Genre-blending in Liebesromanen

Genre-blending in Liebesromanen Dr. Svenja Guhr Bellwether Postdoc at iSchool, UC Berkeley “Digitale Gattungshermeneutik II: Gattungen im Kontext” Uni Rostock, Institut für Germanistik 27. und 28. November 2025 Warum Liebesromane? Warum Liebesromane? •Genreerwartungen der Leser:innen •Emotionale und psychologische Funktionen •Identifikation: Leser:innen erkennen sich in Figuren und Lebenssituationen •Bestätigung weiblicher Werte und Wünsche •Weibliche Handlungsmacht: Heldinnen handeln aktiv und überwinden Hindernisse •Helden zeigen Gefühle, übernehmen Fürsorge •Leseerlebnis: positive Emotionen und Happy End Ramsdell (2012): Romance Fiction: A Guide to the Genre. Harald (2006): Genreflecting: A Guide To Popular Reading Interests Regis (2003): A Natural History of the Romance Novel. Radway (1991): Reading the Romance : Women, Patriarchy, and Popular Literature. Typischer Plot: Regis (2003): A Natural History of the Romance Novel. Harlequin FAQ Q. Was ist der Unterschied zwischen Serienoder Kategorienromanen und anderen Liebesromanen? A. Die Romane der Harlequin-Reihe Romane sind kürzer […] typischerweise 50.000 bis 75.000 Wörter lang. Serienromane sind von Tropen getrieben, und jede Serie liefert ein bestimmtes Leseversprechen, das ein Happy End beinhaltet. Die Romantik steht in diesen Geschichten im Mittelpunkt und nicht nur die Entwicklung einer Figur. Jede Serie hat ihre eigenen erforderlichen Story-Elemente, Sinnlichkeitsgrad und Seitenanzahl. Korpus: Men Made in America •Harlequin-Reihe US-amerikanischer Liebesromane •Geschrieben von 40 Autorinnen •50 Staaten –50 Romane (lokale Schauplätze) •Veröffentlicht zwischen 1982 und 2002 •Autor:innenrichtlinien für Harlequin-Romane 7 Thanks to Hannah Walser who purchased this corpus on a flee market in 2015! Harlequin Submission Guidelines: https://harlequin.submittable.com/submit (The Internet Archive Way Back Machine, Oct. 2015) Genre-blending in the Romance Romance Fiction – Pure Genre? → Romantic Plot + X? Monaco/Algee-Hewitt (2025): “Castle at the Crossroads: A Machine Learning Approach to Generic Mixture in the Nineteenth-Century Gothic Novel”, DH2025, Lissabon. Konzentration auf Szenenwechselpositionen He tipped his hat to her. "Be seeing you." "Goodbye," she murmured through stiff lips. She saw him to the door, managing a frozen smile as she closed the door behind him. Then she put her head in her hands and told herself she wouldn't cry. She wouldn't. Not over Michael Wade. </scene> <scene n=“7”> The doorbell rang. Bailey stiffened, wondering if Gunner might have quickly decided to snatch back one or both of his offers. She pulled the door open again, looking out cautiously. Chili Haskins stood on her porch, his white, bushy mustache like icicles above his lips. "Howdy, Bailey." (Leonard Cowboy Be Mine) Genreklassifikation mit GPT-4o Bamman/Chang/Lucy/Zhou (2024): On Classification with Large Language Models in Cultural Analytics, CHR2024, Paris. •Satzweise Klassifikation ist zu wenig Kontext •Jeweils 5 zufällig ausgewählte 500-Wort-Passagen von 150 Genre-gelabelten Romanen aus Projekt Gutenberg •Relationship between number of shots, Chain-of-Thought reasoning, and accuracy for GPT-4o Close-Reading Annotation Experiment •Aufgabe: Genre-Annotation auf Szenenebene, um Subgenres zu unterscheiden •5 Annotator:innen (LitWis/CompLing/Informatik) •10 ausgewählte Szenen: 5 “domestic space” und 5 “other” •Nur Szenen, keine Nicht-Szenen •Szenen unterschiedlicher Länge in unterschiedlicher Reihenfolge präsentiert) Beispielauszug aus Liebesromanszene "I can't believe any of this is happening." Luke heard the edge of hysteria in Abby's voice and wished he could hold her, assure her that everything would be fine. But, of course, not only could he not hold her, but he had little hope to offer her. He struggled once again, attempting to free himself from the ropes, but there was no give. The room was quickly filling with smoke, a thick black smoke that burned the back of his throat and stung his eyes. "Dammit," he repeated helplessly. "Help. Help us," Abby screamed, but her scream ended with another harsh round of