Daejeon, July 27–31
There is a long tradition of theological research within Digital Humanities (DH), reaching back to Roberto Busa’s Index Thomisticus, one of the foundational projects of the field (Jones 2018). More recently, this strand of research has been consolidated under umbrella terms such as computational theology and digital theology (Nunn/Van Oorschot 2024). Drawing on a literature review and the references compiled in the bibliography, we present results from an ongoing DH-based reappraisal of the sexual abuse in the Catholic Church. While public debate and scholarly research have largely focused on the abuse of children, the experiences of adult women and adolescent girls have been systematically marginalized (Hübel 2020; Gott 2022; Leimgruber 2026). This neglect can be understood as a form of epistemic injustice (Fricker 2007; Hürten 2025). Using AI-supported methods, we present results of a project investigating the German discourse on the sexual abuse of women in media coverage and official Catholic documentation. 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/
Alongside media coverage, social media content, survivor 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.
Table 1. General corpus statistics of five abuse reports.
The entire corpus consists of 2,362 PDF pages and almost 730,000 tokens (after removing noise) with the longest report having 915 pages. We aim to develop AI systems to first identify relevant sections. Relevance is determined as sections that deal with discourse and reporting of the sexual abuse of specifically adult women and/or girls above the age of 14. In these sections, we want to perform further fine-grained classification according to a complex annotation scheme. An expert annotator manually identified all relevant sections in the PDFs, resulting in approximately 6% of pages being marked as relevant, reduced to just the relevant content making 3.6% of the overall text. These annotations indicate that the abuse of adult women and adolescent girls is rarely discussed in the reports.
We employ zero-shot relevance detection using an instruction-tuned LLM (Gemma 3:12b) (Gemma Team et al. 2025). Zero-shot detection refers to the LLM method in which a model assigns classes (here: relevance) to a text unit without having been explicitly trained on labeled examples. This approach was chosen due to the large class imbalance and the limited availability of labeled data. We prioritize high-recall detection because hidden patterns in the texts make it likely that relevant passages are missed even in expert annotations, and relying solely on pre-labeled training data could risk reproducing the very epistemic injustice under investigation. Zero-shot classification has proven particularly effective for rare, implicitly articulated, and semantically complex categories, as it leverages contextual meaning rather than surface-level lexical features (Zhang et al. 2019; Schopf et al. 2023). Our pipeline follows a human-in-the-loop approach: expert-labeled positive examples are analyzed to better understand hidden patterns, identify missed relevant sections, and evaluate the limitations and benefits of generative models. For evaluation, all texts were split into chunks of 10,000 characters, the maximum size of all relevant sections, and chunks containing relevant content were labeled accordingly (Table 1). Through experimentation, we found that a simplified and strict prompt, marking passages as relevant only when certain, produced the best results. Figure 1 presents an English translation of the original German prompt. Two additional prompts were used to collect the model’s explanations and reasoning for further in-depth analysis. Tables 2 and 3 summarize the key evaluation metrics.
Figure 1. Relevance detection prompt (translated from German).
Table 2. Relevance prediction cross table for the LLM and the gold standard (GS) annotation on all chunks.
Table 3. Evaluation metrics overall (accuracy, F1) and specifically for the positive “relevant” class (precision, recall).
The evaluation results indicate a tendency for the model to overclassify chunks as relevant, resulting in low precision but higher recall for the positive minority class. Analysis of false positives and the model’s reasoning yielded several insights:
Overall, the vague and unclear language of the reports about age and situation of survivors, itself a manifestation of hidden patterns, poses challenges to the model. However, we did indeed find relevant texts in the false positives showing possible benefits. We intend to use the model as a high-recall filter rather than a definitive classifier, prioritizing sensitivity over precision to minimize missed cases and learn more about the discourse.
