In this paper, we describe a plot model for German dime novel romances. Starting with the identi昀椀cation of essential structural parts of their plot based on scholarly analysis of romances, we then formalize this conceptual model. A昀琀er a description of the corpus with its 950 novels, mostly from the last 20 years, and the annotation guidelines for the selected plot elements, we automatically detect these elements using a 昀椀ne-tuned German Bert (with LoRA adapters). While it is clear how to evaluate the performance of the automatic extraction of the plot elements, it is less clear how to evaluate the quality of the model in total. We apply it to two texts and compare the result to summaries of these novels based on reading them, to discuss the strengths and weaknesses of the plot model in making aspects of the plot structure visible. For the quantitative evaluation each novel is represented as a multidimensional time series data. Classi昀椀cations of this data distinguish between publishers, genres and series respectively; we see the performance in this task as an indication of the quality of the model. Finally, applying the model to a larger corpus of romance novels, we detect patterns of the genre. The more modern form of the romance novel, published by the publishing houCseora, is characterized by the importance of the physical perception and the reduction of the plot with the high and exclusive focus on the relationship of the lovers.
The formulaic love story with a happy ending, the romance, was for a long time near the bottom
of the ranks of aesthetic appreciation, a place where crude horror and war stories live near the
abyss which is usually inhabited by pornography. But this has changed in recent years, at least
in North America. The democratization of reading, the emancipation of the popular culture,
the loss of status of the elites, all this has made this elevation of the status of romance possible.
Today you 昀椀nd a genuine interest in the genre, because it is so widely read, it creates so very
intense feelings in so many readers and is comparatively under-researched and still not
wellunderstood. This is shown in books from academics like the pioneering studi2e2s][and [
Romance – in its modern meaning – is a world-wide phenomenon as can be seen by the
success the company Harlequin has in over 100 countries1[
In Germany, the market of dime novel romances is divided between three companiBesa:stei, Cora and Kelter. Bastei and Kelter publish almost exclusively German authors. Some of their most successful series are republications from the 1960s to the 1990s. One especially successful series consists of shortened versions of the over 200 novels of the German writer Hedwig Courths-Mahler, who published most of her novels between 1900 and 194C0o.ra publishes a small set of German authors, but mostly translates and shortens novels from its parent company Harlequin. The three publishers of German romance dime novels produce about 60 romance series, each with a weekly issue, around 3.000 novels per year – though an unknown number of them are reissues.
The romance dime novels provide perfect material for Computational Literary Studies (CLS):
There are many, they are comparatively short with little variance in length, and they
instantiate a genre which is known to be formulaic1[
This paper builds on our earlier work, where we discussed di昀erent approaches to the
modeling of plot and argued for a model which is not generic, but only applicable to a corpus with
shared features 1[
In this paper, we will discuss a plot model for the German dime novel romance, building on the research literature and re昀椀ning it with our knowledge of German dime novel romances. This will allow us to explore a question which has been repeatedly discussed in the research on popular literature in general and on dime novels and romance in particular: How formulaic is this literature? One description which can be found quite o昀琀en claims, that “They’re all the same” [10, p. 127]. But as an early researcher of dime novels remarked: If all these novels really are the same, they would certainly not be read. And he already noted that in order to analyze this kind of literature, it is important to identify what remains the same and what varies in a series or genre 4[]. Our plot model allows us to describe an important part of the variation in the way dime novel romances are told.
We use our model to annotate texts manually. Based on that we annotate 950 texts automatically using a 昀椀ne-tuned BERT model.
One of the main problems of modeling plot is to 昀椀nd an adequate way to evaluate the model. We choose a qualitative and a quantitative approach. For the 昀椀rst, we compare the representation of the plot of two stories with a summary and our close reading. This gives us valuable insight which content aspects are retained and which are lost. For the quantitative approach, we use the automatically annotated data to represent each novel. Clustering and classifying them allows us to evaluate how much information is retained in the plot models.
Finally, we use the plot features over the progression of the model to detect typical patterns. Thus we can show that the novels byCora not only emphasize the physical aspects of the relationship between the lovers, but also concentrate the plot structure on their meetings.
