Computational Literary Studies of Fictional Space and Affect

Berenike Herrmann, Daniel Kababgi, Robin Martin Aust, Marie-Christine Boucher

Since the spatial and the emotional/affective turns in literary studies, computational methods have enabled several impressive leaps. Theoretical frameworks of emotion and space, such as the dimensional and discrete models of emotion (Ekman, Plutchik, Russel) and the structuralist theory of space advanced by Yuri Lotman have been operationalized for computational text analysis, enabling novel insights and new scales of analyses. So far, only few approaches have combined the affective and spatial analysis of literary texts, among which is our own work. Building upon this, our project aims to contribute to spatial and affective literary theory, asking about the relation between the spatial dimension of the diegesis and fictionally encoded emotions and sentiment using a mixed-methods approach applied to German-written literary texts within the “DACH” context. One milestone of our project is the creation of literary (prose) corpora in German, with a focus on Austrian, German, and Swiss authors (born between 1750 and 1900), with potential other literatures for comparison (e.g, English and French). For creating the German-language corpus, different existing resources will be used (ELTeC, Kolimo+, TextGrid, DTA, Project Gutenberg) and their metadata will be enriched with new information pertaining to hypothesis-driven research questions. The main research question is about diachronic change in the use and meaning of spatial entities and their affective encoding across ‘national literatures’ as well as genres, subgenres and potentially topics (e.g., novella, novel, ‘Dorfroman’, architecture, pauperism).

  • Do we observe differences between named (Zurich, Krems, Mount Blanc) and unnamed (church, mountain peak, meadow) spatial entities?
  • Do we observe differences between fictive entities and entities based on factual entities?
  • Do we observe differences between fiction and non-fiction?
  • How can we capture the “style of the sublime”?

To gain these new insights, different deep learning models using powerful transformer-based language models like BERT will be trained using an active learning approach. Specifically, we will train /finetune a model that can identify spatial entities as well as a model that classifies discrete emotions and valence and arousal, using existing models. As such, not only is this a study of different national literatures, but it is also a study of the opportunities and challenges of active learning and “traditional” supervised machine learning. To train these models, spatial literary theory will be operationalized( Lotman, Dennerlein, Schumacher).

To summarize, the contributions of this project to the wider area of CLS and spatial literary research shall be (1) the creation of a new corpus resource specifically focused on comparative research of Austrian, German, and Swiss texts, (2) several new machine learning models that can be used and further finetuned for new projects, (3) a methodological study on the benefits and pitfalls of active learning in conjunction with deep learning architecture, and (4) new insights into the differences and similarities of Austrian, German, and Swiss national literature. The team consists of two postdocs (Aust, Boucher) and a pre-doc (Kababgi) as well as the PI (Herrmann). Funding has been acquired from other sources.