In this study, we use the development of emotions in German-language poetry, drama, and prose from 1850 to 1920 to informally test three hypotheses about literature: (1) Literature is a unified field, and therefore genres develop similarly at the same time. (2) The development of literature is led by one genre while the others follow. (3) The three main genres have very diferent developments without any relation to each other. We look at the development of emotions in these genres in general, and then at more fine-grained levels: polarity, six groups of emotions, and the group of love emotions. In the end, our data cannot confirm any of these hypotheses, but do show a closer relationship between poetry and prose, while drama shows a very distinct development. Only in some specific cases, such as the representation of lust and of love, can we see a closer relationship between the genres in general.
The abstract term ‘literature’ is a relatively late invention in English and German, which
conceptually bundles the very diferent histories of the individual major genres (prose, drama, poetry)
and even the terms for some of the major genres are comparatively late abstractions, such as in
the case of the German word for poetry (‘Lyrik’) which came into use around 1800. The
undeniable productivity of the term ‘literature’ lies in the fact that commonalities are emphasized and
marked as essential. The disadvantage is that any use of such an abstract notion may suggest
a unified entity where there are rather heterogeneous sub-fields. Traditional literary studies
and computational literary studies (CLS) tend to see themselves in opposing camps as far as
the tendency to use such abstractions is concerned. On the one hand, even well-established
genre terms are questioned in cultural studies as fictions1(2[]), while CLS uses (e.g. [
The selection of genres that we include in the study is based on the classical triad and each of these major genres is represented by a specific subgenre. Our prose corpus, for example, only contains novels published in print, which excludes entertainment literature published in other formats as well as novellas, short stories and short prose. For pragmatic reason, we must be selective not only in the constitution of the corpus, but also in which of the numerous possible perspectives on the literary texts we take into consideration. Obvious candidates would be aspects like themes, formal structures or character types. Since contemporaries have already determined the change from realism to early modernism by, among other things, how and which emotions are thematized (e.g., 3[8, p. 736]; [9, p. 338f.]; [29, p. 274]), we will concentrate on this perspective in this study. In short, the objective of this study is to investigate the interrelationship between the genres of prose, drama, and poetry during the period spanning from 1850 to 1920 by examining the manner in which these genres thematize and express emotions. We will start our analysis with a closer look at the general frequency of emotions in the three genres from 1850 to 1920. In the next step we look separately at positive and negative emotions and then on six emotion groups: agitation, anger, fear, joy, love, sadness. Our last step is an analysis of the emotions grouped together in ‘love’: ‘lust’, ‘afection’, ‘longing’, and ‘lovesub’. This drilling down into more and more fine-grained perspectives allows us at the one hand to understand what powers the relations we can perceive on the more abstract level, but it also allows to determine whether we can identify counter-histories which subvert the main trends.
Our approach can be linked to various literary-historical studies and in part test their theses: We highlight some of the debates relevant to our study, focusing on scholarship concerned with 19th- and 20th-century German-language literature. We organize them according to the three guiding hypotheses about the development of literature and genres with regard to the representation of emotions.
(1) The literary system develops as a whole, independently of the development of genres. Although the history of emotions in literature has not yet been well studied, some observations can be made how the literary representation of emotions has changed over time. According to Scheuer, a literary history of emotional representations would have to take into account the ‚increasing sentimentalization‘; since the Romantic period there have been literary tendencies of ‚exaggeration of feeling‘ („Gefühlsübersteigerung“3,6[, p. 19]), even into the 20th century. According to this approach, these tendencies manifest themselves around 1900 in phenomena as diverse as the aestheticistic dissection of one‘s own emotional state or in mixtures of sentimentality and brutality in naturalistic texts36([, p. 19f.]). From this perspective, no distinction is made between specific genres.
(2) Literary developments are driven by ‚leading genres‘. In literary histories we often find the assumption that there are such phenomena as leading genres in diferent epochs. These are genres in which the literary innovations that are typical or representative of a period appear ifrst and are particularly pronounced (cf.14[, p. 150]). Even if typicality or representativeness is measured in diferent ways – e.g. quantitatively via the number of publications, qualitatively via the programmatic or discursive relevance in a period – an asymmetrical interdependence between the genres is nevertheless assumed: The leading genre precedes, the others follow if they can. For realism, for example, it is claimed that prose is the leading genre, while drama and poetry are of lesser importance (cf3.,[p. 145f., 327f.]); for naturalism, it is stated that it began around 1885 in poetry (cf. 1[]: 45; [39, p. 533]), then manifested itself in prose and finally, from 1889, significantly influenced drama (cf. [ 1, p. 45]). Poetry is also usually cited as the decisive genre for the symbolist counter-movement to naturalism (c1f,. p[. 47]; [39, p. 533]). Based on our corpus, it is possible to examine how or if the developments of the individual genres relate to each other, even if we only focus on one aspect of genre development, the emotions.
