jacquesfize / mordecai

Full text geoparsing as a Python library

Full text geoparsing as a Python library. Extract the place names from a piece of text, resolve them to the correct place, and return their coordinates and structured geographic information.

Example usage

>>> from mordecai import Geoparser
>>> geo = Geoparser()
>>> geo.geoparse("I traveled from Oxford to Ottawa.")

[{'country_conf': 0.96474487,
  'country_predicted': 'GBR',
  'geo': {'admin1': 'England',
   'country_code3': 'GBR',
   'feature_class': 'P',
   'feature_code': 'PPLA2',
   'geonameid': '2640729',
   'lat': '51.75222',
   'lon': '-1.25596',
   'place_name': 'Oxford'},
  'spans': [{'end': 22, 'start': 16}],
  'word': 'Oxford'},
 {'country_conf': 0.83302397,
  'country_predicted': 'CAN',
  'geo': {'admin1': 'Ontario',
   'country_code3': 'CAN',
   'feature_class': 'P',
   'feature_code': 'PPLC',
   'geonameid': '6094817',
   'lat': '45.41117',
   'lon': '-75.69812',
   'place_name': 'Ottawa'},
  'spans': [{'end': 32, 'start': 26}],
  'word': 'Ottawa'}]

Mordecai requires a running Elasticsearch service with Geonames in it. See "Installation" below for instructions.

Installation and Requirements

1. Mordecai is on PyPI and can be installed for Python 3 with pip:

pip install mordecai

2. You should then download the required spaCy NLP model:

python -m spacy download en_core_web_lg

3. In order to work, Mordecai needs access to a Geonames gazetteer running in Elasticsearch. The easiest way to set it up is by running the following commands (you must have Docker installed first).

docker pull elasticsearch:5.5.2
wget https://s3.amazonaws.com/ahalterman-geo/geonames_index.tar.gz --output-file=wget_log.txt
tar -xzf geonames_index.tar.gz
docker run -d -p 127.0.0.1:9200:9200 -v $(pwd)/geonames_index/:/usr/share/elasticsearch/data elasticsearch:5.5.2

See the es-geonames for the code used to produce this index.

To update the index, simply shut down the old container, re-download the index from s3, and restart the container with the new index.

Citing

If you use this software in academic work, please cite as

@article{halterman2017mordecai,
  title={Mordecai: Full Text Geoparsing and Event Geocoding},
  author={Halterman, Andrew},
  journal={The Journal of Open Source Software},
  volume={2},
  number={9},
  year={2017},
  doi={10.21105/joss.00091}
}

How does it work?

Mordecai takes in unstructured text and returns structured geographic information extracted from it.

• It uses spaCy's named entity recognition to extract placenames from the text.

• It uses the geonames gazetteer in an Elasticsearch index (with some custom logic) to find the potential coordinates of extracted place names.

• It uses neural networks implemented in Keras and trained on new annotated data labeled with Prodigy to infer the correct country and correct gazetteer entries for each placename.

The training data for the two models includes copyrighted text so cannot be shared freely, but get in touch with me if you're interested in it.

API and Configuration

When instantiating the Geoparser() module, the following options can be changed:

• es_ip : Where the Geonames Elasticsearch service is running. Defaults to localhost, which is where it runs if you're using the default Docker setup described above.
• es_port : What port the Geonames Elasticsearch service is running on. Defaults to 9200, which is where the Docker setup has it
• country_confidence : Set the country model confidence below which no geolocation will be returned. If it's really low, the model's probably wrong and will return weird results. Defaults to 0.6.
• verbose : Return all the features used in the country picking model? Defaults to False.
• threads: whether to use threads to make parallel queries to the Elasticsearch database. Defaults to True, which gives a ~6x speedup.

geoparse is the primary endpoint and the only one that most users will need. Other methods are primarily internal to Mordecai but may be directly useful in some cases:

• infer_country take a document and attempts to infer the most probable country for each.
• query_geonames and query_geonames_country can be used for performing a search over Geonames in Elasticsearch
• methods with the _feature prefix are internal methods for calculating country picking features from text.

batch_geoparse takes in a list of documents and uses spaCy's nlp.pipe method to process them more efficiently in the NLP step.

Advanced users on large machines can modify the lru_cache parameter from 250 to 1000. This will use more memory but will increase parsing speed.

Tests

Mordecai includes unit tests. To run the tests, cd into the mordecai directory and run:

pytest

The tests require access to a running Elastic/Geonames service to complete. The tests are currently failing on TravisCI with an unexplained segfault but run fine locally. Mordecai has only been tested with Python 3.

Acknowledgements

An earlier verion of this software was donated to the Open Event Data Alliance by Caerus Associates. See Releases or the legacy-docker branch for the 2015-2016 and the 2016-2017 production versions of Mordecai.

This work was funded in part by DARPA's XDATA program, the U.S. Army Research Laboratory and the U.S. Army Research Office through the Minerva Initiative under grant number W911NF-13-0332, and the National Science Foundation under award number SBE-SMA-1539302. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of DARPA, ARO, Minerva, NSF, or the U.S. government.

Contributing

Contributions via pull requests are welcome. Please make sure that changes pass the unit tests. Any bugs and problems can be reported on the repo's issues page.