Skip to content
Sign in
Artwork for Project Geospatial
Project Geospatial · September 7 · 5 min

FedGeoDay 2025 | Lightning Talk - Jason Gilman

Jason Gilman from Element 84 discusses how to ensure reliability when combining large language models (LLMs) and AI with geospatial data to enable users to ask natural language questions and receive accurate, reliable answers. He introduces "natural language geocoding," a technique developed to translate a user's natural language description of an area (like "within 10 kilometers of the coast") into a precise geographic polygon. This process involves an LLM understanding the request and converting it into a structured graph or tree representing the necessary steps (e.g., database lookups, spatial operations). The key to reliability, he explains, is separating the non-deterministic LLM understanding phase from the deterministic execution phase based on the structured graph. He suggests measuring the accuracy of the LLM's interpretation using a metric called graph edit distance. This approach not only allows for more intuitive user interaction but also enables users to guide AI-driven analyses, such as specifying the type of change they seek in change detection tasks. He also highlights the power of using graphs to represent analysis workflows for execution on distributed computing clusters. • Discusses ensuring reliability when combining LLMs and AI with geospatial data for natural language interactions. • Aims to enable users to ask natural language questions and receive reliable, precise geospatial answers. • Introduces "natural language geocoding" to convert natural language area descriptions (e.g., "within 10 kilometers of the coast") into precise geographic polygons. • Explains the process: an LLM understands the request and converts it into a graph/tree of steps (database lookups, spatial operations). • Highlights the design principle of separating the non-deterministic LLM part from the deterministic execution part for enhanced reliability. • Suggests measuring the accuracy of the LLM's output using graph edit distance. • Demonstrates how this approach allows users to guide AI analysis by specifying parameters, such as the type of change desired in change detection. • Mentions representing workflows as graphs is powerful for execution on clusters. For more content like this check out www.projectgeospatial.com #FedGeoDay #Geospatial #AI #LLMs #NaturalLanguageProcessing #Geocoding #Reliability #Element84 #GeospatialData #MachineLearning #GraphEditDistance

0:00-5:31
Play in

transcript

No transcript — this publisher did not publish one.

show notes

Jason Gilman from Element 84 discusses how to ensure reliability when combining large language models (LLMs) and AI with geospatial data to enable users to ask natural language questions and receive accurate, reliable answers. He introduces "natural language geocoding," a technique developed to translate a user's natural language description of an area (like "within 10 kilometers of the coast") into a precise geographic polygon. This process involves an LLM understanding the request and converting it into a structured graph or tree representing the necessary steps (e.g., database lookups, spatial operations). The key to reliability, he explains, is separating the non-deterministic LLM understanding phase from the deterministic execution phase based on the structured graph. He suggests measuring the accuracy of the LLM's interpretation using a metric called graph edit distance. This approach not only allows for more intuitive user interaction but also enables users to guide AI-driven analyses, such as specifying the type of change they seek in change detection tasks. He also highlights the power of using graphs to represent analysis workflows for execution on distributed computing clusters.


• Discusses ensuring reliability when combining LLMs and AI with geospatial data for natural language interactions.

• Aims to enable users to ask natural language questions and receive reliable, precise geospatial answers.

• Introduces "natural language geocoding" to convert natural language area descriptions (e.g., "within 10 kilometers of the coast") into precise geographic polygons.

• Explains the process: an LLM understands the request and converts it into a graph/tree of steps (database lookups, spatial operations).

• Highlights the design principle of separating the non-deterministic LLM part from the deterministic execution part for enhanced reliability.

• Suggests measuring the accuracy of the LLM's output using graph edit distance.

• Demonstrates how this approach allows users to guide AI analysis by specifying parameters, such as the type of change desired in change detection.

• Mentions representing workflows as graphs is powerful for execution on clusters.


For more content like this check out www.projectgeospatial.com


#FedGeoDay #Geospatial #AI #LLMs #NaturalLanguageProcessing #Geocoding #Reliability #Element84 #GeospatialData #MachineLearning #GraphEditDistance


more episodes

All episodes