(3) Literary genres develop independently. The representation of emotions has been studied most frequently in poetry. This is due to the fact that it has often been suggested that among prose, drama, and poetry, poetry is most closely tied to emotion. The association of poetry with emotionality, often linked to the notion of poetry as a particularly ‘subjective’ genre, was widespread in the 19th century and continued to have proponents in the 20th century (2cf5.,[ p. 56], [42, p. 119–125], [44, p. 25–28], [40, p. 79–81, 84]). For instance, viewing poetry as an expression or representation of emotions was part of the influential aesthetics of Hegel and Vischer ([15, p. 323, 419–421, 424]; [41, p. 1261, 1325]). In the 20th century, however, more scholars emphasized that emotionality is not a necessary feature of poetry, especially when considering poetry beyond the 19th century (cf2.5[, p. 57], [42, p. 120]). The debates outlined so far often take place on a theoretical level and aim to determine characteristics of poetry in terms of a genre definition. Whether poetry empirically represents emotions more often than other genres is a related but distinct question2.4[] found some evidence that German-language poetry from the second half of the 19th and early 20th centuries does indeed represent emotions more frequently than prose and drama. However, the corpus for that study was very small, containing only 5 prose texts and 5 dramas.
With regard to the emotional-historical development of poetry, it has been argued that
modernist poetry represents emotions less often than realist poetry2(,[p. 319]). Other researchers
contend that modernist poetry continues to represent emotions frequently, albeit in a
modified way ([
There is no generally recognized approach to exploring the relationship between genre and emotion in literary studies. Meyer-Sickendiek has proposed a conceptual approach to writing the history of literary genres as a history of emotions. He sees literary genres as ‚media‘ of basic emotions and assigns them each a leading emotion 2([8, p. 35f.]), e.g. the emotion ‚grief‘ to the elegy, ‚surprise‘ to the novella and ‚longing‘ to the melodrama. His starting point are the Poetics. In this approach, the modification of the leading emotions by individual authors over time proves to be relevant to cultural history. In contrast, we take the literary texts as our starting point and analyse the emotions that are actually presented in the individual texts, with these texts being assigned to the three main genres.
Sentiment Analysis and also the analysis of emotions has a comparatively long and rich history
in Computational Literary Studies2.1[] Emotions have also been used to distinguish genres and
subgenres.[
In methodological terms, our study is also comparable to Andrew Piper’s contribution on the concept of fiction, which also draws conclusions from the analysis of historical data for the theoretical discussion of central literary concep3t0s].[
Our poetry corpu1scomprises texts from 20 anthologies from the period under study. The
anthologies feature contemporary poetry. The collections from around 1900 additionally focus
on ‚modern‘ poetry. Often, the anthologies aim to showcase the ‚best‘ texts of their time. For
our corpus, we included all anthologized poems published between 1850 and 1920. For more
details, see [
The drama corpus consists of all texts in the German Drama Corpus (GerDraCor) from the period under study2.In addition, we included all texts from Project Gutenberg edition 16 from the same time span, which were marked as drama and which were not already in our corp3u4s].[ The prose corpus contains novels from DT3A,TextGrid4 and Project Gutenberg. 1Code and Data: https://github.com/cophi-wue/Unity-of-Literary-Change.git 2https://dracor.org/ger 3https://www.deutschestextarchiv.de/ 4https://textgrid.de/
In order to examine developments over time, the corpus texts had to be dated. For the plays taken from Dracor, we use the dates provided by DraCor. The additional plays, the poems and the prose texts have been dated by us. While it was relatively easy to find dates for plays and prose, it was much more difÏcult for poetry, in part because many poems were written by authors who are little known today and whose works are not or only partially catalogued or digitized. From the poems with reliable dating, we have calculated that an author is on average 33 years old at the time of publication. For the poems that cannot be dated precisely, we estimate the first publication to be 33 years after the author’s birth. In total, we were able to compile dates manually for 130 plays, 2821 poems and 1908 prose te5x. ts 5Including those that fall outside our period of investigation (86).
