LLMs in Geocoding: Converting Social Media Self-Described Locations into Geographic Coordinates

Abstract

Twitter (Now X) users have been increasing every year. As one of the biggest platforms world wide, the amount of location-related data has increased, yet only a small fraction of posts are geotagged. Traditional geocoding systems and methods struggle on reading social media texts due to its noisy and ambiguous nature. However, recent advancement in AI and Large Language Models (LLMs) has influenced the geospatial field by demonstrating a strong potential for prediction geoinformation. Hybrid approaches combining LLMs and gazetteers especially have gained significant interest. This dissertation explores building a hybrid pipeline that integrates LLMs, social network community voting mechanism to create an end-to-end geocoding system that predicts geo-coordinates and enhance its predictions performance. We found out the LLM models such as GPT-4o had performance best in its base form in both coordinate predictions and polygons at 55.54% accuracy for coordinates, and seen an increase to 62.62% enhancement upon integrating it with network signal. The findings provide valuable insight on encouraging on the integration of LLMs into hybrid geospatial approaches, particularly in enhancing the incorporation of free-form text within geospatial databases and informing the development of advanced hybrid models.