2026-05-20T12:31:03Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2015382025-05-12T19:35:22Zoai_pmh:alloai_pmh:repo_items

201538 https://hdl.handle.net/2286/R.2.N.201538 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 186 pages Doctoral Dissertation Academic theses en Tian, Yuanyuan Li, Wenwen Goodchild, Michael Kedron, Peter Baral, Chitta Arizona State University Partial requirement for: Ph.D., Arizona State University, 2025 Field of study: Geography Generative Artificial Intelligence (GenAI) has transformed how people search, analyze, and create content. For example, the OpenAI ChatGPT became the first application that attracted 100 million monthly active users within two months after launch. Perplexity exemplifies people’s growing interest in conversational question-answering. Under the hood, Large Language Models (LLMs) are the core technology of those advancements, which have been integrated into information retrieval systems to enhance tasks such as feature construction, semantic search, and question-answering. Geographic Information Retrieval (GIR) expands Information Retrieval (IR) by focusing on retrieving information with a location component, addressing challenges unique to geospatial data. Despite LLMs demonstrating capabilities in geographic tasks, such as named entity recognition and automating workflows, their adoption in GIR remains underexplored. This dissertation aims to bridge this gap by investigating how LLMs can enhance GIR. The dissertation comprises three interconnected studies, answering questions of what to retrieve, when to integrate, and how to focus. The first study establishes fundamental key elements of effective text-based searches. Using the UCGIS GIS&T Body of Knowledge as a case study, it examines the challenges of linking similar documents within a GIScience encyclopedia. A novel summarizer is proposed to break the bottleneck of semantic search, providing a more efficient solution than manual similarity measure. Building upon the semantic search foundation, the second study introduces spatial and temporal contexts to tackle the problem of similar geographic event recommendation. It designs a retrieval and re-ranking framework that integrates spatial proximity, temporal context, and semantic similarity. The proposed method significantly improves event discovery and monitoring in the Local Environmental Observation Network, surpassing the traditional word-matching-based method or pure dense retrieval. The third study extends it to include multimodal data and spatial-temporal scaling, further enhancing geographic event recommendation. By combining textual and visual content through LLMs, it creates richer, context-aware representations of geographic events, thereby improving the accuracy and practical relevance of retrieved events. The developed methods facilitate more accurate, scalable, and contextual retrieval. Practically, these advancements contribute to improved knowledge content organizing and event monitoring, empowering decision-makers with quick access to relevant spatiotemporal information for addressing critical environmental and societal challenges and informed decision-making. Geographic Information Science and Geodesy Geography GIS Information retrieval Large Language Model Natural Language Processing recommendation system Enhancing Geographic Information Retrieval by Generative AI and Large Language Models