.2016) Kraków 2016. HTML. [online] Handbuch Literatur & Emotionen. Ed. by Martin von Koppenfels / Cornelia Zumbusch. Berlin et al. 2016. Richard J. Landis / Gary G. Koch: The Measurement of Observer Agreement for Categorical Data. In: Biometrics 33 (1977), i. 1, pp. 159–174. DOI: 10.2307/2529310 Peter J. Lang / Margaret M. Bradley / Bruce N. Cuthbert: International Affective Picture System (IAPS). Technical Manual and Affective Ratings. Technical Report A-8. University of Florida, Gainesville, FL. (2008). [online] Stella Lange: Gefühle schwarz auf weiß. Implizieren, Beschreiben und Benennen von Emotionen im empfindsamen Briefroman um 1800. Heidelberg 2016. Heather C. Lench / Sarah A. Flores / Shane W. Bench: Discrete Emotions Predict Changes in Cognition, Judgment, Experience, Behavior, and Physiology: A Meta-analysis of Experimental Emotion Elicitations. In: Psychological Bulletin 137 (2011), pp. 834–855. DOI: 10.1037/a0024244 Robert W. Levenson: Basic Emotion Questions. In: Emotion Review 3 (2011), i. 4, pp. 379–386. DOI: 10.1177/1754073911410743 Claude Levi-Strauss: Myth and Meaning. New York 1995. Kirsten A. Lindquist: Language and Emotion: Introduction to the Special Issue. In: Affective Science 2 (2021), pp. 91–98. DOI: 10.1007/s42761-021-00049-7 Jurij Michajlovič Lotman: Über das Wechselverhältnis von ›primär‹ und ›sekundär‹ in kommunikativen modellbildenden Systemen. In: Kunst als Sprache. Untersuchungen zum Zeichencharakter von Literatur und Kunst. Leipzig 1981, pp. 111–115. Jurij Michajlovič Lotman: Die Struktur literarischer Texte. 4th ed. München 1993. Jana Lüdtke / J. Berenike Herrmann: childTale-A: A corpus of eighty fairy tales from the 7th edition by the Brothers Grimm, manually annotated for textually encoded emotions. Data set. Zenodo. DOI: 10.5281/zenodo.7737329 Jana Lüdtke / Arthur Jacobs: The Emotion Potential of Simple Sentences: Additive or Interactive Effects of Nouns and Adjectives? In: Frontiers in Psychology 6 (2015), pp. 1–15. DOI: 10.3389/fpsyg.2015.01137 Max Lüthi: Märchen. 10th edition, revued and ed. by Heinz Rölleke. Metzler 2004. DOI: 10.1007/978-3-476-04156-2 Max Lüthi: Das europäische Volksmärchen. 11th edition. Tübingen 2005. Max Lüthi: Affekte. In: Enzyklopädie des Märchens Online. Ed. by Rolf Wilhelm Brednich / Heidrun Alzheimer / Hermann Bausinger / Wolfgang Brückner / Daniel Drascek / Helge Gerndt / Ines Köhler-Zülch / Klaus Roth / Hans-Jörg Uther. Berlin et al. 2016. DOI: 10.1515/emo Andrew Lynch: The History of Emotions and Literature. In: The Routledge Companion to Literature and Emotion. Ed. by Patrick Colm Hogan / Bradley J. Irish / Lalita Pandit Hogan. (= Routledge Literature Companions). London et al. 2022, pp. 98–109. Katja Mellmann: Emotionalisierung – Von der Nebenstundenpoesie zum Buch als Freund. Eine emotionspsychologische Analyse der Literatur der Aufklärungsepoche. Paderborn 2006. Katja Mellmann: Emotionalisieren. In: Erzählen: Ein interdisziplinäres Handbuch. Ed. by Matías Martínez. Stuttgart 2017, pp. 243–49. DOI: 10.1007/978-3-476-05364-0_36 Burkhard Meyer-Sickendiek: Affektpoetik. Eine Kulturgeschichte literarischer Emotionen. Würzburg 2005. Saif Mohammad: From Once Upon a Time to Happily Ever After: Tracking Emotions in Novels and Fairy Tales. In: Proceedings of the 5th ACL-HLT Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities. (LaTeCH '11: Portland, 24.06.2011) Portland, OR 2011, pp. 105–114. PDF. [online] Julia Nantke: Annäherungen an eine digitale Semiotik: Chancen und Grenzen computergestützter Untersuchungsmethoden für die semiotische Analyse literarischer Texte. In: Zeitschrift für Semiotik 39 (2017), i. 1-2, pp. 83–108. Stefan Neuhaus: Märchen. 