Our corpus consists of 950 romance dime novels from 昀椀ve sub-genres, released by three publishers (see Figure2). We can distinguish between several sub-genres. The most represented is the group of ‘pure’ or ‘generic’ romance novels. In the other romance genres, a romantic couple relationship is mandatory, but less dominant in the plot. The ‘Heimatroman’ is quite speci昀椀c to the German-speaking countries. Its setting are the alpine mountain villages, where old-fashioned values are still kept alive (similar to the children’s book ‘HeMideid’)i.cal novels additionally depict everyday life in doctoral o昀케ces or hospitals, aRnodmantic Suspense novels combine elements of the romance with those of horror and sometimes contain supernatural elements such as ghosts or werewolves.
This selection of series represents a cross-section of curren2tpluyblished dime novels by the dominating publishers. But this also includes reprints of texts 昀椀rst published up to 100 years ago (e.g. in the Courths-Mahler series). Unfortunately, publishers of dime novels usually avoid communicating a novel’s age in any way. Therefore, we exclude any historical analyses for now. The authorship of the texts is also di昀케cult to determine. Several strategies of attribution are used simultaneously: clear names, pseudonyms of one auth3oarn,d pseudonyms owned by publishers 7[].4
In the following we describe a conceptual model of the plot structure of romance dime novels
based on scholarly de昀椀nitions and descriptions of this genre10[
One important structural unit for our understanding of plot isstchenee. We understand a scene as part of the discours of a narrative that represents a part of thheistoire in such a way that both are simultaneous, the location does not change, a particular action is central, and the constellation of characters remains33[]. Some of the events of our plot model are characteristics of one single scene, for example the 昀椀rst contact or the happy ending, while others appear in more than one scene. The plots of popular romances can be understood as a sequence of scenes, and each scene shows either a meeting of the protagonists or not. Using scenes as basic narrative units allows us to create an information structure which ties all plot elements to a plot unit. Additionally, we can add genre speci昀椀c information to this plot unit: do the lovers meet or no5t.The variations between romances can now be described using plot elements likeFirst Contact and the information about the scene structure.
The main goal of the narrative is the emotional engagement of the reader usually by showing the emotions of the protagonists, by depicting the process of falling in love. In romance novels this love is overwhelmingly heteronormativ2e7][. That is also true for the German dime novel. Falling in love is depicted as a process that is accompanied by perceptions of the character and the body of the loved one. These perceptions are shown in more than one scene and can refer to 5Snitow describes the whole plot of Harlequin romances as scenes, in which the lovers are together or in which the female protagonist is ‘essentially’ alon2e6[] but this is too limited to describe all romances we read. any aspect of a personality, but we can distinguish between positive and negative dimensions. Because we observed that the mutual declaration of love was in some interesting cases not identical with the happy ending, we decided to keep them separately.
In the US the term romance is nowadays sometimes understood to refer to a story which
contains one or more erotic encounters between the lovers. But if we look at the genre in total,
we see this is o昀琀en not the case: Historically, the boom of the erotic romance started in the
1970s [
For our annotation, we adapted the conceptual model in the following way: We assume that there is a “main love plot” in the romances. All categories are annotated in relation to this main love plot, not in relation to minor characters or other plot lines. Prior to annotation, the texts were automatically segmented into scenes using LLPr7o.Captured is 1) information related to the unit of the whole scene; 2) information related to smaller units of text, usually phrases. Based on research literature and our reading experience, we assumed that there usually is a heterodiegetic narrator and that the narration is depicting scenes generally in their chronological order.
Figure 4 provides an overview of the annotated information captured in each novel on the level of the scene. It is based on the description of the romance above with some important exclusions and simpli昀椀cations: We did not annotate theobstacle, because there are so many di昀erent ones, especially if one also includes non-contemporary romances.
