Searching for places by their spatial configuration is useful in domains that are grounded in the physical world, like urban planning, civil engineering, and travel. However, many spatial search use cases in these domains are not well-served by modern search engines and mapping applications, with many spatial search questions naturally specified using a visual query pattern, which requires computationally expensive spatial pattern matching to resolve. The types of spatial relationships that can be defined are also heterogeneous in nature, making it difficult to reason over them consistently, and capturing such a query typically requires using a pictorial, sketch-map, or graph-based query format, which is incompatible with text-centric search engines and mapping platforms.

In this proposal, we address these challenges, taking an approach that is approximate in nature to address the computational cost, flexible enough to handle heterogeneous relations, and text based to facilitate integration with mainstream search platforms. We leverage techniques in artificial intelligence and natural language processing to develop an ensemble of complementary models and approaches, each capable of performing efficient spatial reasoning over a different type of spatial relationship, which can be described in natural language. Finally, we introduce a system that showcases the viability of our approach to enable fast, robust approximate search over textually-specified spatial queries.