📄 From News Streams to Narrative Intelligence Briefs: LLM-Assisted Political Discourse Analysis in the Hungarian 2026 Pre-Election Context

Summary

Teams monitoring political news cannot keep up with pre-election coverage, especially in less-resourced languages. Standard NLP outputs such as topic labels or claim classes are also hard to act on. This case study asks whether LLMs, held to explicit analytical schemas and staged validation, can turn Hungarian news from the 2026 pre-election period into structured "narrative intelligence briefs."

The authors collected articles from 21 outlets across the political spectrum and filtered them for election relevance. They then ran a three-stage, auditable pipeline:

• GPT-4o extracts Narrative Policy Framework frames (actors, roles, targets, causal claims) and SemEval-taxonomy manipulation techniques from each article.
• The frames are embedded and clustered into narratives.
• Claude writes a five-part brief per cluster using only the extraction outputs. Its counter-strategy advice is restricted to curated external evidence.

Human experts and an LLM judge rated descriptive sections higher than prescriptive ones. Most errors came from missing context rather than fabricated content. The work offers a reproducible template for semi-automated analyst support, explicitly not for automated decision-making.

Key points

• Data: RSS collection from 21 Hungarian sources (pro-government, independent, opposition-aligned, general portals) in February 2026 produced 3,708 deduplicated articles. A single-pass GPT-4o classifier kept 574 as election-relevant.
• NPF extraction: GPT-4o (temperature 0, Pydantic structured output) produced 4,266 frame elements, about 7.4 per article. The role set adds "threat" and "proxy" for entities like Brussels or migration.
• Manipulation detection: This uses the SemEval 2023/SlavicNLP 2025 taxonomy (6 categories, 25 labels). Prompt iterations raised F1 from 0.008 to 0.227, comparable to the best shared-task system (macro F1 0.21). Still, 12 of 25 categories got zero recall.
• Clustering: Frames were embedded with paraphrase-multilingual-mpnet-base-v2, reduced with UMAP and clustered with HDBSCAN. This gave 80 clusters (silhouette 0.600, 26.4% noise), mapped to 7 thematic domains. A BERTopic baseline recovered topics but not narrative framing.
• Brief sections: Each brief describes the narrative, maps the characters involved, profiles the manipulation techniques used, assesses the risk of escalation, and proposes a counter-strategy. The counter-strategy section draws on RESIST, GAMMMA+, inoculation theory and the counter- vs. alternative narrative distinction. The code is in the repository aletheos-ngo/politicalNLP2026.