2nd revued edition. Tübingen 2017. Bo Pang / Lillian Lee: Opinion Mining and Sentiment Analysis. In: Foundations and Trends in Information Retrieval 2 (2008), i. 1–2, pp. 1–135. DOI: 10.1561/1500000011 Jaak Panksepp / Douglas Watt: What is Basic about Basic Emotions? Lasting Lessons from Affective Neuroscience. In: Emotion Review 3 (2011), i. 3, pp. 387–396. DOI: 10.1177/1754073911410741 Franz Petermann / Silvia Wiedebusch: Emotionale Kompetenz bei Kindern. 3rd edition. Göttingen 2016. Jonathan Posner / James A. Russell / Bradley S. Peterson: The Circumplex Model of Affect: an Integrative Approach to Affective Neuroscience, Cognitive Development, and Psychopathology. In: Development and Psychopathology 17 (2005), i. 3, pp. 715–734. DOI: 10.1017/S0954579405050340 Gerald Prince: A Grammar of Stories: An Introduction. The Hague 1973. DOI: 10.1515/9783110815900 Vladimir Propp: Morphology of the Folktale. 2nd edition, revised and edited. Austin 1968. R Core Team: R: A Language and Environment for Statistical Computing. 2021. HTML. [online] Andrew J. Reagan / Lewis MItchell / Dilan Kiley / Christopher M. Danforth / Peter Sheridan Dodds: The Emotional Arcs of Stories are Dominated by Six Basic Shapes. In: EPJ Data Science 5 (2016), i. 1, pp. 1–12. DOI: 10.1140/epjds/s13688-016-0093-1 Rainer Reisenzein / Gernot Horstmann / Achim Schützwohl: The Cognitive-Evolutionary Model of Surprise: A Review of the Evidence. In: Topics in Cognitive Science 11 (2019), i. 1, pp. 50–74. DOI: 10.1111/tops.12292 Lutz Röhrich: Märchen mit schlechtem Ausgang. In: Hessische Blätter für Volkskunde 49–50 (1958), pp. 236–248. Lutz Röhrich: Grausamkeit. In: Enzyklopädie des Märchens Online. Ed. by Rolf Wilhelm Brednich / Heidrun Alzheimer / Hermann Bausinger / Wolfgang Brückner / Daniel Drascek / Helge Gerndt / Ines Köhler-Zülch / Klaus Roth / Hans-Jörg Uther. Berlin et al. 2016. DOI: 10.1515/emo Heinz Rölleke: Die Märchen der Brüder Grimm: Eine Einführung. Stuttgart 2004. Gabriela Rotari: Digital Analysis of Emotions in the Brothers Grimms Fairy Tales. In: EADH 2018: »Data in Digital Humanities«. (EADH 2018: Galway, 07.12.–09.12.2018) Galway 2018. PDF. [online] Paul Rozin / Edward B. Royzman: Negativity Bias, Negativity Dominance, and Contagion. In: Personality and Social Psychology Review 5 (2001). pp. 296–320. DOI: 10.1207/S15327957PSPR0504_2 James A. Russell: A Circumplex Model of Affect. In: Journal of Personality and Social Psychology 39 (1980), pp. 1161–78. DOI: 10.1037/h0077714 Rudolf Schenda: Volk ohne Buch. Studien zur Sozialgeschichte der populären Lesestoffe 1770–1910. München 1977. Thomas Schmidt / Katrin Dennerlein / Christian Wolff: Emotion Classification in German Plays with Transformer-based Language Models Pretrained on Historical and Contemporary Language. In: Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature. (LaTeCH-CLfL 2021: Punta Cana, 07.11.–11.11.2021) Punta Cana 2021, pp. 67–79. DOI: 10.18653/v1/2021.latechclfl-1.8 David S. Schmidtke / Tobias Schröder / Arthur M. Jacobs / Markus Conrad: ANGST: Affective Norms for German Sentiment Terms, Derived from the Affective Norms for English Words. In: Behavior Research Methods 46 (2014), i. 4, pp. 1108–1118. DOI: 10.3758/s13428-013-0426-y Monika Schwarz-Friesel: Das Emotionspotenzial literarischer Texte. In: Handbuch Sprache in der Literatur. Ed. by Anne Betten / Berbeli Wanning / Ulla Fix. Berlin 2017. DOI: 10.1515/9783110297898-016 Nadine Stamm: Klassifikation und Analyse von Emotionswörtern in Tweets für die Sentimentanalyse. Zürich 2014. PDF. [online] Jessica L. Tracy / Daniel Randles: Four Models of Basic Emotions: A Review of Ekman and Cordaro, Izard, Levenson, and Panksepp and Watt. In: Emotion Review 3 (2011), i. 4, pp. 397–405. DOI: 10.1177/1754073911410747 UNESCO: Memory of the World. Kinder- und Hausmärchen (Children’s and Household Tales). October 2004. Revised version of July 2011. PDF. [online] Hans-Jörg Uther: The Types of International Folktales. A Classification and Bibliography. Based on the System of Antti Aarne and Stith Thompson. Helsinki 2004. Hans-Jörg Uther: Handbuch zu den »Kinder- und Hausmärchen« der Brüder Grimm: Entstehung - Wirkung - Interpretation. 