The Flow of Emotions. Intratextual Emotion Progression and German-Language Poetry, c. 1850-1910

1. Introduction

For most literary phenomena, it is not only relevant how often they appear in a text, but also in what order. A clear example is the representation of emotions: A text that first represents happiness and then turns to sadness conveys a different emotionality than a text that first represents sadness and then turns to happiness.

Our study analyzes the progression of emotions within literary texts. We examine a large corpus of German-language poetry from about 1850 to 1910. By focusing on the order in which emotions are represented, we complement previous work by our research group, which has already addressed the general frequency of emotions in poetry, but not their intratextual progression (Konle et al. 2022). On a general level, our contribution provides first fundamental insights into the flow of emotions in German poetry. In addition, we test a specific hypothesis, namely that modernist poetry (in our case, c. 1885-1910) is less likely than realist poetry (c. 1850-1884) to move from negative to positive emotions and more likely to move from positive to negative emotions at the intratextual level. This hypothesized tendency away from positive turnarounds aligns with similar claims in traditional literary studies (e.g., Friedrich 1956: 173), and is inspired by previous research indicating that modernist poetry represents less positive emotions in general (e.g., Konle et al. 2022). We will show that, contrary to the hypothesis, there is no substantial difference between realist and modernist poems in the progression of (positive or negative) emotions. Instead, we find another tendency common to both periods, namely a general increase in the frequency of emotion toward the end of the poems.

2. Related Work

Sentiment analysis as a methodological approach to literary texts has been widespread in the digital humanities since Jockers (2014) at the latest. Rebora (2023) provides a detailed overview on how the field developed since. As far as the analysis of sentiment in poetry is concerned, the work of Haider et al. (2020) should be mentioned for the German language; in contrast to our work, the focus here is on the reader's emotion, not the emotion depicted on the surface of the text. Similar efforts have also been made in other languages, e.g. for Arabic (Alsarif et al. 2013), English (Haider et al. 2020, Ahmad et al. 2020) or Latin (Sprugnoli et al. 2023).

3. Resources

Our corpus consists of 6619 poems from 20 anthologies of contemporary German-language poetry published between 1859 and 1911. 8 anthologies represent the poetry of realism (c. 1850-1885), 12 anthologies the poetry of early modernism (c. 1885-1910).

We manually annotated 1521 poems from our corpus for emotion. The goal was to annotate the emotions represented in the text itself, but not readers’ emotions. The annotators used a list of 6 discrete emotions (agitation, anger, fear, love, joy, sadness), inspired by the emotion hierarchy in (Shaver et al. 1987). It was possible to assign exactly one, but also none or multiple emotions to individual lines of a poem. First, each poem was annotated independently by two annotators, then they merged annotations manually into a consensus annotation. The inter-annotator agreement, measured with γ (Mahet et al. 2015), was 0.7491, which is quite solid for emotion annotation.

We performed an emotion classification task, which we modeled as a series of binary classifications. The basis of our classification is the German BERT (Devlin et al. 2018) model gbert-large (Chan et al. 2020), which we adapted to our target domain via continued pre-training. Subsequently, we performed fine-tuning on the binary emotion classification tasks. Overall, the classification performance shows an acceptable level of uncertainty. ¹ More detailed information on annotation and emotion detection can be found in previous papers from our working group (Konle et al. 2022; annotation guidelines: Kröncke et al. 2022a, Kröncke et al. 2022b). In the following analyses, we use the manual annotations and the results of the automated classification task for all remaining texts.

4. Analysis

To give a first overview of the intratextual progression of emotions, we distinguish exclusively between positive and negative emotions (love + joy = positive, anger + fear + sadness = negative, agitation is discarded). We divide each poem into 10 parts and measure which parts contain positive and/or negative emotions. ² Finally, we combine the values from all poems (Fig. 1).

Fig. 1: Progression of positive and negative emotions. Example: Of all the first deciles of the poems, 37% (realism) or 25% (modernism) contain at least one positive emotion, while 18% (realism) or 21% (modernism) contain at least one negative emotion. ‘Combined Sentiment’ equals ‘Positive Sentiment’ minus ‘Negative Sentiment’. The transparent area behind each line shows the 90% confidence interval.

Throughout the text, realist poems tend to represent positive emotions more often and negative emotions less often than modernist poems. As a result, combined sentiment is consistently higher for realism. This outcome was predictable, as previous studies have shown that modernist poems typically represent fewer positive emotions in general (Konle et al. 2022). A new and important finding, however, is that realist and modernist poems have fundamental similarities (rather than differences) in their emotional progression. In particular, it is true for both periods that positive and negative emotions tend to occur more frequently at the end of the text than at the beginning. Loosely speaking, the texts become "more emotional" toward the end (p < .001). ³ There are also some differences. In particular, the relative frequency of negative emotions is almost the same in realism and modernism at the beginning of the texts, but it is substantially higher in modernism at the end. The general increase in negative emotions at the intratextual level is therefore even more pronounced in modernism.