We annotated the representation of emotions in 1352 poems, 28 prose texts, and 23 plays. The
goal was not to annotate readers’ emotions, but rather the emotions represented in the text
itself, e.g., whether a character is happy, sad, in love, etc. The annotators used a list of 40
discrete emotions (e.g., longing, lust, joy, surprise, envy, regret). The selection is based on
existing emotion models (e.g. 7[], [
As we will analyze the emotion group ‚Love‘ in more detail later on, we also report the annotations of (selected) individual emotions that constitute this group: the individual emotion love in the narrower sense, e.g. in ‚I love you‘, but also longing, afection, and (sexual) lust. Therefore, we use the label ‚Love‘ to refer to the individual emotion love s(ulbo)v, ebut also to the emotion group.
In order to generate sufÏcient material for a statistical analysis of our entire corpus, we need to automate the annotation process. The task is redefined so that instead of annotating any kind of text span, one or more emotions are assigned to a predefined segment. The segments are chosen with respect to genre: stanza in poetry; lines in plays and paragraphs in novels. This reduces the complexity of the task from segmentation (identifying emotion-carrying text passages) and subsequent labelling (selecting one or more emotions) to a single multi-labelling process. As foundation model we choose SauerkrautLM-7B-Her6.OThe LLM is created by merging two mistral-7B models and further training on a German dataset. This choice is based on the fact that this model is freely available and can be handled with the computational resources at our disposal. Among the German models that meet these criteria, it performs best in general benchmark7 at the time of this study. Larger free models (e.g. Llama-3-SauerkrautLM-70bInstruct), as well as commercial services (e.g. gpt4), are ruled out due to the large amount of predictions which need to be generated for the novels in the corpus alone. We reduce the computational costs of finetuning the model by making use of QLoRA6[].
The model is quantised to 8bit and the rank of its learnable adaptation matrix is 4 with an alpha of 8. The training is performed over 30 epochs, the batchsize is 140 and the initial learningrate3 −4. Imbalance in label distribution is countered by random under-sampling. The quality of the model is determined separately for each genre in a 5-fold cross-validation (see table4).
The second task aims to further diferentiate the emotion of love from Shaver’s model. The task is to find one or more sub-emotions in segments with love emotion that caused the categorisation in the love group. We restricted the sub-emotions to identify to sluobv,elonging, afection and lust. Training is carried out as described for the first task. Tabl5e shows the evaluation results. Predictions are only generated for segments, that were labeled with love by the first model. 6https://huggingface.co/VAGOsolutions/SauerkrautLM-7b-HerO 7https://huggingface.co/datasets/VAGOsolutions/MT-Bench-TrueGerman
The data set generated is inherently complex, as the interplay of emotions, genres, and their evolution over time makes it difÏcult to provide a single comprehensive analysis that addresses all aspects equally.
Firstly, emotions difer in their frequency, and this variation also changes across diferent genres. Additionally, genres vary significantly in terms of the number of texts, the number of segments within those texts (e.g., novels typically have more paragraphs than poems have stanzas), and the length of these segments (paragraphs tend to be longer than speech acts or stanzas).
In the figures below, we opted to calculate the average per document rather than per segment within each genre. However, this approach has its limitations: treating a long novel and a short poem as single data points may place disproportionate weight on shorter texts. Moreover, genres with shorter texts exhibit more variability, as they are more likely to contain no or only one emotional instance. In contrast, genres with longer texts containing more segments are more likely to include a full range of emotions. Despite this, we chose to average by document, as it feels more intuitive and better suited to the subject matter than averaging by segments. The document-based average reflects the emotions encountered when picking up a random book and reading it, whereas averaging by segments would resemble reading a single page from a book, which is not a realistic representation.
This approach does not fully resolve the issue of varying emotion frequencies across genres. To visualize trends between genres more efectively, we applied z-transformations. Since these transformations express changes in terms of standard deviations from the mean within each genre, rather than relative or absolute frequencies, it allows us to compare values despite the difering characteristics of the genres.
The only exception arises when calculating emotion density. To provide a normalized value, we computed emotions per token. However, this measure is only a rough approximation, as not all tokens have the same likelihood of carrying emotion, due to the varying segment lengths. Smaller segments result in more frequent measurements per token, as we can only detect whether a segment contains an emotion or not.
In order to calculate statistical correlations between time series, stationarity must first be checked. Stationarity here means that the time series shows almost no trend and constant variability over the period under investigation. This property determines which statistical methods are permissible.
