Populating Philippine OpenStreetMap Building Type using Large Language Models

Abstract. Accurate building information is essential to a wide range of applications requiring preliminary inputs for humanitarian initiatives, city planning, scientific studies, and navigation systems (Atwal et al., 2022). In the Philippines, which includes the preparation of the Comprehensive Land Use Plan (CLUP), collecting building types can contribute to the development of the demographic maps, exposure maps, and transportation maps, among others. Data collection is an extensive process; thus, community mapping tools such as OpenStreetMap (OSM) present a convenient avenue for collecting descriptive attributes of building types faster. However, the data entry process involves varied tasks beyond basic entry, which results in menial and tedious recording of information that is occasionally inaccurate and lacking. To fill this gap, the study proposes to use Large Language Models (LLMs) trained in Philippine infrastructure contexts. Comparing the LLMS and a classical machine learning model in text classification demonstrates that the classical model performs well on localized features, while LLMs specialize in more complex contextual features. The BERT model, RoBERTa, and the Random Forest model showcased 98.25%, 98.42%, and 92.39% accuracy, respectively, on the testing dataset. This highlights the potential of textual classification using LLMs in urban planning studies.