Browse Teach Curate

Benchmarking Prompted LLMs Against Commercial Moderation APIs for German Hate Speech

An empirical evaluation of non-fine-tuned tools for detecting hate speech in German online-newspaper reader comments. On the domain-specific HOCON34k dataset, GPT-4o (zero-/one-/few-shot prompting) is compared against Google's Perspective API, OpenAI's Moderation API, and the HOCON34k baseline, scored by a combined Matthews-correlation-coefficient and F2 metric. GPT-4o beats both commercial moderation services and exceeds the baseline by roughly five percentage points without task-specific training.

Demerzel Text

Digital Guardians: Can GPT-4, Perspective API, and Moderation API reliably detect hate speech in

An empirical benchmark of off-the-shelf, non-fine-tuned tools for detecting hate speech in German online-newspaper reader comments. Using the HOCON34k dataset (1,592 annotated comments built specific…