Because the bulk of German dime novel romances are written for wome7n], [the point of view
is most of the time the female protagonis8t. Therefore, if we want to model the process of
‘falling in love’, it seems to us a justi昀椀ed limitation to focus on the female protagonist. The
process of ‘falling in love’ includes many di昀erent aspects; we operationalized it by focusing
on the female protagonist’s perceptions of, emotions caused by and physical attraction to the
male protagonist. In our approach to the annotation of emotions, we follow the model that
Simone Winko developed for analyzing emotions in poetry and that Claudia Hillebrand adapted
for narrative texts2[
love plot. An example: The female protagonist is happy because she has a date with the male protagonist. Only the presence of a positive emotion is recorded. Under ‘perception’ we capture the perception of the male protagonist by the female protagonist. Here, we distinguish between the perception of character traits and physical appearance, such as body, voice, or smell. Just as with emotions, only the polarity of the perception – not what is perceived – is recorded. Thirdly, by the tag ‘physical attraction’, we capture phrases that express that the female protagonist is physically drawn to the male protagonist. As mentioned above, we do not directly annotate the ‘obstacle’ that stands in the way of the protagonists’ love. However, at least indirectly, the obstacle is likely to be partly re昀氀ected in the categories we annotate. For the protagonist’s negative perceptions and negative emotions may be caused in part by it.
We annotated 15 novels at scene and phrase level in full. Four of those were annotated by all four annotators to calculate inter-annotator agreement (Multi-Annotat1i)o.nSsince each novel contains only one happy ending and one consent scene, we have additionally given annotators the task of identifying happy endings in a larger collection of novels. This results in 137 happy ending and 83 consent scenes. For an overview of the annotated resources see Table1. The inter-annotator agreement for these features were in most cases quite high, with a notable exception: the IAA on character perceptions and inner emotions is comparatively low. The di昀erence is not particularly surprising and can be explained by the higher complexity, both conceptual and linguistic, of emotion and character compared to the perception of physical properties. The disagreement is not in the distinction between positive and negative perceptions or emotions, but in the identi昀椀cation of relevant text passages.
The automatic annotation of the texts is performed either on scenes or on sentences. Each scene is annotated with the information whether there is a meeting between the lovers in this scene or not (and we assume that the 昀椀rst contact can be labeled asFirst Contact). Additionally, we classify scenes as aConsent Scene, that is a mutual declaration of love, orHaappy Ending or as a negative. On the sentence level, the polarity of character perceptions, physical perception, and emotions are attributed; and indicators for physical attraction are detected.
In order to generate automatic annotations for the whole corpus, we utilize machine learning and 昀椀t one model for each tag of the tagset (see section 2.1). Our foundation model is 昀椀ctiongbert-large9 a variant of gbert-large1[] adapted to narrative texts. The 昀椀netuning is performed using LoRA adapters (rank=4, alpha=8)9[]. The use of LoRA, instead of full-昀椀netuning, has the advantage that by reducing the number of weights that must be trained, the amount of computation required is greatly reduced, and the model is more robust against over昀椀tting. Each task is trained over 20 epochs with a linear decreasing learning rate of 0.0001. Since all types of annotations are highly unbalanced (e.g. there are much more phrases containing no annotation), random under-sampling is utilized in every epo1c0h.
10Code and Data: https://github.com/LeKonArD/On-Character-Perceptiona-nd-Plot-Structure-of-German-Rom ance-Novels
Table 3 denotes the performance of the meeting classi昀椀cation compared to a baseline. The baseline simply assumes: If the names of the two main protagonists are present in a scene, there is a meeting. We show that this approach is not su昀케cient.
The identi昀椀cation of happy ending and consent scenes is di昀케cult due to its highly imbalanced data distribution (one happy ending per novel and usually only one consent scene). Training both tasks leads to high recall/low precision results on the testset.
We deal with this problem by using the knowledge that happy endings emerge towards the end of a novel as post-processing over the classi昀椀cation results, so that the last scene classi昀椀ed as a potential happy ending is always the one happy ending. For the consent scene we have no such information, so we simply use the scene with the highest so昀琀max value for determining the consent scene. These steps lead to a mediocre happy ending and consent scene detection. Table4 denotes the performance on complete, held out novels.
