Geographic Question Answering and Spatial Representation Learning

Two papers from UCSB’s STKO Lab (with Ni Lao) on spatial AI — answering geographic questions and encoding spatial features.

Papers Merged

Geographic Question Answering: Challenges, Uniqueness, Classification, and Future Directions — Mai, Janowicz, Zhu, Cai, Lao — AGILE Conference 2021 (peer reviewed)
Multi-Scale Representation Learning for Spatial Feature Distributions Using Grid Cells (Space2Vec) — Anonymous authors (under ICLR 2020 review)

Comprehensive survey analyzing why geographic questions are difficult:

Geographic entities require spatial operations (distance, containment, adjacency)
Questions often involve implicit geographic knowledge
Classification framework for geographic question types
Review of existing GeoQA systems and approaches
Future research directions for spatial QA

Inspired by Nobel Prize-winning neuroscience research on grid cells in mammals:

Multi-scale periodic representation for encoding spatial positions
Analogous to word embeddings but for geographic coordinates
Applications: POI search, land use classification, spatial interpolation, location-aware image classification
Jointly models distributions with vastly different characteristics (common in GIS data)

Extends Ni Lao’s knowledge representation work from Knowledge Vault and Path Ranking Algorithm into the spatial domain. The grid cell encoding connects to the broader theme of neural-symbolic approaches to AI.