2nd revued edition. Berlin 2013. Jos Josephus Augustinus van Berkum: A Survey of Emotion Theories and their Relevance to Language Research. Berlin 2022. DOI: 10.1515/9783110347524-001 Teun Adrianus van Dijk / Walter Kintsch: Strategies of Discourse Comprehension. New York 1983. Michael Vauth / Hans Ole Hatzel / Evelyn Gius / Chris Biemann: Automated Event Annotation in Literary Texts. In CHR 2021: Computational Humanities Research Conference. Amsterdam, The Netherlands (2021), pp. 333–345. PDF. [online] Heinz-Günter Vester: Emotion, Gesellschaft und Kultur: Grundzüge einer soziologischen Theorie der Emotionen. Opladen 1991. Melissa L. H. Võ / Markus Conrad / Lars Kuchinke / Karolina Urton / Markus J. Hofmann / Arthur M. Jacobs: The Berlin Affective Word List Reloaded (BAWL-R). In: Behavior Research Methods 41, (2009), pp. 534–38. DOI: 10.3758/BRM.41.2.534 Christiane Voss: Narrative Emotionen: Eine Untersuchung über Möglichkeiten und Grenzen philosophischer Emotionstheorien. Berlin, Boston 2004. DOI: 10.1515/9783110896268 Amy B. Warriner / Victor Kuperman / Marc Brysbaert: Norms of Valence, Arousal, and Dominance for 13,915 English Lemmas. In: Behavior Research Methods 45, (2013), pp. 1191–207. DOI: 10.3758/s13428-012-0314-x Raphael Winkelmann / Jonathan Harrington / Klaus Jänsch: EMU-SDMS: Advanced Speech Database Management and Analysis in R. In: Computer Speech & Language 45 (2017), pp. 392–410. DOI: 10.1016/j.csl.2017.01.002 Simone Winko: Kodierte Gefühle: Zu einer Poetik der Emotionen in lyrischen und poetologischen Texten um 1900. Berlin 2003. Simone Winko: Literature and Emotion. In: Language and Emotion: An International Handbook (= Handbücher zur Sprach- und Kommunikationswissenschaft, 46, 3) 3 vol. Ed. by Gesine Lenore Schiewer / Jeanette Altarriba / Bee Chin Ng. Berlin 2022-2023 et al. Vol. 3, pp. 1417-1436. DOI: 10.1515/9783110795486-004 Markus Antonius Wirtz / Franz Caspar: Beurteilerübereinstimmung und Beurteilerreliabilität: Methoden zur Bestimmung und Verbesserung der Zuverlässigkeit von Einschätzungen mittels Kategoriensystemen und Ratingskalen. Göttingen 2002. Albin Zehe / Martin Becker / Fotis Jannidis / Andreas Hotho: Towards Sentiment Analysis on German Literature. In: KI 2017: Advances in Artificial Intelligence. Ed. by Gabriel Kern-Isberner / Johannes Fürnkranz / Matthias Thimm. Cham 2017, pp. 387–394. DOI: 10.1007/978-3-319-67190-1_36 Albin Zehe / Martin Becker / Lena Hettinger / Andreas Hotho / Isabella Reger / Fotis Jannidis: Prediction of Happy Endings in German Novels Based on Sentiment Information. In: Interactions between Data Mining and Natural Language Processing. 3rd Workshop on Interactions between Data Mining and Natural Language Processing. (ECML-PKDD 3: Riva del Garda, 19.09.–23.09.2016) Riva del Garda 2016, S. 9–16. PDF. [online] Jack Zipes: Why Fairy Tales Stick: The Evolution and Relevance of a Genre. New York 2006.

