Year: 2026
Type: article
Abstract: Analyzing media representation requires scalable data extraction, but proprietary LLMs pose data governance and cost risks, limiting adoption in journalism. We address this by developing and evaluating a locally deployable, hybrid LLM pipeline using small, open-source models to extract attributes such as occupations, quotes, and demographics. We test this architecture on 200 human-annotated German news articles from a 100,000-article corpus, spanning diverse genres such as interviews, reports, and reviews. We compare its performance against monolithic, single-prompt LLM baselines such as GPT-4o. Results show the hybrid pipeline significantly outperforms all baselines, solving the critical recall deficits of single-prompt methods. It remains robust to genre variations while baseline performance degrades on narrative texts. This research demonstrates that strategic task decomposition within a local LLM pipeline yields superior extraction performance, establishing a highly accurate and governable alternative to commercial LLMs.
Authors: Yu, Joe, Born, Nadja, Treffers, Theresa
Institution: Technical University of Munich
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164704440