If two time series are stationary, we test their correlation for the same period (lag = 0) using the Pearson correlation coefÏcient. We can map whether there are time-shifted dependencies (lag > 1) using cross-correlation and test for significance with the Granger Causality te1s3t][. If both time series are not stationary, neither method can be used. Instead, we check for cointegration with the Engle-Granger test1[0].8 For both tests, we limit the maximum number of lags to 4 (corresponds to 12 years), as we consider a connection with a period further in the 8A table with all test results can be found in the repository. past to be implausible in this context. In the case that one time series is stationary and the other is non-stationary, we do not perform tests because the underlying statistical processes that generate the time series are so diferent that we rule out a relationship.
To have enough support for each data point in the time series, we work with 3-year bins (1850-1852, 1853-1855, etc.). For our visualizations in figure 3 to figure 9 we additionally smooth the time series by applying a moving average over a window of 3 bins. This reduces the variance the data show, but makes it much easier to visually identify interesting changes.
In the following sections, we analyze the representation of emotions in prose, drama, and poetry, progressing from the abstract to the specific. We begin by examining the overall frequency of emotions (5.1). Next, we distinguish between positive and negative emotions (5.2), followed by an analysis of the six emotion groups (love, joy, etc.) (5.3). Finally, we focus on the ‘Love’ emotion group, delving into its individual emotions, such as sexual lust and longing (5.4).
First, we look at the ‘emotional density’ of the genres (figure2). In terms of emotions per token, poetry represents emotions more often than drama and prose. This finding is consistent with the common notion that poetry is a particularly emotional genre. However, since a relevant proportion of poems do not represent emotions at all, emotionality is certainly not a necessary feature of poetry. Nevertheless, even at the beginning of the 20th Century, when the proportion of emotion in poetry has decreased substantially, it is still above the values in plays and and prose.
Figures3 and 4 show how the frequency of emotion representations changes over time. Figure 4 transforms the values from figure3 into the number of standard deviations from the genre mean. This z-transformation allows for another meaningful comparison of trends across genres by placing them on a common scale.
As expected, there is a decline in the frequency of emotion in poetry during early modernism, although this trend does not continue into the 1910s. The amount of emotion per token in prose also declines over the long term, at least until the 1890s. The frequency of emotion representation in drama follows a markedly diferent pattern. It increases rather than decreases. When we statistically test whether the time series follow similar patterns, we see no significant result in any case.
Our next step is to distinguish between positive and negative emotions (Figur5)e. We consider the emotions in the ‘Love’ and ‘Joy’ groups to be positive, and the emotions in the ‘Sadness’, ‘Fear’ and ‘Anger’ groups to be negative. ‘Agitation’ is discarded.
For poetry and prose the amount of positive emotions decreases up to 1890, but only for poetry this process continues. Drama shows a diferent pattern, with positive emotions becoming more frequent over the long term. The genres also follow diferent trajectories for negative emotions. While there is a long-term increase in negative emotions in drama, there is a decrease in prose, which seems to stop around 1900. From about 1880, poetry follows a similar pattern to prose. Statistical testing shows that the only significant relationship exists between prose and drama at 1 lag for both positive and negative emotions. The result is due to the opposite tendencies of the two genres: while positive and negative emotions decrease in prose, they increase in drama.
We further increase the level of detail in our analysis and focus on the six emotion groups,
which are based on the emotion hierarchy in3[
Examining the corresponding time series, we observe heterogeneous results (Fig7u)r.eThe development of emotions across diferent genres does not follow a uniform or straightforward pattern. There are some similarities, but only in limited cases. For example, both poetry and prose show a decrease in joy that stops around 1900 and a pattern of decline and increase in agitation from the 1880s onward, while their development with respect to other emotions is not very similar.
To further investigate whether emotion frequencies follow similar trajectories across genres, we plot their cross-correlation (Figu8r)e.9 The unlagged correlation of the poetry time series with the prose and drama time series is shown on the right side of the plots (lags = 0). High values at this point indicate that the frequency of a given emotion is increasing or decreasing simultaneously in poetry and prose, or poetry and drama. Additionally, we examine the correlation for diferent time lags by shifting the prose and drama time series by 1 to 4 lags (one lag equals three years), then recalculating the correlation. This approach reveals whether the time series follow similar patterns but with a delay. For instance, if there is a similar trend of increase and subsequent decline in a given emotion for both poetry and prose, but poetry begins this trajectory 1 or 2 lags earlier, the correlation without time shifts might be close to zero. However, with a shift of 1 or 2 lags, we might see that the correlation is quite high.
