In May 2016 Google released SyntaxNet, a syntactic parser whose performance beat previously proposed approaches.

In this post I will show you how to have SyntaxNet’s syntactic dependencies and other morphological information in Python, precisely how to load NLTK structures such as DependencyGraph and Tree with SyntaxNet’s output.

In this example will use the Portuguese model, but as you will see this can be easily adapted to any language, provided you have already a pre-trained model.


First you need to install SyntaxNet:

Then, you need to download a pre-trained model, from the list of all the available models<language>.zip

As the authors show in the tutorial after installing SyntaxNet and downloading a pre-trained model, one can parse a sentence with the following command:

cat sentences.txt | syntaxnet/models/parsey_universal/ \
$MODEL_DIRECTORY > output.conll

Now I will show you how to parse a file with a sentence per line and use it within Python NLTK.

cat sentences.txt

Quase 900 funcionários do Departamento de Estado assinaram memorando \
que critica Trump.
Meo, Nos e Vodafone arriscam-se a ter de baixar preços a milhões \
de clientes.

First, we load all the sentences into a list and joined them into a single string separated by the newline ‘\n’ character.

import subprocess
import os
import sys

from nltk import DependencyGraph

with open(sys.argv[1], 'r') as f:
    data = f.readlines()
    sentences = [x.strip() for x in data]

Then we will use python subprocess to call SyntaxNet, process the loaded sentences, and fetch the parsed sentences from stdout.

all_sentences = "\n".join(sentences)

# redirect std_error to /dev/null
FNULL = open(os.devnull, 'w')

process = subprocess.Popen(
    'MODEL_DIRECTORY=/Users/dbatista/Downloads/Portuguese; '
    'cd /Users/dbatista/models/syntaxnet; '
    'echo \'%s\' | syntaxnet/models/parsey_universal/ '
    '$MODEL_DIRECTORY 2' % all_sentences,

output = process.communicate()

We process the captured stdout, for each token, the dependencies and other morphological information. Each token is represented by a list with all its syntactic and morphologic information. A list of lists makes the sentence.

processed_sentences = []
sentence = []
for line in output[0].split("\n"):
    if len(line) == 0:
        sentence = []
        word = line.split("\t")

We then join each word/token information in a string separated by a ‘\tab’ character, each word/token in a different line.

deps = []
for sentence in processed_sentences:
    s = ''
    for line in sentence:
        s += "\t".join(line) + '\n'

We then pass this string into the NLTK’s DependenccyGraph and can then see all the dependency triples or an ASCII print of the tree.

for sent_dep in deps:
    graph = DependencyGraph(tree_str=sent_dep.decode("utf8"))
    print "triples"
    for triple in graph.triples():
        print triple
    tree = graph.tree()
    print tree.pretty_print()

For the first sentence we have the following triples and tree:

((u'assinaram', u'VERB'), u'nsubj', (u'funcion\xe1rios', u'NOUN'))
((u'funcion\xe1rios', u'NOUN'), u'nummod', (u'900', u'NUM'))
((u'900', u'NUM'), u'advmod', (u'Quase', u'ADV'))
((u'funcion\xe1rios', u'NOUN'), u'name', (u'Departamento', u'PROPN'))
((u'Departamento', u'PROPN'), u'case', (u'do', u'ADP'))
((u'funcion\xe1rios', u'NOUN'), u'name', (u'Estado', u'PROPN'))
((u'Estado', u'PROPN'), u'case', (u'de', u'ADP'))
((u'assinaram', u'VERB'), u'ccomp', (u'memorando', u'VERB'))
((u'memorando', u'VERB'), u'ccomp', (u'critica', u'VERB'))
((u'critica', u'VERB'), u'mark', (u'que', u'SCONJ'))
((u'critica', u'VERB'), u'dobj', (u'Trump.', u'PROPN'))

      funcionários               memorando
   ________|___________              |
 900  Departamento   Estado       critica
  |        |           |       ______|_______
Quase      do          de    que           Trump.

And for the second sentence:

((u'pode', u'VERB'), u'nsubj', (u'galinha', u'NOUN'))
((u'galinha', u'NOUN'), u'det', (u'Uma', u'DET'))
((u'pode', u'VERB'), u'dobj', (u'ovos', u'NOUN'))
((u'ovos', u'NOUN'), u'case', (u'por', u'ADP'))
((u'ovos', u'NOUN'), u'nummod', (u'250', u'NUM'))
((u'ovos', u'NOUN'), u'nmod', (u'ano.', u'NOUN'))
((u'ano.', u'NOUN'), u'case', (u'por', u'ADP'))

   |             ovos
   |      ________|____
galinha  |        |   ano.
   |     |        |    |
  Uma   por      250  por

I’m still trying to figure out how to have SyntaxNet running as a daemon or service, where we can give a sentence and have as a result, for instance, a JSON object with the syntactic and morphologic information.