Summary:
Question answering (QA) is a computer science discipline within the fields of information
retrieval and natural language processing that is concerned with building systems that
answer questions posed in natural language. Question answering is not just a scientific
challenge, question answering techniques have seen widespread use and adoption in the
industry. They are used in all well-known search engines (e.g., Google and Bing) and
digital assistants (e.g., Siri, Alexa, and Google Assistant) for answering questions like
“What’s the time in New York?” or “How many children did Albert Einstein have?”.
Today, users often are interested in posing questions or information requests with a geo-
spatial dimension to search engines, chatbots and virtual personal assistants. However,
such technologies, like Google and ChatGPT, still struggle to give immediate and precise
answers to complex geographic questions of the form: “Which rivers longer than 10kms
cross London and at least one other city?”, “How many churches exist in a 2-mile radius
of the city center of Austin, Texas?”, “Which countries border Greece, have the euro as
their currency and their population is greater than the population of Greece”.
In this thesis, we deal with the problem of answering such questions over geospatial know-
ledge graphs i.e., knowledge graphs (KGs) which represent knowledge about geographic
features or simply features in the terminology of GIS systems. Geospatial knowledge in
KGs is encoded using latitude/longitude pairs representing the center of features (as e.g.,
in DBpedia and YAGO2), but also more detailed geometries (e.g., lines, polygons) since
these are more appropriate for modeling the geometries of features such as rivers, roads,
countries etc. (as in Wikidata, YAGO2geo, WorldKG and KnowWhereGraph).
We present the geospatial QA system GeoQA3 which is based on GeoQA and its revi-
sion GeoQA2. GeoQA3 represents a radical evolution of the GeoQA family of engines,
introducing Large Language Models in the question-answering pipeline to improve natural
language understanding and facilitate dynamic query generation. This allows GeoQA3 to
understand and correctly answer a larger variety of questions. The engine is available as
open source. As its predecessor, it targets the union of the knowledge graph YAGO2 and
the geospatial knowledge graph YAGO2geo.
Keywords:
Geospatial Question Answering, Geospatial Data, Neural Network, Knowledge Graph, SPARQL