We see that anger in poetry and plays is not correlated at lag 0. But the correlation increases at lag -1. Which means that one time step after drama poetry shows a similar change of values. This fits with the observation in Figure 7 that there is an increase in anger after 1900 in all three genres, though it starts a few years earlier in drama. A similar pattern can be found for agitation.
The statistical tests show that for most combinations of two time series, their development
9We only compute cross-correlation for those time series which are stationary.
is not significantly related. The exceptions are the relationships between prose and drama
regarding joy and sadness with 1 to 3 lags. While both emotions decrease in prose, they tend
to increase in drama.
5.4. Love
In this section, we take a closer look at the ‘Love’ emotion group and its individual emotions.
Here, too, the results are heterogeneous. The plot on ‚lust‘ shows a development that is
consistent with literary-historical theses3(6[, p. 20]; [35, p. 153f.], [
The cross-correlation, shown in Figur1e0, reveals that there are some correlations between poetry and drama, but they are not very strong. In the case of afection, the correlation increases at -2 and -3 lags, but it must be borne in mind that afection is represented very rarely in poetry.
When we statistically test the relationships between time series, we see that there is a significant connection between play and prose in lust, and between diferent combinations of all three genres in lovseub. Among others, drama does not follow the decline of lsouvbein prose and drama.
The data don’t support the first hypothesis, that all genres move more or less in sync. The overall trends we observe in most of the data don’t support the second hypothesis either, that the development of one genre is the model for the other genres that follow with some time lag. In individual cases, however, we see developments that seem to fit the idea of a leading genre.
In particular, one of the most important innovations of modern literature, the representation of lust, is very pronounced in poetry and appears only with a time lag in drama and prose. But even the third hypothesis of a completely independent development of the genres can hardly be reconciled with the observations based on the data. Rather, we see that poetry and prose, especially after 1890, sometimes have a lot in common, but that drama almost always shows its own development. The fact that drama behaves diferently from prose and poetry has been noted not only by our research group, but also in a study of word choice in 19th and early 20th century English literature1(8[]). Whether these are signs of a more general ’special status of drama’ may be a topic for future research. All in all, the data already allows us to be quite sceptical about the three hypotheses which express assumptions often found in literary studies.
On the other hand, our study has some shortcomings that may call into question the reliability of the results. The relatively small number of plays (only about 350 compared to about 1600 novels and about 5000 poems) may distort the results, especially the low numbers in the ifrst decades, so increasing the number of plays in our corpus will be a high priority. Another problem is the proportion of subgenres, for example comedy in drama or elegy in poetry. At the moment we don’t control our data for this variable. On the other hand, probably a change in the preference of a subgenre is already an indicator for a change in which emotions are preferred. So if we were to keep the proportions of the subgenres equal, we would dampen some of the fluctuations in the area of emotions.
We believe that the issue of making valid comparisons across significantly diferent segment lengths and numbers requires further empirical investigation. In future research, we aim to move away from using segments as a classification unit and instead focus on token-level classification. Previous eforts to train models within this framework have unfortunately been unsuccessful, but we are optimistic that progress can be made by utilizing larger language models and prompting.
Other future work will look into the reasons for the closer similarity between the development of prose and poetry and the more singular development of drama. Is it a question of the medium? Prose and poetry are distributed in print and typically meant to be read in solitude, while plays are performed on a stage? Or does the diference lie in the fact that poetry and prose have an intermediate communicative instance (the narrator or speaker), while drama manages without this instance? So far we have only looked at the relationship between one genre and another, but in the future we want to model the process by including all time series at once. This would allow us to determine the weight of temporary influences that genres have on each other. Finally, this could be a first step towards modeling the influence of external factors, such as ideological or socio-economic developments.
This work was funded by the Deutsche Forschungsgemeinschaft as part of theSPP 2207 Computational Literary Studieisn the project: Literary Change. German Poetry between Realism and Early Modernism and Its Relation to Literary, Cultural and Social Developments.
Leonard Konle: Software, Formal Analysis, Visualization, Writing – original draft Merten Kröncke: Data Curation, Methodology, Writing – original draft Simone Winko: Data Curation, Conceptualization, Supervision, Writing – original draft, Funding acquisition Fotis Jannidis: Conceptualization, Supervision, Writing – original draft, Funding acquisition