Discourse-Guided Lateral Reading for Automated News Trustworthiness Assessment
International Conference on AI Research · 18.11.2026 · 7 (1) · 213-221
DOI: 10.34190/icair.7.1.5684 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Conference on AI Research, online erschienen am 18.11.2026. Verfasst von Oliwia Majtyka und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
The rapid proliferation of misinformation across digital news platforms has created an urgent demand for assistive technologies that empower users to evaluate the credibility of online content. Lateral reading, the investigative practice of verifying claims via external sources, is widely recognized as a highly effective strategy for assessing trustworthiness. However, existing automated approaches to question generation (QG) often falter in this context, producing generic or holistic inquiries that fail to target the specific claims, evidence, and named entities within an article, thereby limiting their utility for rigorous real-world fact-checking. In this paper, we present a two-stage pipeline for discourse-guided lateral reading question generation, specifically developed for the TREC 2026 DRAGUN (Detection, Retrieval, and Augmented Generation for Understanding News) Track. Our approach addresses a critical limitation of document-level QG by shifting the focus to fine-grained discourse segments. In the first stage, we utilize a few-shot prompted large language model (LLM) to decompose news articles into functional units, such as claims, evidence, fact descriptions, and supporting context, grounded in the principles of Rhetorical Structure Theory (RST). By identifying the hierarchical and relational roles of these segments, the second stage of our system generates targeted investigative questions tailored to the specific discourse type of each unit. Preliminary evaluation of our discourse-guided pipeline demonstrates a significant improvement in question specificity and relevance when compared to a traditional single-stage baseline. To provide a more formal and standardized validation of these results, we are currently employing the TREC AutoJudge framework. This meta-track evaluation allows us to score our system against importance-weighted rubrics and expert-derived investigative priorities, ensuring the generated questions facilitate objective verification. Our hypothesis is that this structural decomposition significantly reduces low-quality, "summary-style" questions in favor of precise inquiries that ground each prompt in a verifiable claim. These findings suggest that discourse-guided segmentation is a scalable and effective strategy for enhancing the precision of automated fact-checking, with broader implications for retrieval-augmented generation (RAG), media literacy tools, and the development of more transparent, verifiable AI systems.
Zitieren
APA
Majtyka, O., & Liu, T. (2026). Discourse-Guided Lateral Reading for Automated News Trustworthiness Assessment. International Conference on AI Research, 7(1), 213-221. https://doi.org/10.34190/icair.7.1.5684
Vancouver
Majtyka O, Liu T. Discourse-Guided Lateral Reading for Automated News Trustworthiness Assessment. International Conference on AI Research. 2026;7(1):213-221. doi:10.34190/icair.7.1.5684
BibTeX
@article{Majtyka2026,
author = {Majtyka, Oliwia and Liu, Ting},
title = {Discourse-Guided Lateral Reading for Automated News Trustworthiness Assessment},
journal = {International Conference on AI Research},
year = {2026},
volume = {7},
number = {1},
pages = {213--221},
doi = {10.34190/icair.7.1.5684}
}
EUVIMED