Our team focused on Subtask 2 (narrative classification) and tried several conceptually straightforward approaches: (1) prompt engineering of LLMs, (2) a zero-shot approach based on sentence similarities, (3) direct classification of fine-grained labels using SetFit, (4) fine-tuning encoder models on fine-grained labels, and (5) hierarchical classification using encoder models with two different classification heads. We list results for all systems on the development set, which show that the best approach was to fine-tune a pre-trained multilingual model, XLM-RoBERTa, with two additional linear layers and a softmax as classification head.

We propose a framework for quantitative-qualitative research in corpus-assisted discourse studies (CADS), which operationalises the central process of manually forming groups of related words and phrases in terms of “discoursemes” and their constellations. We introduce an open-source implementation of this framework in the form of a REST API based on Corpus Workbench. Going through the workflow of a collocation analysis for fleeing and related terms in the German Federal Parliament, the paper gives details about the underlying algorithms, with available parameters and further possible choices. We also address multi-word units (which are often disregarded by CADS tools), a semantic map visualisation of collocations, and how to compute assocations between discoursemes.

We use query results from manually designed corpus queries for fine-tuning an LLM to identify argumentative fragments as a text mining task. The resulting model outperforms both an LLM fine-tuned on a relatively large manually annotated gold standard of tweets as well as a rule-based approach. This proof-of-concept study demonstrates the usefulness of corpus queries to generate training data for complex text categorisation tasks, especially if the targeted category has low prevalence (so that a manually annotated gold standard contains only a small number of positive examples).

We are concerned with mapping the discursive landscape of conspiracy narratives surrounding the COVID-19 pandemic. In the present study, we analyse a corpus of more than 1,000 German Telegram posts tagged with 14 fine-grained conspiracy narrative labels by three independent annotators. Since emerging narratives on social media are short-lived and notoriously hard to track, we experiment with different state-of-the-art approaches to few-shot and zero-shot text classification. We report performance in terms of ROC-AUC and in terms of optimal F1, and compare fine-tuned methods with off-the-shelf approaches and human performance.