Awesome

lstm-parser

Transition based dependency parser with state embeddings computed by LSTM RNNs INCLUDING THE MORPHOLOGICAL FEATURES IN THE CONLL DATA SETS.

Required software

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)

Build instructions

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

Train a parsing model

Having 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 ArcStdMorphologyOracle.jar.jar -t -1 -l 1 -c training.conll > trainingOracle.txt
java -jar ArcStdMorphologyOracle.jar.jar -t -1 -l 1 -c development.conll > devOracle.txt

parser/lstm-parse -T trainingOracle.txt -d devOracle.txt --hidden_dim 100 --lstm_input_dim 100 -w sskip.100.vectors --pretrained_dim 100 --rel_dim 20 --action_dim 20 -t -P

Link to the word vectors that we 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 for both training and parsing.

Note-2: the training process should be 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 eval.pl script from the CoNLL-X Shared Task to get the correct numbers.

Parse data with your parsing model

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

java -jar ArcStdMorphologyOracle.jar.jar -t -1 -l 1 -c test.conll > testOracle.txt

parser/lstm-parse -T trainingOracle.txt -d testOracle.txt --hidden_dim 100 --lstm_input_dim 100 -w sskip.100.vectors --pretrained_dim 100 --rel_dim 20 --action_dim 20 -P -m parser_pos_2_32_100_20_100_12_20-pidXXXX.params

The model name/id is stored where the parser has been trained. The parser will output the conll file with the parsing result.

Pretrained models

TODO

Citation

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,
}

License

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

Contact

For questions and usage issues, please contact cdyer@cs.cmu.edu and miguel.ballesteros@upf.edu.