Transition based dependency parser with state embeddings computed by LSTM RNNs. This version is an easier-to-use refactoring of the original code from the ACL 2015 paper.

For the EMNLP character-based model, please check out from the branch char-based and follow the instructions in the README file of that branch.

• A C++ compiler supporting the C++11 language standard
• Boost libraries
• Eigen (newer versions strongly recommended)
• CMake
• gcc (only tested with gcc version 5.3.0, may be incompatible with earlier versions)

mkdir build
cd build
cmake .. -DEIGEN3_INCLUDE_DIR=/path/to/eigen
make -j2

Given a training.conll file and a development.conll formatted according to the CoNLL data format, to train a parsing model with the LSTM parser type the following at the command line prompt:

java -jar ParserOracleArcStdWithSwap.jar -t -1 -l 1 -c training.conll > trainingOracle.txt
java -jar ParserOracleArcStdWithSwap.jar -t -1 -l 1 -c development.conll > devOracle.txt
parser/lstm-parse -P -t trainingOracle.txt -d devOracle.txt --hidden_dim 100 --lstm_input_dim 100 -w sskip.100.vectors --rel_dim 20 --action_dim 20

Link to the word vectors used in the ACL 2015 paper for English: sskip.100.vectors.

Note-1: you can also run it without word embeddings by removing the -w option during training.

Note-2: the training process should be manually stopped when the development result does not substantially improve anymore. Normally, after 5500 iterations.

Note-3: the parser reports (after each iteration) results including punctuation symbols while in the ACL-15 paper we report results excluding them (as it is common practice in those data sets). You can find the eval.pl script from the CoNLL-X Shared Task to get the correct numbers.

There is a pretrained model for English here. It was trained as described above (though training was continued for about 6800 iterations) on the standard train/dev split from the Penn Treebank, converted to Universal Dependencies using the Stanford dependency converter. When tested on Section 23 of the Penn Treebank, it achieves a LAS of 92.16% and a UAS of 93.55%.

Given a test.conll file formatted according to the CoNLL data format:

parser/lstm-parse -m english_pos_2_32_100_20_100_12_20.params -t test.conll

If you are not using the pretrained model, you will need to replace the .params argument with the name of your own trained model file.

The parser will output the CoNLL file with the parsing result.

If you make use of this software, please cite the following:

@inproceedings{dyer:2015acl,
  author={Chris Dyer and Miguel Ballesteros and Wang Ling and Austin Matthews and Noah A. Smith},
  title={Transition-based Dependency Parsing with Stack Long Short-Term Memory},
  booktitle={Proc. ACL},
  year=2015,
}

This software is released under the terms of the Apache License, Version 2.0.

The EOS binary archive library for Boost is included under the MIT License.

For questions and usage issues, please contact cdyer@cs.cmu.edu and miguel.ballesteros@upf.edu. For questions specifically about this easier-to-use version, please contact jdunietz@cs.cmu.edu.