Table 5 shows the performance of sentence level tasks on the test set. The F-scores of the categories are in an acceptable range (and feature low variance between splits) except for negative physical perception. The reason for the outlier in this category is not due to factors like linguistic variance, but simply re昀氀ects the lower number of examples (see Tab1l)e. The case is di昀erent for physical attraction, where the value is higher than in the other categories, although there are not signi昀椀cantly more examples. Possible explanations are the restriction to exclusively positive examples or a low linguistic variance of its phrases.
Evaluation of plot models is uncharted territory because research usually tests whether the automatically extracted features align with a manual annotation of these features, but not whether the model itself is an appropriate representation of the plot. Sometimes researchers try to approach this problem by associating spikes in their graphs with plot highlights, for exam6p]l.e [ We take two approaches: a qualitative one in which we compare the results of close reading two stories with their representation in our model and a quantitative one in which we look at a clustering of our collection of romance dime novels based on a representation of our plot model as a multidimensional time series, and by looking at the quality of a classi昀椀cation of the novels represented in this way. It should tell us, how informative and discriminative this representation is.
Even if our model of plot is richer as more generic ones, because it is hand-tailored to a speci昀椀c corpus, it does not retain some information which is considered by readers as an important part of the plot. To better understand this loss, we will discuss below the relationship between our formal model of a romance dime novel plot and a more conventional summary. Let us start with the summaries of two novels: ‘Seduction under palm trees’ a novel from 2008 by Maureen Child and the much older novel ‘Melody of a summer’ by Leni Wüst from 1964.
Verführung unter Palmen (Seduction under palm trees): The multimillionaire Max Striver meets Janine Shaker, a 氀昀orist from California, in the club of theFantasies resort. Contrary to her usual habits, Janine 昀氀irts with Max and spends a passionate night with him – without using contraception. A昀琀er they make love again the next night, Max proposes to her that she plays his wife for the next few weeks in exchange for a large payment, in order to 昀椀nally scare o昀 Max’s ex-wife, Elisabeth, who is stalking him. Because, as he knows, Janine has been robbed of her fortune by her ex-昀椀ancé, and is in need of money, she agrees, but is very disappointed by his behavior. They successfully pretend to be married, but though they are attracted to each other, they are also kept apart by Max mistrust of people in general. He proposes to marry without love, but she refuses because she loves him and does not want to live in a marriage without love. Janine discovers that she is pregnant. When she informs Max about the pregnancy, he does not believe her and accuses her of trying to deceive him. Janine leaves. Le昀琀 alone, a昀琀er a violent argument, Max 昀椀nds the positive pregnancy test, travels a昀琀er her, confessing his love for her, which he has now realized, and proposes to her. She accept2s][.
Melodie eines Sommers (Melody of a summer): The famous singer Antonia Segestus is staying near the German
city Würzburg for a health cure. There she meets her former boyfriend Johannes Witthold. The two became
acquainted during one of Antonia’s engagements years ago and fell in love. Since Johannes had to support his
mother, while Antonia had a great career ahead of her, the two separated and each married someone else. During
the summer they now fell in love again. So when Antonia’s husband, the famous doctor Martin Haßlacher arrives
at her hotel, she tells him that she wants a divorce. Haßlacher is furious and tells her she is insane like her mother.
