Document Geocoding for Inferring Spatial Familiarity Within Reddit Text (Short Paper)

Abstract

In this paper, we reframe social media user geolocation as a task of inferring a user’s spatial familiarity. We propose an LLM-based approach that identifies places a user is most familiar, including the strength of familiarity. We evaluate this approach using two Reddit evaluation corpora constructed from subreddit user-flair tags: Reddit-London, where users are associated with London boroughs, and Reddit-USA, where users are associated with US states. We compare these datasets with GeoText benchmark dataset. The results show that LLMs can recover place-associated labels from social media text, achieving the strongest performance on Reddit-USA state prediction and moderate performance on fine-grained London borough prediction. We further show that inferred familiar places exhibit clear spatial dependency. Using polygon-based metrics, familiar places inferred from Reddit-London users have a mean distance to labeled boundary of 3.48 km, while Reddit-USA shows stronger alignment with state labels compared to GeoText, with higher inside-label share, and lower spatial dispersion. These findings suggest spatial familiarity provides a useful alternative framing for social media user geolocation, and that Reddit flair-based labels are more closely aligned with expressed spatial familiarity than home location labels in GeoText benchmark dataset.