An enormous amount of information about historical events sits in unstructured text — news articles, government reports, and local bulletins — but extracting it by hand at scale is impractical. GroundSource, a methodology from Google Research, addresses this for flooding by turning news coverage into a structured, localized event archive using the Gemini language model. A related overview of the prediction side of this work appears in this piece on AI-powered flash-flood forecasting.
How GroundSource works
The pipeline begins by collecting news reports in which flooding is the primary topic. The primary text of each article is extracted and, because coverage spans many languages, standardized into English with the Google Cloud Translation API. The central step is semantic extraction with Gemini, guided by a carefully designed prompt that walks the model through a structured analytical process:
- Classification: distinguishing reports of real, current, or past floods from articles that only discuss future warnings, policy meetings, or general risk modeling.
- Temporal reasoning: anchoring relative references such as “last Tuesday” to an article’s publication date to pin down when an event actually occurred.
- Spatial precision: identifying fine-grained locations such as neighborhoods and streets and mapping them to standardized geographic areas using the Google Maps Platform.
Validation and accuracy
Technical validation supports the dataset’s use in high-stakes research. In manual reviews, 60% of extracted events were precisely accurate in both location and time, and 82% were accurate enough to be practically useful — for example, capturing the correct administrative district or placing an event within the same day as the reported peak. As an external check, GroundSource captured between 85% and 100% of the severe flood events recorded by the Global Disaster Alert and Coordination System (GDACS), a joint United Nations and European Commission initiative, between 2020 and 2026, indicating that the method picks up smaller localized events as well as major disasters.
Scale and impact
By converting unstructured media into data, GroundSource produced roughly 2.6 million flood-event records drawn from millions of news articles across more than 150 countries — a substantial expansion over traditional monitoring archives, which hold far fewer entries. The dataset has been released openly and is being used to help train flood-forecasting models, contributing to predictions that can be issued in advance of an event.
Limitations and what to watch
A news-derived dataset inherits the biases of its sources. Regions and events with little media coverage will be underrepresented, so an absence of records is not evidence that nothing happened, and reporting volume can reflect attention rather than severity. Automated extraction also makes mistakes: with 60% of events precisely accurate on both location and time, a meaningful share contain errors, which is why the data is best treated as a large-scale signal to be validated against authoritative sources rather than as a definitive record. Translation and place-name resolution add further uncertainty, particularly for fine-grained locations. Used with these caveats in mind, the approach still shows how language models can unlock historical information that was previously trapped in text.
