.. image:: https://travis-ci.org/dmcc/PyStanfordDependencies.svg?branch=master :target: https://travis-ci.org/dmcc/PyStanfordDependencies
.. image:: https://badge.fury.io/py/PyStanfordDependencies.png :target: https://badge.fury.io/py/PyStanfordDependencies
.. image:: https://coveralls.io/repos/dmcc/PyStanfordDependencies/badge.png?branch=master :target: https://coveralls.io/r/dmcc/PyStanfordDependencies?branch=master
Python interface for converting Penn Treebank <http://www.cis.upenn.edu/~treebank/>
trees to Universal Dependencies <http://universaldependencies.github.io/docs/>
and Stanford Dependencies <http://nlp.stanford.edu/software/stanford-dependencies.shtml>
_.
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 <http://nlp.stanford.edu/software/corenlp.shtml>
or
Stanford Parser <http://nlp.stanford.edu/software/lex-parser.shtml>
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()
.
If you have the asciitree <https://pypi.python.org/pypi/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 <http://graphviz.org/>
to render your graphs. You'll need the Python graphviz <https://pypi.python.org/pypi/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 <https://pypi.python.org/pypi/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.
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 <https://pypi.python.org/pypi/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.
PyStanfordDependencies supports most features in Universal Dependencies <http://universaldependencies.github.io/docs/>
(see issue #10 <https://github.com/dmcc/PyStanfordDependencies/issues/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).
clearnlp-converter <https://pypi.python.org/pypi/clearnlp-converter/>
(uses clearnlp <http://www.clearnlp.com/>
instead of Stanford CoreNLP <http://nlp.stanford.edu/software/corenlp.shtml>
_ for
dependency conversion)Licensed under Apache 2.0 <http://www.apache.org/licenses/LICENSE-2.0>
_.
Written by David McClosky (homepage <http://nlp.stanford.edu/~mcclosky/>
, code <http://github.com/dmcc>
)
Bug reports and feature requests: GitHub issue tracker <http://github.com/dmcc/PyStanfordDependencies/issues>
_
convert.py
programJPype
, handle version mismatches
in IBM Javagraphviz
formatting, CoreNLP 3.5.1,
better Windows portabilityCCprocessed
support