Antonia is deeply a昀ected by this. When she tells Johannes about the separation, he also decides to separate from
his wife. To do so, however, he has to travel to nearby Würzburg. He promises Antonia to come back to her in
the evening. When Johannes does not arrive in the evening as expected, Antonia thinks he will not part with
his wife a昀琀er all. Still under the impression that the conversation with her husband has le昀琀 on her, Antonia
attempts suicide. However, she is found in time by a housemaid and taken to the hospital, where she lies in a
coma from then on. Now it turns out that Johannes had an accident on his way to see her and therefore could
not arrive in time. At Antonia’s bedside, Johannes and Haßlacher have long conversations: Haßlacher is now
more conciliatory and agrees to the separation. Finally, Antonia awakens from her coma and is able to be with
Johannes [
In the 昀椀gures, we see that Janine and Antonia are not doing so well emotionally over long stretches of the plot. The fact that the female protagonists feel so many negative emotions was an impression that emerged again and again during the qualitative analysis. But what emotions it is here that they feel and what causes them, we again do not see in the 昀椀gures. In Janine’s case, the negative emotions are usually a direct reaction to Max’s behavior: He makes her feel angry, sad, or hurt. Johannes on the other hand is not responsible for Antonia’s negative feelings. Instead, it is her husband Martin Haßlacher and the external circumstances that make her feel depressed. Another major di昀erence between the two novels, that is visible in the 昀椀gures, is the position of the consent scene: in ‘Seduction under palm trees’ it coincides with the happy ending. In ‘Melody of a Summer’, however, the partners already come together in the 昀椀rst third of the novel. A昀琀er that, the story tells of the obstacles that stand in the way of their being together. Which obstacles must be overcome in the two novels we do not see in our annotations at the moment.
All in all the comparison shows that the model retains some of the important aspects of the plot, but it is also obvious that equally important aspects, especially those that refer to the causal link between events or to the motivation, are missing.
For the evaluation of the representational power of the model, we use dimensionality reduction
and classi昀椀cation. We do know from previous research that the categories genre, publisher, and
series are important for the segmentation of the corpus, and that information about all three
categories is part of the text 1[
Figure7, on the right, shows that the publishers are dominating the structure. Especially the
di昀erence between Cora on the one hand andBastei and Kelter on the other hand is obvious;
this 昀椀ts with results from previous studies on the important role of publishers in the 昀椀eld of
11https://github.com/wannesm/dtaidistance
12Parameters min_dist: 0.8, metric: euclidean, n_neighbors: 15
German dime novels 1[
The results are noticeably worse for the classi昀椀cation of the four genres (see tab7l)ewith an F1-score of 0.57 (macro avg), random baseline: 0.25. If we look at the results for the separate classes, we see that the very low recall of the Medical Romances contributes a lot to this result. There are two series in the genreMedical Romance, one from the publisherCora, the other from the publisherBastei. Though they share the same setting they have very di昀erent plot structure as we will see in more detail below.
precision recall f1-score support Medical Romance Heimatroman Romance Romantic suspense accuracy macro avg weighted avg
Lastly there is the classi昀椀cation of the eleven series with an F1-score of 0.46 (macro avg), random baseline: 0.09. Because here the publisher is not a confounding variable, it works rather well.
Table8 and the corresponding confusion matrix in Figu8reshow some interesting patterns which give us an insight into this literary 昀椀eld: The novels from one series of the publisher Cora are easily confounded with the novels from another series by the same publisher. Almost two thirds of the seriesBaccara for example are attributed toJulia and Ärzte zum Verlieben, while almost half of the novels of the last series are classi昀椀ed as being from tBhaeccara series. The same is true, but to a markedly lesser extent, for the series from the publishBearsstei and Kelter: Their identity is, even under this abstract view of a plot structure, more speci昀椀c and less interchangeable – with some interesting exceptions, about a quarter of the general romance seriesSilvia is classi昀椀ed as Medical Romance, even though or because the setting is not part of the classi昀椀cation information. 5.2.2. Discussion To better understand how well our plot model was evaluated by our classi昀椀cation of the plot representations, we compare it to a classi昀椀cation using the whole text. The classi昀椀cation of the romances by publishers, genre or series using a bag of word representation works very well: publisher: 0.97, genre: 0.87 and series: 0.73 (F1-score, macro avg.), which is not surprising because this representation provides the classi昀椀er with information on settings, typical character names, style and a lot more (full results in appendix). It is quite interesting that the plot model with only a few features, but with the information over the progression, is, at least in the case of classi昀椀cation by publisher with 0.83 also quite good.
Figure9 shows how much each of the features of the plot model contributes to the classi椀昀cation: Only physical attraction, which is quite similar to positive physical perception, and happy ending and consent scene which are quite similar in most texts, do not contribute new information. This also supports our assumption that this model is informative.