After the identification of relevant passages, we intend to perform more detailed fine-grained analysis of the texts. For this purpose, we developed a complex hierarchical annotation scheme in an iterative process (Reiter 2020). The scheme consists of 16 main categories and 122 sub-categories covering all relevant research aspects of the topic, including information about survivors and accused persons, the actions of Church authorities, and the framing of the crime (see figure 2 for an example annotation). We annotated all sections marked as relevant by the expert annotator according to the scheme resulting in 1,658 annotations of varied lengths. For annotation, we used the CATMA tool (Gius et al. 2025) which is an established tool used in DH research for the flexible annotation of large form texts and shows wide usage in similar projects (Dennerlein et al. 2023; Heyns and van Zaanen, 2024; Guhr et al. 2025).
Figure 2. Annotation example. Figure 2 provides insight into the annotation practice (Bignasca et al. 2023, 81). Annotated elements include: in blue the accused Hansjörg Vogel (gender: male, status: cleric), in light green the victim (gender: female, age: adult), in brown the description of the crimes (here: consequences of pregnancy, annotated as description of the crimes_reproductive), in yellow sexual acts including penetration (must have happened previously), in dark green the bishop’s reaction in the form of his resignation, and in pink the year of the reaction. Please note that the original annotation was done for German text and this is an English translation.
After narrowing the texts to relevant passages, class imbalance and the lack of labeled data are less severe for individual annotation classes. Accordingly, we treat discriminative transformer-based language models as state-of-the-art for classification tasks. Nevertheless, an imbalance persists, as most of the text within relevant passages does not correspond to any annotation category. We formulate the task as binary classification between a given category and non-annotated text, splitting both annotations and relevant text into sentences and assigning the corresponding labels (resulting in 1,422 non-annotated sentences). For evaluation, we tested the largest German-language BERT models, including gbert-large (Chan et al. 2020) and ModernGBERT (Wunderle et al. 2025). To address class imbalance, we experimented with strategies such as class weighting, oversampling, back-translation, and paraphrasing (Taskiran et al. 2025). Without such optimization, the models consistently collapsed toward the negative majority class, in contrast to the generative LLMs discussed in section 3. Table 4 presents the results of the best-performing models in a 5×5 evaluation for the five most frequently annotated main categories.
Table 4. Classification results.
PR-AUC values above 0.7 indicate good integration of the minority class, and balanced accuracy values of around 0.8 are encouraging. Our evaluations showed various insights and best practices for small data settings. The best configuration with the model bert-base-german-cased includes the application of class weights and oversampling through back-translation (German à English à German). Further model information: Fine-tuning for 4 epochs, learning rate of 4e-5, Adam optimizer, trained on a A100 GPU. Back-translation was performed with sentencepiece (https://github.com/google/sentencepiece)
This study marks important components of a larger effort to uncover the marginalized voices in German discourse on sexual abuse in the Catholic Church. All annotations, data and results are publicly available on a GitHub repository. We intend to develop and optimize pipelines like the one presented for other text types such as media texts and survivor testimonies. With the limited set of texts, we have already gained insights through annotation scheme development and annotation analysis. Our machine learning results are centered around the challenge of small data sets and class imbalance. The high-recall relevance detection using generative LLMs is encouraging, and we intend to explore the models further prioritizing human judgement before AI in a human-in-the-loop approach, with additional annotations as well as advanced few-shot-learning (giving the model a small set of labeled examples to improve classification). The fine-grained classification model shows promising results, and we look forward to applying optimized models on large unlabeled data to further our advances in turning “silenced data” into analytically and socially meaningful knowledge.
Sexual abuse has been systematically covered up in the Catholic Church, and the abuse of women continues to be formally recognized only under narrow conditions. Uncovering “hidden knowledge” about such cases therefore is crucial to gain insight into the phenomenon, develop preventive measures and promote (epistemic) justice for the survivors. To our knowledge, there is currently little comparable computational work focused specifically on systematically identifying and analyzing marginalized survivor perspectives within the Catholic Church abuse discourse, highlighting the methodological and exploratory contribution of this project to both DH and computational social analysis. By demonstrating how computational methods can surface epistemic injustice embedded in institutional documentation, this work contributes to broader DH discussions about the ethical dimensions of data interpretation and the role of digital methods in amplifying marginalized voices.