Generating knowledge graphs from news articles via multi-agent claim extraction and attribution

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

News sources play a fundamental role in democratic societies by shaping how citizens access information about facts, events, and competing viewpoints. It is therefore essential to study how different outlets in different countries fulfil this role. A central aspect of such analyses is the examination of claims expressed by specific voices in the media. However, the large-scale identification, attribution, and extraction of claims from news content remain open challenges, due to the diverse ways in which claims and agents can be expressed, and the distinctive characteristics of journalistic writing. In this paper, we address this challenge by introducing SEER, a comprehensive framework for constructing large-scale knowledge graphs from news articles. To ensure high-fidelity extraction, SEER incorporates a novel Multi-Agent Ensemble module specifically designed to handle the complexity of claim extraction and attribution better than single-model approaches. The resulting knowledge graph is based on an extension of the News Classification Ontology (NCO), which captures the key characteristics of claims in news media. We also present a prototype implementation of the proposed architecture and evaluate it on a corpus of articles drawn from four domains: climate change, UK immigration, the war in Ukraine, and US politics. We perform an evaluation on manually annotated data, showing that SEER’s multi-agent extraction module achieves strong performance (78.22% Macro F1), outperforming ten alternative single-model LLM-based agents on the claim extraction task. We conducted an ablation study and identified recurring error types that highlight remaining challenges in journalistic claim extraction.