gendermeme-core

A first open-sourced version of GenderMeme: https://gendermeme.org/

This repo contains the code required to run GenderMeme.

Pre-requisites:

• CoreNLP version 3.8.0. Get it here; download and extract.
• Python libraries (install with pip install)
  • pycorenlp
  • numpy

Rough guide for understanding this repo:

This repo contains two folders:

• nlp: This folder contains a file called utils.py. utils.py calls annotate_corenlp, a function which takes the text of an article as an input, and calls pycorenlp, which is a Python wrapper around CoreNLP. The main NLP tasks are performed here (coreference resolution, named entity recognition, dependency parsing, quote detection, etc.) by CoreNLP, and the CoreNLP annotations are returned.
• analysis: This folder contains the following files:
  • utils.py: Contains functions that parse the output of annotate_corenlp to:
    • Identify mentions of people in our article
    • Figure out which mentions refer to the same person, and hence get a list of individuals mentioned in our article (entity resolution)
    • Guess the gender of each individual
    • Attribute quotes and associated verbs to people to figure out who says something in the article.
  • gender.py and gender_babynames.py: Two files of first names and their most likely gender (we use first names to infer gender in some cases). gender.py has been taken from this useful public repo, and gender_babynames.py has been derived by us from R’s library babynames, based on Census data in the US over the years.
  • analysis.py: Contains a utility wrapper function called get_article_info, which allows the user to pass a piece of text and run the whole GenderMeme pipeline on it, and get a JSON output.

How to run this code:

• Start a CoreNLP server (more details here):

  • cd into the directory that you unzipped CoreNLP to, and run:

    java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 150000

• import the function get_article_info from analysis.py and call it as get_article_info(text_to_analyze) with the text to be analyzed passed as a string.
• The output is a JSON string, structured as follows. For each individual, we assign a unique id, and produce a JSON object with the following keys:

  {
  ‘name’: full name of person
  ‘mentions’: a list of positions 
  ‘num_times_mentioned’: int
  ‘gender’: string
  ‘gender_method’: string, one of ‘HONORIFIC’, ‘COREF’ or ‘NAME_ONLY’
  ‘quotes’: list of CoreNLP tokens (words) that this person said
  ‘is_speaker’: whether the person speaks in the article, and a list of reasons for why we think so
  ‘associated_verbs’: list
  }
  

  The overall output is a list of such objects

Example Usage

Run the following from the analysis directory:

>>> from analysis import get_article_info
>>> text = 'Ann Smith and her husband Jim went to the movies. "It was okay," he said.'
>>> get_article_info(text)

It will return a JSON object created from a Python dictionary which, on prettifying, looks like:

[{'associated_verbs': [u'go'],
    'gender': 'FEMALE',
    'gender_method': 'COREF',
    'is_speaker': (False, {'Reasons': []}),
    'mentions': [{'end': 2, 'sent_num': 1, 'start': 1}],
    'name': u'Ann Smith',
    'num_times_mentioned': 1,
    'quotes': []},
{'associated_verbs': [u'go', u'say'],
    'gender': 'MALE',
    'gender_method': 'COREF',
    'is_speaker': (True,
                   {'Reasons': ['Quoted saying 4 words', 'Subject of say']}),
    'mentions': [{'end': 6, 'sent_num': 1, 'start': 6}],
    'name': u'Jim Smith',
    'num_times_mentioned': 1,
    'quotes': [{u'after': u' ',
                u'before': u'',
                u'characterOffsetBegin': 51,
                u'characterOffsetEnd': 53,
                u'index': 2,
                u'lemma': u'it',
                u'ner': u'O',
                u'originalText': u'It',
                u'pos': u'PRP',
                u'speaker': u'7',
                u'word': u'It'},
               {u'after': u' ',
                u'before': u' ',
                u'characterOffsetBegin': 54,
                u'characterOffsetEnd': 57,
                u'index': 3,
                u'lemma': u'be',
                u'ner': u'O',
                u'originalText': u'was',
                u'pos': u'VBD',
                u'speaker': u'7',
                u'word': u'was'},
               {u'after': u'',
                u'before': u' ',
                u'characterOffsetBegin': 58,
                u'characterOffsetEnd': 62,
                u'index': 4,
                u'lemma': u'okay',
                u'ner': u'O',
                u'originalText': u'okay',
                u'pos': u'JJ',
                u'speaker': u'7',
                u'word': u'okay'},
               {u'after': u'',
                u'before': u'',
                u'characterOffsetBegin': 62,
                u'characterOffsetEnd': 63,
                u'index': 5,
                u'lemma': u',',
                u'ner': u'O',
                u'originalText': u',',
                u'pos': u',',
                u'speaker': u'7',
                u'word': u','}]}]