Daejeon, July 27–31
In this poster contribution, we present an overview of an ongoing Digital Humanities (DH) research project The project “Von Epistemic Injustice zu Epistemic Awareness. Prozessbasierte Methodenforschung zu Missbrauch an erwachsenen Frauen in der katholischen Kirche” (From Epistemic Injustice to Epistemic Awareness. Process‑based Methodological Research on Abuse of Adult Women in the Catholic Church) is funded by the DFG (German Research Association, project number 539293243): https://missbrauchsmuster.de/forschen/projekte/forschungsprojekt-von-epistemic-injustice-zu-epistemic-awareness-prozessbasierte-methodenforschung-zu-missbrauch-an-erwachsenen-frauen-in-der-katholischen-kirche/
We present selected results from each component to demonstrate the project’s potential and to outline its current direction through illustrative research findings.
In a first pilot study with five abuse reports (2,363 PDF pages) commissioned by different dioceses and bishops conferences in German speaking regions and published between 2017 and 2025, Most of the German dioceses have by now commissioned independent researchers or law firms to investigate abuse cases within the diocese. The research design and scope of the reports vary greatly. Yet, the majority of reports focusses on the abuse of minors by clerics examining the period from after the Second World war to the present and drawing primarily on archival material provided by the dioceses and, in some cases, supplemented by interviews with survivors, church officials and other contemporaries. The reports are generally made available to the public online as is the case with the five reports that have been analyzed for the pilot study. All data were anonymized by the editors prior to the publication in accordance with European data protection regulations. The term survivor is widely preferred in international abuse research because it avoids positioning individuals in a passive or objectified role—a connotation that the term victim can inadvertently reinforce.
Analysis of the annotations has produced initial insights into the silencing of women’s experiences in the reports. It reveals that even reports focused on the abuse of minors document abuse against adults. However, these cases scarcely register in the public reception of the reports (see also Leimgruber 2026, 96f.). Moreover, the case-based analysis shows that several perpetrators abused victims from more than one age group. Because each case is defined by a single perpetrator, individual cases may include multiple victims of different ages. In seven of the cases that were annotated as relevant, the ages of all victims could be determined, and each of these involved both minor and adult victims. Two additional cases likewise included both minors and adults but also involved further victims whose ages could not be established (Figure 1). This overlap of survivor age groups refutes the assumption of predominantly pedocriminal offenders within the Catholic Church This assumption was already refuted by Dreßing et al. (2018, 128), who demonstrated that pedophilic preference disorders could only be assumed for a subset of offenders within the Catholic Church.
Figure 1. Age group combinations per relevant case
Only 3.4% of the original texts were manually marked as relevant and annotated (highlighting the neglect of this survivor group in the reports). This massive class imbalance and lack of labeled data pose a challenge to all machine learning tasks. In a first step, we aim to design a high-recall relevance detection system using generative AI (Zhang et al. 2019; Schopf et al. 2023). In experiments with an instruction-tuned LLM (Gemma 3:12b) (Gemma Team et al. 2025) we achieve macro F1-scores of 0.6 with high recall and low precision for the positive class. Due to the sensitive nature of the material concerning sexual abuse in the Catholic Church, all LLM-based analyses were conducted on local university server infrastructure. This work is carried out either by experts with long-standing experience in researching abuse in the Catholic Church or by trained student assistants working in close collaboration with them. New team members receive guidance in self‑care practices, and emotionally demanding material is regularly addressed in collegial debriefings. Further model information: German BERT model with fine-tuning for 4 epochs, learning rate of 4e-5, Adam optimizer, trained on a A100 GPU. Optimization via class weights and oversampling through back-translation.
Another important text type is public media reports on the web. Using the Google Search API https://developers.google.com/custom-search/v1/overview?hl=de The tool is based on Open Search and PostgresSQL.
Figure 3. Screenshot of annotation tool. Annotators can perform advanced searches over scraped corpora and annotate relevance information.
To illustrate the potential of web-based research, we present results from a case study of Gottes-Suche.de (“The search for God”), a website that manually aggregates reports on sexual abuse in the Catholic Church. We scraped all short reports available on the platform (approximately 10,000 texts) and conducted a semantic analysis. Figure 4 shows that references to women (and semantically related terms) increase over time, with a notable peak in 2019, indicating a growing awareness of this survivor group within media discourse.
Figure 4: Percentage of the usage of words like women throughout the collected media reports on Gottes-Suche.de from 2010 to 2024.
These example results from Gottes-Suche.de demonstrate the potential of large-scale analysis once sufficiently robust AI systems are in place. At the same time, given the ethical sensitivity of the research domain, it is essential to prioritize human expert judgment. We therefore continue to pursue a human-in-the-loop approach in which AI systems are iteratively improved through continuous reflection on and validation of their outputs. All data, annotations, analyses and more information about the AI application are made publicly available via a GitHub repository to support transparency and further research in this field.