From public discourse to place semantics: modeling semantic similarity of places using geosocial media data

Understanding how people describe, experience, and attach meaning to places is central to advancing human-centered approaches in GIScience. This study introduces a data-driven framework for inferring and modeling place semantics from geosocial media discourse to assess semantic similarity between urban places. Place semantics are conceptualized as distributions of latent topics extracted via transformer-based topic modeling, capturing cognitive-affective, socio-cultural, and functional dimensions of meaning. A key methodological innovation is the integration of geospatial context into topic modeling through geo-text feature engineering. This involves constructing joint vector representations that combine sentence-transformers with spatial information encoded using H3 indexing and Word2Vec-based location embeddings. This enables topic extraction across varying semantic and spatial granularities, enhancing geographic specificity. Topic distributions are aggregated into geospatially indexed place embeddings to support semantic comparison. The framework was implemented in a place-based semantic search system featuring thematic querying and similarity-based place retrieval. Applied to geosocial media data from Flickr, Instagram, and X for the city of Dresden, results show that geospatial context improves topic resolution, reveals both localized and distributed place semantics, and enables exploration of place similarity. This study advances place-based GIScience by bridging natural language modeling and spatial analysis through scalable, human-centered tools for understanding urban places.