Getting Started with Google's SyntaxNet for POS-tagging

Google just released (12 March 2016) its open-source project SyntaxNet which they say as "the world's most accurate parser". SyntaxNet has been developed using Google's Tensorflow Framework.

This is a tutorial on OSX to get started with SyntaxNet to tag part-of-speech(POS) in English sentences.

Here are the steps for installation:

1. Install bazel:

   • Install JDK 8.
   • Install Xcode command line tools.
   • Install bazel version 0.2.2 (NOT 0.2.2b) (use darwin version for OSX)

2. Install homebrew and then swig: $ brew install swig

3. Install protocol buffers supported by tensorflow: $ pip install -U protobuf==3.0.0b2

4. Install asciitree to draw parse trees i=on console: $ pip install asciitree

Now check the build and test SyntaxNet using:

$ git clone --recursive https://github.com/tensorflow/models.git
$ cd models/syntaxnet/tensorflow
$ ./configure
$ cd ..
#For testing on Mac run following: (will take some time for tests to pass)
$ bazel test --linkopt=-headerpad_max_install_names \
    syntaxnet/... util/utf8/...

Google has already provided a trained model for POS-tagging English sentences called Parsey McParseface. This is located under syntaxnet/models. To test it inside the terminal we can use the script syntaxnet/demo.sh which provides a basic interface to Parsey McParseface.

# try this in terminal to POS-tag a sentence
$ echo 'Did you see that man?' | syntaxnet/demo.sh

# the following should be the output
Input: Did you see that man ?
Parse:
see VB ROOT
 +-- Did VBD aux
 +-- you PRP nsubj
 +-- man NN dobj
 |   +-- that DT det
 +-- ? . punct