There is one patch file for each evaluation dataset used in our experiments. The format of each file contents is tab-separated values:
- toponym
- latitude, longitude in dataset
- unified latitude, longitude using WikiData
There may be more than one examples for each toponym in the respective datasets hence the unique rows in the patch files would be less than total evaluation examples.
#2 (i.e. latitude in dataset , longitude in dataset) is provided only for reference to connect back to the original datasets. For evaluation itself we use #1 and #3 only.
The patches are released under GNU General Public License v3.0, same as the one used in open-sourced evaluation datasets on which these are based on.
The dataset is associated with the paper on Spatial Language Representation with Multi-Level Geocoding
If you use these data fixes, please cite:
@misc{mlg_geocoding,
title={Spatial Language Representation with Multi-Level Geocoding},
author={Sayali Kulkarni and Shailee Jain and Mohammad Javad Hosseini and Jason Baldridge and Eugene Ie and Li Zhang},
archivePrefix={arXiv},
primaryClass={cs.CL}
}