PyStanfordDependencies

Python interface for converting Penn Treebank trees to Universal Dependencies and Stanford Dependencies.

Example usage

Start by getting a StanfordDependencies instance with StanfordDependencies.get_instance():

>>> import StanfordDependencies
>>> sd = StanfordDependencies.get_instance(backend='subprocess')

get_instance() takes several options. backend can currently be subprocess or jpype (see below). If you have an existing Stanford CoreNLP or Stanford Parser jar file, use the jar_filename parameter to point to the full path of the jar file. Otherwise, PyStanfordDependencies will download a jar file for you and store it in locally (~/.local/share/pystanforddeps). You can request a specific version with the version flag, e.g., version='3.4.1'. To convert trees, use the convert_trees() or convert_tree() method (note that by default, convert_trees() can be considerably faster if you're doing batch conversion). These return a sentence (list of Token objects) or a list of sentences (list of list of Token objects) respectively:

>>> sent = sd.convert_tree('(S1 (NP (DT some) (JJ blue) (NN moose)))')
>>> for token in sent:
...     print token
...
Token(index=1, form='some', cpos='DT', pos='DT', head=3, deprel='det')
Token(index=2, form='blue', cpos='JJ', pos='JJ', head=3, deprel='amod')
Token(index=3, form='moose', cpos='NN', pos='NN', head=0, deprel='root')

This tells you that moose is the head of the sentence and is modified by some (with a det = determiner relation) and blue (with an amod = adjective modifier relation). Fields on Token objects are readable as attributes. See docs for additional options in convert_tree() and convert_trees().

Visualization

If you have the asciitree package, you can use a prettier ASCII formatter:

>>> print sent.as_asciitree()
 moose [root]
  +-- some [det]
  +-- blue [amod]

If you have Python 2.7 or later, you can use Graphviz to render your graphs. You'll need the Python graphviz package to call as_dotgraph():

>>> dotgraph = sent.as_dotgraph()
>>> print dotgraph
digraph {
        0 [label=root]
        1 [label=some]
                3 -> 1 [label=det]
        2 [label=blue]
                3 -> 2 [label=amod]
        3 [label=moose]
                0 -> 3 [label=root]
}
>>> dotgraph.render('moose') # renders a PDF by default
'moose.pdf'
>>> dotgraph.format = 'svg'
>>> dotgraph.render('moose')
'moose.svg'

The Python xdot package provides an interactive visualization:

>>> import xdot
>>> window = xdot.DotWindow()
>>> window.set_dotcode(dotgraph.source)

Both as_asciitree() and as_dotgraph() allow customization. See the docs for additional options.

Backends

Currently PyStanfordDependencies includes two backends:

• subprocess (works anywhere with a java binary, but more overhead so batched conversions with convert_trees() are recommended)
• jpype (requires jpype1, faster than the subprocess backend, also includes access to the Stanford CoreNLP lemmatizer)

By default, PyStanfordDependencies will attempt to use the jpype backend. If jpype isn't available or crashes on startup, PyStanfordDependencies will fallback to subprocess with a warning.

Universal Dependencies status

PyStanfordDependencies supports most features in Universal Dependencies (see issue #10 for the most up to date status). PyStanfordDependencies output matches Universal Dependencies in terms of structure and dependency labels, but Universal POS tags and features are missing. Currently, PyStanfordDependencies will output Universal Dependencies by default (unless you're using Stanford CoreNLP 3.5.1 or earlier).

Related projects

• clearnlp-converter (uses clearnlp instead of Stanford CoreNLP for dependency conversion)

More information

Licensed under Apache 2.0.

Written by David McClosky (homepage, code)

Bug reports and feature requests: GitHub issue tracker

Release summaries

• 0.3.1 (2015.11.02): Better collapsed universal handling, bugfixes
• 0.3.0 (2015.10.09): Support copy nodes, more input checking/debugging help, example convert.py program
• 0.2.0 (2015.08.02): Universal Dependencies support (mostly), Python 3 support (fully), minor API updates
• 0.1.7 (2015.06.13): Bugfixes for JPype, handle version mismatches in IBM Java
• 0.1.6 (2015.02.12): Support for graphviz formatting, CoreNLP 3.5.1, better Windows portability
• 0.1.5 (2015.01.10): Support for ASCII tree formatting
• 0.1.4 (2015.01.07): Fix CCprocessed support
• 0.1.3 (2015.01.03): Bugfixes, coveralls integration, refactoring
• 0.1.2 (2015.01.02): Better CoNLL structures, test suite and Travis CI support, bugfixes
• 0.1.1 (2014.12.15): More docs, fewer bugs
• 0.1 (2014.12.14): Initial release