coughing. Flames leapt out of the wastebasket and licked the edge of the sofa, blackening the fabric and creating more thick smoke. Abby coughed, a wrenching sound that tore at Luke's heart. "Abby, I'm so sorry. I should have realized. I knew there was something not right, that Rusty had something up his sleeve. I should have done something. I should have sensed the danger.” (Cassidy Midnight Wishes) Set Up 1) Priming: •“The short passages below are scenes taken from 20th-century American novels. These scenes were identified and extracted according to the sceneannotation guidelines in Gius et al. (2021).” •“Think about different novel genres you know and list the genres that come to your mind:” 2) Annotation Task: •“Given the following novel extracts, please read each scene carefully. a) Decide on a genre for each scene (there is no single ‘right’ answer; choose the genre that seems most appropriate to you). b) Highlight the words, phrases, or sentences that influenced your decision c) Write one sentence explaining your choice in the “Justification” column. Also indicate whether there is another genre that could plausibly fit the scene. 3) Reflection: •“After completing the annotations, think about novel genres again. Was it easy to decide on a genre? Please add any additional genres that came to your mind:” Ergebnisse Annotationsevaluation (Krippendorff's Alpha) •IAA: α ≈ 0.31 (geringe Übereinstimmung zwischen 5 Annotator:innen) → Goldstandarderstellung Freitextkommentare der Annotator:innen •A1: “adventure, action, crime, fantasy, mystery; Some scenes had keywords that I would add to particular genres like guns, fire, explosions to crime or action or adventure genres, but then there were love scene-elements as well in all of them which makes it difficult to grasp.” •A2: “I completed annotations of three passages. The first two I felt were relatively easy to categorize, but the third was not. When I was reading the third passage, it made me realize that there are many books I know that I don’t know the genre of. I thought of westerns as a potential other genre but I’m not sure if that’s appropriate.” •A3: “I felt like after reading one that I felt confident was romance, everything after that also seemed like romance. I also feel like the genres I chose were coarse enough where I didn’t feel like adding any more.” •A4: “yes, mostly (either it was very clear or I had no idea – not much in between). additional genres: fantasy” I Asked Gemini 3 (local UCB model) Annotationsevaluation (Krippendorff's Alpha) •IAA: α ≈ 0.31 (geringe Übereinstimmung zwischen 5 Annotator:innen) •IAA: α ≈ 0.62 (Übereinstimmung zwischen Goldstandard und GEMINI) ("gold","9780373360055_058","action"), ("gold","9780373360086_077","romance"), ("gold","9780373360147_001","thriller"), ("gold","9780373360147_062","romance"), ("gold","9780373360246_074","thriller"), ("gold","9780373360383_025","action"), ("gold","9780373360406_027","romance"), ("gold","9780373360468_042","thriller"), ("gold","9780373360468_052","literary fiction"), ("gold","9780373360536_094","crime"), ("Gemini", "9780373360055_058", "action"), ("Gemini", "9780373360086_077", "romance"), ("Gemini", "9780373360147_001", "literary fiction"), ("Gemini", "9780373360147_062", "romance"), ("Gemini", "9780373360246_074", "thriller"), ("Gemini", "9780373360383_025", "action"), ("Gemini", "9780373360406_027", "romance"), ("Gemini", "9780373360468_042", "western"), ("Gemini", "9780373360468_052", "literary fiction"), ("Gemini", "9780373360536_094", "thriller") Fazit und Ausblick •Szenen sind eine homogenere Textabschnittsgröße: groß genug für umfassend Kontext, in sich geschlossen, klein genug fürs Training oder direkter Verwendung mit LLMs •Geminis Klassifikationsergebnis entspricht oft den Kompromissentscheidungen der menschl. Annotator:innen •Wir wissen nur nicht, worauf die Klassifikationsentscheidungen basieren. •Offene Fragen: •Wie kommt man an ausreichend qualitativ-hochwertige Trainingsdaten? •Manuelle Annotation von Szenen ist zu zeitaufwendig (ca. 12 Szenen/h) •Wie können wir die Klassifikationsaufgabe validieren? •Was könnten nächste Schritte sein?