Zero-shot prompt-based classification: topic labeling in times of foundation models in German Tweets

Filtering and annotating textual data are routine tasks in many areas, including social media and news analytics.Automating these tasks enables scaling analyses with respect to speed and breadth while reducing manual effort.Recent advancements in Natural Language Processing, particularly the success of large foundation models, provide new tools for automating annotation processes through text-to-text interfaces with written guidelines, eliminating the need for training samples.This work assesses these advancements in a real-world setting by empirically testing them on German Twitter data about social and political European crises.We compare promptbased results with human annotations and established classification approaches, including Naive Bayes and BERT-based fine-tuning with domain adaptation.Despite hardware limitations during model selection, our prompt-based approach achieves comparable performance to fine-tuned BERT without requiring annotated training data.These findings highlight the ongoing paradigm shift in NLP toward task unification and the elimination of pre-labeled training data requirements.