Geo-analytical transformations for answering questions with data

Projektbeschreibung

Geografische Fragen von KI beantwortet

Geografische Informationssysteme (GIS) spielen eine zentrale Rolle bei der Beantwortung komplexer räumlicher Fragen in verschiedenen Wissenschaftsbereichen. Der Aufschwung der KI in der Geografie und den Geowissenschaften hat die Hoffnung geweckt, dass geografische Daten problemlos und ohne technische GIS-Kenntnisse wiederverwendet werden können. Besonders vielversprechend sind geografische Frage-Antwort-Methoden (GeoQA), die es den Nutzenden ermöglichen, räumliche Daten in natürlicher Sprache abzufragen. Viele geografische Fragen übersteigen jedoch das reine Abrufen gespeicherter Karten und erfordern die dynamische Umwandlung von Daten zur Erstellung neuer Karten. In diesem Zusammenhang leistet das ERC-finanzierte Projekt GeoTrAnsQData Pionierarbeit mit einem Ansatz, der es der KI ermöglicht, räumliche Fragen mittels analytischer Transformationen zu interpretieren. Insgesamt soll es Wissenschaftlern ermöglicht werden, geografische Phänomene mithilfe der auf natürlicher Sprache basierenden Kartenerstellung zu quantifizieren.

Ziel

Geographic information systems (GIS) and corresponding data sources are of major importance for answering the questions posed by data scientists from various domains of application. The recent trend of adopting Artificial Intelligence (AI)-based methods in Geography and Geoscience has spurred hopes that geographic information can be successfully reused across disciplines without requiring the technical skills of GIS. In this context, geographic question-answering (GeoQA) methods are particularly promising because they enable users without a technical background to answer their questions about geographic space using natural language. However, most geographic questions data scientists may want to answer require some form of geo-analytical transformation of maps. Answer maps need to be (re)generated from data, rather than retrieved from storage. In contrast, stored maps are frequently lacking, outdated, biased, or of insufficient quality. For example, instead of retrieving statistical facts about noise intensity in a city, we might want to know about the coverage of noisy areas in this city. Yet, while the latter map might not be readily available, it could be generated from available sources. State-of-the-art factoid-based GeoQA methods are insufficient for this purpose because they directly match questions to facts, skipping over the problem of how unknown answers may be generated from these sources. In this project, we will develop an entirely novel paradigm for data-based question-answering, called transformative GeoQA, including a theory of geo-analytical transformations. The latter not only allows us to interpret questions in terms of the procedures for answering them but also to describe geodata as a source or purpose of a transformation. In GeoTrAnsQData, we lay the methodical and theoretical foundations for a groundbreaking data ecosystem that enables the quantification of geographic phenomena by generating maps using natural language.