Deep Learning for Toponym Resolution: Geocoding Based on Pairs of Toponyms

Geocoding aims to assign unambiguous locations (i.e., geographic coordinates) place names toponyms) referenced within documents (e.g., spreadsheet tables or textual paragraphs). This task comes with multiple challenges, such as dealing referent ambiguity (multiple places a same name) reference database completeness. In this work, we propose geocoding approach based on modeling pairs of toponyms, which returns latitude-longitude coordinates. One the input toponyms will be geocoded, and second one is used context reduce ambiguities. The proposed deep neural network that uses Long Short-Term Memory (LSTM) units produce representations from sequences character n-grams. To train our model, use toponym co-occurrences collected different contexts, namely in Wikipedia articles) geographical inclusion proximity Geonames data). Experiments areas interest—France, United States, Great-Britain, Nigeria, Argentina Japan—were conducted. Results show models trained co-occurrence data obtained higher accuracy, relations combination can help obtain slightly accuracy fewer sources.