The Course of News Events: Comparison of Bottom-Up and Top-Down Approaches for Collecting Text-Based Disaster Data

Jul 16, 2026

News articles provide essential evidence for understanding disaster impacts and adaptation. The core methodological issue in socioenvironmental research is selecting a representative news sample. Two strategies are compared: top-down retrieval of disaster-related articles using an existing disaster inventory, and bottom-up NLP clustering of news texts by temporal and spatial signals. A German landslide dataset is used to evaluate how event-coverage choices affect downstream work in media-inequality studies, disaster monitoring, and inventory enrichment.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

Introduction

Related Work

Data and Event Matching

The Course of News Events: A Comparison of Bottom-Up and Top-Down Approaches for Collecting Text-Based Data about Disasters Brielen Madureira1,2 , Andreas Niekler1,3 , Mariana Madruga de Brito2,1 1LeipzigLab – Climate Discourse, Leipzig University, Germany 2Helmholtz Centre for Environmental Research, Germany 3Computational Humanities, Leipzig University, Germany Correspondence: [brielen.madureira@uni-leizig.de](brielen.madureira@uni-leizig.de) arXiv :2607 .00849v 1 [ cs .CL] 1 Jul 2026 Abstract News articles are an important source of information on disaster impacts and adaptation. A key methodological challenge in socioenvironmental studies is how to select a representative data sample. Two approaches are common: querying news databases top-down with the aid of an existing disaster inventory or using NLP methods to cluster news texts bottom-up based on temporal and spatial features. Using a dataset of German news about landslides worldwide, we compare these approaches and discuss variations in event coverage. Such research design decision can influence the resulting news sample, affecting its use in studies of inequality in media coverage, disaster monitoring and inventory enrichment. 1 Introduction Understanding how humans experience and respond to disasters calls for interdisciplinary collaboration between environmental and (computational) social science (Meehl et al., 2000 ; Albeverio et al., 2006 ; McPhillips et al., 2018 ; de Brito and Sodoge, 2023) . At this intersection blossom fundamental questions about the consequences of climate hazards to society, the relation between adaptation measures and risk reduction, and inequalities in exposure and impact across societal groups. Ensuing findings can ultimately inform allocation of disaster-relief funds (Chapman et al., 2022) and democratic decision making (Soroka et al., 2012) . Synthesizing such knowledge requires gathering information that sprouts on various sources, from sensor-based measurements to digital documents. When text data is in view, the Natural Language Processing field joins that interdisciplinary table with methods for e.g. classification, information extraction, geoparsing and modelling complex social constructs in language use (de Brito et al., 2026) . A fundamental challenge comprises identifying the time, location and impact of disasters ([e.g. land-](e.g. land-) Top-Down Disaster Inventory Bottom-Up targeted queries News Database selection News Events about Disasters Figure 1: Illustrative comparison of two approaches to detect references to disasters in the news. slides, wildfires and floods) . Numerous studies (see Section 2) use global disaster inventories like EM-DAT (Delforge et al., 2025), which is based on high-quality, manually curated data but is also incomplete and unavoidably biased. As a consequence, when the scientific community overly relies on a single source serving as ground truth, the derived collective knowledge may overfit traits of the database rather than the actual phenomena. To counteract this problem, news databases can provide additional event information via two approaches (Figure 1): top-down procedures use known events in external disaster inventories to query news databases for targeted content (e.g. Caiet al., 2025), whereas bottom-up procedures identify, geolocate and cluster news into segmented events, which can then be aligned to or validated against external inventories (e.g. Valkenborg et al., 2026) . But neither is infallible: while the first overlooks events not recorded in inventories, the latter ignores events that were not deemed newsworthy by the media represented in the news database. This paper looks more closely into this matter. We compare news events identified via top-down and bottom-up approaches in dataset of German news about landslides and discuss their advantagesand shortcomings. Such methodological insights, grounded in empirical observations, can strength

FAQs

What problem does the paper address in disaster-related news research?

It addresses how to select a representative data sample from news sources when studying disaster impacts and adaptation.

How do top-down and bottom-up approaches differ for collecting disaster text data?

Top-down methods query news databases using a known disaster inventory, while bottom-up methods identify, geolocate, and cluster news texts into events based on temporal and spatial features.

What dataset is used to compare the two approaches?

The study uses a dataset of nearly 55k German news articles about worldwide landslides, covering the period 2000 to 2024.

Jul 16, 2026

Upload by Olivia Brown