Semantic similarity is not enough: A novel NLP-based semantic similarity measure in geospatial context

Abstract

Summary: In this study, we addressed two primary challenges: firstly, the issue of domain shift, which pertains to changes in data characteristics or context that can impact model performance, and secondly, the discrepancy between semantic similarity and geographical distance. We employed topic modeling in conjunction with the BERT architecture. Our model was crafted to enhance similarity computations applied to geospatial text, aiming to integrate both semantic similarity and geographical proximity. We tested the model on two datasets, Persian Wikipedia articles and rental property advertisements. The findings demonstrate that the model effectively improved the correlation between semantic similarity and geographical distance. Furthermore, evaluation by real-world users within a recommender system context revealed a notable increase in user satisfaction by approximately 22% for Wikipedia articles and 56% for advertisements.

Keywords

• Geographical information science
• Computer science
• Machine learning