A spatially-aware algorithm for location extraction from structured documents

Place names facilitate locating and distinguishing geographic space where human activities and natural phenomena occur. Extracting place names at multiple spatial resolutions from text is beneficial in several tasks such as identifying the location of events, enriching gazetteers, discovering connections between events and places, etc. Most modern place name extraction approaches generalize the linguistic rules and lexical features as a universal rule and ignore patterns inherent in place names in the geographic contexts. As a result, they lack spatial awareness to effectively identify place names from different geographic contexts, especially the lesser-known place names. In this research, we develop a novel Spatially-Aware Location Extraction (SALE) algorithm for place name extraction from structured documents that uses a hybrid approach comprising of knowledge-driven and data-driven methods. We build a custom named entity recognition (NER) system based on the conditional random field (CRF) and train/ fine-tune it using spatial features extracted from a dataset based on a given geographic region. SALE uses multiple pathways, including the use of the spatially tuned NER to enhance the efficacy in our place names extraction. The experimental results using a large geographic region show that our algorithm outperforms well-known state-of-the-art place name recognizers.