To get a more nuanced view, we now distinguish the progression of emotion into several types (from positive to negative, from negative to positive, etc.). For readability, we divide each poem into 3 (instead of 10) parts and note for each part whether it contains mostly positive ( +), mostly negative ( ‒), or no emotions at all ( x). Finally, we assign each poem to exactly one of the resulting "flow types" and measure the relative frequency of each flow type (Fig. 2). ⁴

Fig. 2: Relative frequencies of flow types. Example: The flow type ' +++' (positive emotions dominate in the first, second, and third thirds of the text) characterizes 15% of realist poems and 9% of modernist poems.

Before analyzing the results, we provide a more summary view of the 27 flow types. To do this, we combine similar flow types into more comprehensive “superflow types” (Tab. 1). The frequency of each superflow type is the sum of the frequencies of the corresponding flow types (Fig. 3).

Fig. 3: Relative frequencies of superflow types. Example: The superflow type 'neg → pos' (roughly, texts that begin with negative emotions and end with positive emotions) characterizes 12% of realist poems and 10% of modernist poems.

Figures 2 and 3 show that realist poems are more likely to represent consistently positive emotions and less likely to represent consistently negative or consistently no emotions than modernist poems. ⁵ As indicated, this was to be expected in light of previous studies. What is surprising, however, is that there is no significant difference between realism and modernism with respect to the superflow types 'neg → pos' and 'pos → neg'. ⁶ The hypothesis proposed in the introduction states that modernist poetry is less likely to move from negative to positive emotions and more likely to move from positive to negative emotions than realist poetry, but our results do not support this contention.

Some scholars have made more specific claims about the progression of emotions in poetry. Hugo Friedrich, for example, focuses on poems that begin with fear and suggests that modernist poems are less likely to represent the overcoming of fear than poems from earlier periods (Friedrich 1956: 173). We aim to shed light on this thesis by taking another look at (super)flow types (table 2). We limit ourselves to texts that begin with fear and distinguish two superflow types, one that represents the overcoming of fear, i.e., the turn from fear to positive emotions (‘Relief’), and one that does not (‘no Relief’).

Fig. 4: Relative frequencies of superflow types (fear). Example: The superflow type 'no Relief' (roughly, texts that begin with fear and that do not end with positive emotions) characterizes 77% of realist poems and 80% of modernist poems, considering only poems that begin with fear.

In modernism, the superflow type ‘no Relief’ is much more frequent than the superflow type ‘Relief’ which means that modernist poems that begin with fear are not very likely to end with positive emotions. However, this is also true of realist poems. Even though the superflow type ‘no Relief’ is slightly more common in modernism than in realism, the difference is very small and not statistically significant. ⁷ Therefore, our results do not support Friedrich's claim. One possible reason for the disagreement is that Friedrich examines a substantially different corpus, i.e., modernist poems up to the mid-20th century by highly canonical and not necessarily German-speaking authors. We tested whether the frequency of the superflow types ‘Relief’ and ‘no Relief’ changes meaningfully when we consider only highly canonical authors from our (modernist) corpus (Rilke, Hofmannsthal, George, etc.), but it does not.

5. Discussion

The scholarly hypothesis that there is a substantial difference between realist and modernist poems in terms of emotion progression could not be verified. As noted, this finding may change when other languages, time periods, and perhaps authors are examined, although, at least in our corpus, the results were not fundamentally different when we considered only canonical authors. 

On the other hand, our results show a distinct pattern common to both periods, namely an increase in the frequency of emotions toward the end of the poems. It is unclear how general this finding is in relation to time and language. In the context of German literature one could form the hypothesis that this is a specific rhetorical structure with the intention to emotionalize the reader and thus we would assume that poetry starts to show this structure after 1750. It is much more difficult to set a date for the end of this hypothetical long-term structure and it probably depends on the underlying notion of poetry. A corpus based on a concept of poetry which includes what is nowadays called lyrics may still show the same trends while a corpus based only on poetry printed in books and journals is probably showing a different structure since postmodernism. 

As poetry is a very peculiar text type, the model we trained cannot be used to annotate emotions in other text types even of the same period as experiments with prose and plays showed without domain adaptation (Kröncke et al. 2024). 

In the future we plan to examine whether the progression of emotions is related to factors other than literary period, such as genre or author gender.