the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 05 Jun 2026
From text to geoinformation – A modular approach for extraction of disaster information from web text data
Abstract. The implementation of effective disaster management measures requires comprehensive information about a given flooding situation. Text data from web news offer potentially large volumes of information for this purpose. However, the extraction and spatiotemporal analysis of flood event-related information is inherently demanding due to the immense volume of unstructured text. Addressing this challenge, we present a modular and scalable method that allows the extraction of disaster-relevant information from a large text corpus. This is accomplished by combining domain specific entity extraction with dictionaries, a machine learning model for toponym identification, and hand-crafted rules for entity linking in a modular workflow. The extracted information is augmented with geolocations in order to support spatial analysis. Using the West Germany flooding event 2021 as a case study, we evaluate the capacity of our approach to extract relevant geospatial information at a variety of spatial granularity levels and in the form of various thematic descriptors. By doing so, we outline the capabilities and limitations of this approach for text extraction and analysis. Furthermore, we demonstrate the potential for systematic utilization of text data for improved situational awareness and for disaster management support.
- Preprint
(2844 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-2361', Anonymous Referee #1, 24 Jun 2026
-
RC2: 'Comment on egusphere-2026-2361', Anonymous Referee #2, 16 Aug 2026
This manuscript presents a modular workflow for extracting and geolocating disaster-relevant information from web text data, using the 2021 flood event in western Germany as a case study. The topic is relevant to disaster information extraction and situational awareness, and the integration of text mining with geospatial analysis is potentially useful. However, the methodological contribution and validation of the proposed workflow need to be strengthened before the manuscript can be considered for publication. The following issues should be addressed.
1. The workflow mainly combines existing components, including dictionary-based extraction, Flair NER, gazetteer/Nominatim geocoding, and rule-based entity linking. What is the main methodological novelty of this study compared with existing disaster-related geoparsing and information extraction approaches?
2. The evaluation focuses mainly on the final toponym–DRI relation pairs, while errors from DRI extraction, toponym recognition, and geocoding are not evaluated separately. It would be useful to report the performance of the major modules individually.
3. The F1 scores of the three linking strategies are relatively low, ranging from 33.4% to 42.6%. How does this level of accuracy affect the reliability of the subsequent spatial analysis and its potential use for disaster management?
4. The same-sentence rule (t3) achieves the highest F1 score, but the authors finally use the pruning rule (t2). Please provide a clearer quantitative justification for this choice.
5. The evaluation of geocoding mainly reports whether a toponym can be matched to a location. However, successful matching does not necessarily mean correct geocoding, as illustrated by the erroneous allocation of “Qatar” to an embassy in Bonn. Geocoding accuracy should be evaluated against manually verified locations.
6. The temporal analysis uses the publication or registration date of web documents rather than the actual time of the reported disaster information. Therefore, Figure 7 may represent media attention dynamics rather than the temporal evolution of the flood event. This distinction should be clarified.
7. The comparison with CEMS building-damage data in Section 4.4 is mainly visual. A quantitative spatial comparison, such as correlation, hotspot agreement, or affected-area classification, would provide stronger evidence for the usefulness of the extracted information.
8. Toponyms referring to municipalities, counties, states, rivers, and specific facilities are all represented as points. Using administrative centroids may introduce artificial spatial precision and could influence the geographic-centre analysis in Figure 9. This uncertainty should be considered.
9. The manually annotated reference dataset contains 122 documents and 5,087 toponym–DRI pairs. Please provide more information about the annotation procedure, number of annotators, annotation guidelines, and inter-annotator agreement.
10. The disaster-relevant dictionary is a core component of the method, but its construction is not sufficiently reproducible. Please clarify the TF-IDF selection procedure, manual keyword addition, suffix expansion, and preferably provide the complete dictionary as supplementary material.
11. The method is evaluated only for one flood event and German-language web texts. The claims regarding scalability and transferability should therefore be moderated, or an additional event or study area should be used for independent validation.
12. GDELT-based news data may contain duplicate reports and uneven media coverage. Were duplicated or syndicated articles removed? The authors should also discuss whether the spatial distribution of DRI reflects actual disaster impacts or, partly, differences in media attention.
13. The manuscript repeatedly emphasizes potential applications in situational awareness and disaster management. However, the case study is mainly a retrospective analysis. The authors should clarify whether the proposed workflow is intended for real-time operational use or primarily for post-event analysis.
Citation: https://doi.org/10.5194/egusphere-2026-2361-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 77 | 42 | 14 | 133 | 10 | 7 |
- HTML: 77
- PDF: 42
- XML: 14
- Total: 133
- BibTeX: 10
- EndNote: 7
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This paper presents a workflow for extracting disaster related information from Web text. The authors demonstrated this workflow in a case study on the 2021 flooding event in West Germany. Overall, this paper is more like a project report and lacks a clear research contribution.
- Web text is known to have various biases. Depending on the disaster, some aspects of the disaster may be reported while some other aspects may be ignored. This study specifically used Web news from the GDELT project, and the extracted information will inherit all these biases. This is not to mention the NLP and other methods used to process the text, which have their own algorithmic biases.
- How long is the time delay in the Web news in GDELT? If the news in GDELT is largely delayed, then the extracted information is unlikely going to be useful for disaster response.
- Related to the previous points, how would the information extracted from news be useful for disaster managers? Can the authors provide some concrete examples in which the information extracted from the news is something unknown to disaster managers?
- This paper proposes a general workflow for extracting multiple types of information from Web text. However, there is no comparison with previous methods. A concrete research paper would focus on extracting one or two types of information (e.g., topics or locations) and compare the proposed new method with existing methods.
- The methods used in the workflow seem to be outdated, such as TF-IDF which is an old information retrieval technique. It is unclear what the methodological innovation of this paper is.
- The performance of the workflow as reported in Figure 5 is low, with F1-score being about 0.4. It is unclear whether this workflow can extract information accurately.
- This paper is based on a single case study on a flooding event. A stronger paper would have another case study, ideally on another type of disaster, to demonstrate the generalizability of this workflow.