Home

Awesome

RBGParser

=========

This project is developed at Natural Language Processing group in MIT. The project contains a Java implementation of a syntactic dependency parser with tensor decomposition, described in the following papers:

[1] Tao Lei, Yu Xin, Yuan Zhang, Regina Barzilay and Tommi Jaakkola. Low-Rank Tensors for Scoring Dependency Structures. ACL 2014. PDF

[2] Yuan Zhang, Tao Lei, Regina Barzilay, Tommi Jaakkola and Amir Globerson. Steps to Excellence: Simple Inference with Refined Scoring of Dependency Trees. ACL 2014. PDF

[3] Yuan Zhang*, Tao Lei*, Regina Barzilay and Tommi Jaakkola. Greed is Good if Randomized: New Inference for Dependency Parsing. EMNLP 2014. PDF

This project is implemented by Tao Lei (taolei [at] csail.mit.edu) and Yuan Zhang (yuanzh [at] csail.mit.edu).

=========

<br>
1. Compilation

To compile the project, first do a "make" in directory lib/SVDLIBC to compile the SVD library. Next, make sure you have Java JDK installed on your machine and find the directory path of Java JNI include files. The directory should contains header files jni.h and jni_md.h. Take a look or directly use the shell script make.sh to compile the rest of the Java code. You have to replace the "jni_path" variable in make.sh with the correct JNI include path. Also, create a "bin" directory in the project directory before running make.sh script.

<br>
2. Data Format

The data format of this parser is the one used in CoNLL-X shared task, which describes a collection of annotated sentences (and the corresponding gold dependency structures). See more details of the format at here and here

<br>
3. Example Usage
3.1 Basic Usage

Take a look at run.sh as an example of running the parser. You could also run the parser for example as follows:

java -classpath "bin:lib/trove.jar" -Xmx20000m parser.DependencyParser \
  model-file:example.model \
  train train-file:example.train \
  test test-file:example.test \
  output-file:example.out

This will train a parser from the training data example.train, save the dependency model to the file example.model, evaluate this parser on the test data example.test and output dependency predictions to the file example.out.

3.2 More Options

The parser will train a 3rd-order parser by default. To train a 1st-order (arc-based) model, run the parser like this:

java -classpath "bin:lib/trove.jar" -Xmx20000m parser.DependencyParser \
  model-file:example.model \
  train train-file:example.train \
  test test-file:example.test \
  model:basic

The argument ``model:MODEL-TYPE'' specifies the model type (basic: 1st-order features, standard: 3rd-order features and full: high-order global features).

There are many other possible running options. Here is a more complicated example:

java -classpath "bin:lib/trove.jar" -Xmx20000m parser.DependencyParser \
  model-file:example.model \
  train train-file:example.train \
  test test-file:example.test \
  output-file:example.out \
  model:standard  C:1.0  iters:5  pruning:false R:20 gamma:0.3 thread:4

This will run a standard model with regularization C=1.0, number of training iteration iters=5, rank of the tensor R=20, number of threads in parallel thread=4, weight of the tensor component gamma=0.3, and no dependency arc pruning pruning=false. You may take a look at RBGParser/src/parser/Options.java to see a full list of possible options.

3.3 Using Word Embeddings

To add unsupervised word embeddings (word vectors) as auxiliary features to the parser. Use option "word-vector:WORD-VECTOR-FILE":

java -classpath "bin:lib/trove.jar" -Xmx20000m parser.DependencyParser \
  model-file:example.model \
  train train-file:example.train \
  test test-file:example.test \
  model:basic \
  word-vector:example.embeddings

The input file example.embeddings should be a text file specifying the real-value vectors of different words. Each line of the file should starts with the word, followed by a list of real numbers representing the vector of this word. For example:

this 0.01 0.2 -0.05 0.8 0.12
and 0.13 -0.1 0.12 0.07 0.03
to 0.11 0.01 0.15 0.08 0.23
*UNKNOWN* 0.04 -0.14 0.03 0.04 0
...
...

There may be a special word *UNKNOWN* used for OOV (out-of-vocabulary) word. Each line should contain the same number of real numbers.