probablepeople is a python library for parsing unstructured romanized name or company strings into components, using advanced NLP methods. This is based off usaddress, a python library for parsing addresses.
What this can do: Using a probabilistic model, it makes (very educated) guesses in identifying name or corporation components, even in tricky cases where rule-based parsers typically break down.
What this cannot do: It cannot identify components with perfect accuracy, nor can it verify that a given name/company is correct/valid.
probablepeople learns how to parse names/companies through a body of training data. If you have examples of names/companies that stump this parser, please send them over! By adding more examples to the training data, probablepeople can continue to learn and improve.
In the terminal,
pip install probablepeople
Parse some names/companies!
tag are differet methods:
import probablepeople as pp name_str='Mr George "Gob" Bluth II' corp_str='Sitwell Housing Inc' # The parse method will split your string into components, and label each component. pp.parse(name_str) # expected output: [('Mr', 'PrefixMarital'), ('George', 'GivenName'), ('"Gob"', 'Nickname'), ('Bluth', 'Surname'), ('II', 'SuffixGenerational')] pp.parse(corp_str) # expected output: [('Sitwell', 'CorporationName'), ('Housing', 'CorporationName'), ('Inc', 'CorporationLegalType')] # The tag method will try to be a little smarter # it will merge consecutive components, strip commas, & return a string type pp.tag(name_str) # expected output: (OrderedDict([('PrefixMarital', 'Mr'), ('GivenName', 'George'), ('Nickname', '"Gob"'), ('Surname', 'Bluth'), ('SuffixGenerational', 'II')]), 'Person') pp.tag(corp_str) # expected output: (OrderedDict([('CorporationName', 'Sitwell Housing'), ('CorporationLegalType', 'Inc')]), 'Corporation')
Probablepeople uses parserator, a library for making and improving probabilistic parsers - specifically, parsers that use python-crfsuite's implementation of conditional random fields. Parserator allows you to train probablepeople's model (a .crfsuite settings file) on labeled training data, and provides tools for easily adding new labeled training data.
git clone https://github.com/datamade/probablepeople.git cd probablepeople pip install -r requirements.txt python setup.py develop make all nosetests .
If there are name/company formats that the parser isn't performing well on, you can add them to training data. As probablepeople continually learns about new cases, it will continually become smarter and more robust.
NOTE: The model doesn't need many examples to learn about new patterns - if you are trying to get probablepeople to perform better on a specific type of name, start with a few (<5) examples, check performance, and then add more examples as necessary.
For this parser, we are keeping person names and organization names separate in the training data. The two training files used to produce the model are:
To add your own training examples, first put your unlabeled raw data in a csv. Then:
parserator label [infile] [outfile] probablepeople
[infile] is your raw csv and
[outfile] is the appropriate training file to write to. For example, if you put raw strings in
my_companies.csv, you'd use
parserator label my_companies.csv name_data/labeled/company_labeled.xml probablepeople
label command will start a console labeling task, where you will be prompted to label raw strings via the command line. For more info on using parserator, see the parserator documentation.
If you've added new training data, you will need to re-train the model. To set multiple files as traindata, separate them with commas.
parserator train [traindata] probablepeople
probablepeople allows for multiple model files -
person for person names only,
company for company names only, or
generic (both). here are examples of commands for training models:
parserator train name_data/labeled/person_labeled.xml,name_data/labeled/company_labeled.xml probablepeople --modelfile=generic parserator train name_data/labeled/person_labeled.xml probablepeople --modelfile=person parserator train name_data/labeled/company_labeled.xml probablepeople --modelfile=company
If something is not behaving intuitively, it is a bug and should be reported. Report it here by creating an issue: https://github.com/datamade/probablepeople/issues
Help us fix the problem as quickly as possible by following Mozilla's guidelines for reporting bugs.
Your patches are welcome. Here's our suggested workflow:
Copyright (c) 2014 Atlanta Journal Constitution. Released under the MIT License.