The advantage of our representation of plot is its ability to show developments over time and perhaps even patterns of development. So in this section, we will show how plot aspects like character perception and meetings develop in the novel progression for the whole corpus because this gives us some new insight into typical patterns. Additionally, we examine in more detail how large di昀erences between the progression patterns of various series disappear in this aggregated view. Finally we look at the longest distance between two meetings of the lovers and the di昀erence of the amount of positive physical perception to describe typical patterns.
The most interesting 昀椀nding of this paper, which, as far as we know, has not been described before, is the fact that in the case of theCora novels, the plot of the erotic novels relates almost exclusively to the relationship between the two main characters. This has structural consequences. Above, we already saw the high values for the positive physical perception for selectedCora romances, which is even more clear, when we look at Figu1r2ewith the average values of this feature. In Figure13 we see box plots for the average numbers of scenes without a meeting of the protagonists. AlClora series show on average very low values. In other words, there are only very few scenes without a meeting of the protagonists in these novels. This means that these Texts are much more focused on the love plot.
It was one of our main goals to build an instrument that allows us to highlight the variation in a genre o昀琀en described as formulaic or uniform. We were able to show that novels by the publisherCora have two quite signi昀椀cant plot features: the high importance of the physical perception, which has been known to scholars, and the high and exclusive focus on the relationship of the lovers, which is new. In this case distant reading was able to show a new and interesting pattern. There are many follow-up questions which could be pursued: How stable is this pattern across di昀erent series, when did it evolve? One could tentatively hypothesize that the Harlequin ‘formula’ works not only by enriching the texts with sexual encounters, but also by concentrating on the meeting of the two protagonists, to the point where even rivals no longer play a major role, but only the emotions of the two lovers.
The list of future work is quite long. One could include information about typical settings into the plot or identify where the narration deviates from the chronology. Though the classi椀昀cation gives us some insight into the degree to which this model is informative and discriminative, it still begs the question, how to test the quality of the plot model more rigorously.
A particularly challenging problem is the modeling of the ‘obstacle’ which stands between the relationship of the two lovers and how they overcome it. Some obstacles like the scheming of a rival, the lack of consent of ambitious parents, or the fear of a closer bond due to bad experiences can be found quite o昀琀en, but the variety of these plot points is rather large.
In this study, we concentrated on plot elements and disregarded the aspect ‘character type’, but it is well-known that in these novels character types abound. Based on previous work on character types, it would be possible to add this to the ontology of the romance.
The reported results on the performance of the automatic annotation models leave room for improvement. Since the tasks trained here individually share some properties at sentence level (except for inner emotions, they are perceptions: the female protagonist is the one perceiving, the male protagonist is the one perceived; polarity is captured for emotions and perception), a model that stores the patterns of these properties across tasks should be superior to our approach. One way to achieve this could be the use of Adapter Fusion19[].
The scene level tasks, on the other hand, certainly bene昀椀t from models28[] that allow to process larger contexts, since e.g. a happy ending is also de昀椀ned by the fact that no further obstacle appears a昀琀erwards. Our models are currently blind to this information, since they classify each scene separately.
Finally, the corpus we used here is the product of a very contingent process. Work on the creation of a corpus, which is balanced with regard to sub-genres, mode of narration, time of publication, publishers etc., is ongoing but far from its completion.
CRediT Roles: Konle: Formal Analysis, Investigation, So昀琀ware, Writing; Hilger: Data Curation, Investigation, Methodology, Writing; Jannidis: Conceptualization, Methodology, Supervision, Writing. Annotators: Jule Beck, Oana Heckl, Agnes Hilger, Lilli-Grace Moutschka
precision recall f1-score support precision recall f1-score support Medical romance Heimatroman Romance Romantic suspense accuracy macro avg weighted avg accuracy macro avg weighted avg precision recall f1-score support Alpengold (B) Baccara (C) Courths-Mahler (B) Das Berghotel (B) Der Bergdoktor (B) Dr. Stefan Frank (B) Irrlicht (K) Julia (C) Silvia (B) Ti昀any (C) Ärzte zum Verlieben (C) accuracy macro avg weighted avg