List of Figures and Tables Tab. 1: Annotation schemas and specific instructions for each category. Tab. 2: Reliability of annotations for valence, arousal, and the six basic emotions calculated seperately for each fairy tale as averages of Krippendorff’s alpha and percentage of agreement. Fig. 1: Histogram, density curve and Boxplot for the Avergae Valence values of all annotated fairy tales. The y-axis shows the relative number of texts per valence segment, the dashed auxiliary line depicts the overall mean = -0.008. Fig. 2: A – Relative frequency of the negative, neutral, and positive sentence (in %) in all fairy tales and the results of pairwise comparisons, B – Histogram and boxplot for the Emotion Potential (relative frequency (in %) of emotional sentences in the fairy tales, black dashed line indicates 50 %, red dashed line indicates the overall corpus meanEP = 58.31), C - Scatterplot of the relative frequency of positive and negative sentences (in %) per fairy tale (individual texts coded by color, dashed lines indicate 50 %). [Graphic: Berenike Herrmann / Jana Lüdtke 2023] Fig. 3: Emotional Arcs of all fairy tales in ChildTale-A based on DCT-smoothed valence annotations. Narrative time is normalized to a time window from 1 to 100. The Emotional Arcs are ordered according to the ascending range of the emotion trajectories, defined as the difference between the highest and lowest smoothed valence value. The assignment of the IDs to the fairy tales can be found in Table S1. Black dashed lines indicate the theoretical mean of the valence scale. [Graphic: Berenike Herrmann / Jana Lüdtke 2023] Fig. 4: Relationship between fairy tale length and range of DCT-smoothed valence annotations (A), fairy tale length and range of the original valence annotations (valence span) (B), and Emotion Potential (percent of positive and negative sentences) and range of DCT-smoothed valence annotations (C). [Graphic: Berenike Herrmann / Jana Lüdtke 2023] Fig. 5: Trajectory of the changes in the original valence annotation (values between the dashed auxiliary lines correspond to sentences annotated as neutral) and the trajectory of the Emotional Arc, i. e. the DCT-smoothed valence annotations (m = 5) for the fairy tales ID34 and ID02. [Graphic: Berenike Herrmann / Jana Lüdtke 2023] Fig. 6: Boxplots of relative frequencies of the six basic emotions in all fairy tales (absolute frequencies divided by number of sentences per fairy tale). Dashed line represents the overall mean. Mean values of each basic emotion are compared to overall mean. [Graphic: J. Berenike Herrmann / Jana Lüdtke 2023] Fig. 7: Correlation matrix with spearman correlation coefficients between the relative frequency of the six basic emotions, the Emotion Potential (percent of emotional sentences), Average Valence, range in the Emotional Arc (the DCT-smoothed valence values), and length (number of sentences) for all fairy tales (higher positive and negative correlation coefficient are marked red or blue, respectively; significant values are marked with asterisks: * p < .05, ** p < .01, *** p < .001). [Graphic: Berenike Herrmann / Jana Lüdtke 2023] Fig. 8: Emotion Profiles for all fairy tales in ChildTale-A. Fairy tales are ordered as in Figure 3, according to the range of the Emotional Arcs (the difference between the highest and lowest DCT-smoothed valence values). For the legend of the fairy tale IDs, see Table S1. [Graphic: Berenike Herrmann / Jana Lüdtke 2023] Tab. 3: Relative frequency (in %) of the (non-)occurrence of the six basic emotions for negative neutral and positive sentences. Tab. S1: Overview of all annotated fairy tales in ChildTale-A (N = 80): ID (used in Figures 3 and 5), German and English title, CHT-Index, text length (number of sentences), Average Valence (mean of valence annotations of all sentences), percentages of negative, neutral, and positive sentences (sent.), and Krippendorff’s alpha (Kripp. α) for valence annotations of the sentences.