@inproceedings{loginova-etal-2026-news,
    title = "From News Streams to Narrative Intelligence Briefs: {LLM}-Assisted Political Discourse Analysis in the {H}ungarian 2026 Pre-Election Context",
    author = "Loginova, Ekaterina  and
      Ermakov, Maksim  and
      Khramov, Stephan",
    editor = "Afli, Haithem  and
      Bouamor, Houda  and
      Zaghouani, Wajdi  and
      Ghannay, Sahar  and
      Hossain, Shehenaz",
    booktitle = "Proceedings of the 3rd Workshop on Natural Language Processing for Political Sciences ({P}olitical{NLP} 2026)",
    month = may,
    year = "2026",
    address = "Palma, Mallorca (Spain)",
    publisher = "ELRA Language Resources Association (ELRA)",
    url = "https://preview.aclanthology.org/ingest-nlpsi/2026.politicalnlp-1.1/",
    doi = "10.63317/228tnds8d6kd",
    pages = "1--16",
    abstract = "Civil-society organisations, journalists, and fact-checkers monitoring elections require scalable ways to convert high-volume political news into actionable narrative intelligence, yet most NLP pipelines stop at classification outputs that are difficult to operationalise. We present a methodology-driven case study assessing whether large language models, constrained by an explicit analytical schema and multi-stage validation, can reliably transform Hungarian pre-election news into structured narrative intelligence briefs. Using RSS-scraped content from 21 Hungarian-language sources (574 election-relevant articles), we implement a multi-stage pipeline: (1) per-article extraction of narrative event frames grounded in the Narrative Policy Framework (actor{--}action{--}target with role assignment and causal claims) and manipulation techniques from the SemEval propaganda taxonomy; (2) embedding-based clustering of narrative frames with domain classification; and (3) constrained brief generation producing five structured sections{---}narrative summary, character map, manipulation profile, escalation assessment, and counter-strategy{---}where counter-strategies are grounded in verified external sources via curated contextual cards and constrained by evidence-based de-escalation principles. We evaluate brief quality through dual-track evaluation combining three human domain experts and three LLM judges on a single brief, with a scaled 29-brief LLM-as-judge assessment, and document key failure modes across a defined taxonomy. We conclude with implications for trustworthy human-in-the-loop political NLP and the practical limits of LLM-assisted narrative intelligence."
}