Code for the Recurrent Neural Network Grammars paper (NAACL 2016), by Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith, after the Corrigendum (last two pages on the ArXiv version of the paper). The code is written in C++.
@inproceedings{dyer-rnng:16,
author = {Chris Dyer and Adhiguna Kuncoro and Miguel Ballesteros and Noah A. Smith},
title = {Recurrent Neural Network Grammars},
booktitle = {Proc. of NAACL},
year = {2016},
}
- A C++ compiler supporting the C++11 language standard
- Boost libraries
- Eigen (latest development release)
- CMake
- EVALB (latest version. IMPORTANT: please put the EVALB folder on the same directory as
get_oracle.pyandsample_input_chinese.txtto ensure compatibility)
Assuming the latest development version of Eigen is stored at: /opt/tools/eigen-dev
mkdir build
cd build
cmake -DEIGEN3_INCLUDE_DIR=/opt/tools/eigen-dev ..
make -j2
sample_input_english.txt (English PTB) and sample_input_chinese.txt (Chinese CTB)
The oracle converts the bracketed phrase-structure tree into a sequence of actions.
The script to obtain the oracle also converts singletons in the training set and unknown words in the dev and test set into a fine-grained set of 'UNK' symbols
python get_oracle.py [training file] [training file] > train.oracle
python get_oracle.py [training file] [dev file] > dev.oracle
python get_oracle.py [training file] [test file] > test.oracle
python get_oracle_gen.py [training file] [training file] > train_gen.oracle
python get_oracle_gen.py [training file] [dev file] > dev_gen.oracle
python get_oracle_gen.py [training file] [test file] > test_gen.oracle
The discriminative variant of the RNNG is used as a proposal distribution for decoding the generative model (although it can also be used for decoding on its own). To save time we recommend training both models in parallel.
On the English PTB dataset the discriminative model typically converges after about 3 days on a single-core CPU device.
nohup build/nt-parser/nt-parser --cnn-mem 1700 -x -T [training_oracle_file] -d [dev_oracle_file] -C [original_dev_file (PTB bracketed format, see sample_input_english.txt)] -P -t --pretrained_dim [dimension of pre-trained word embedding] -w [pre-trained word embedding] --lstm_input_dim 128 --hidden_dim 128 -D 0.2 > log.txt
IMPORTANT: please run the command at the same folder where remove_dev_unk.py is located.
If not using pre-trained word embedding, then remove the --pretrained_dim and -w flags.
The training log is printed to log.txt (including information on where the parameter file for the model is saved to, which is used for decoding under the -m option below)
build/nt-parser/nt-parser --cnn-mem 1700 -x -T [training_oracle_file] -p [test_oracle_file] -C [original_test_file (PTB bracketed format, see sample_input_english.txt)] -P --pretrained_dim [dimension of pre-trained word embedding] -w [pre-trained word embedding] --lstm_input_dim 128 --hidden_dim 128 -m [parameter file] > output.txt
Note: the output will be stored in /tmp/parser_test_eval.xxxx.txt and the parser will output F1 score calculated with EVALB with COLLINS.prm option. The xxxx in /tmp/parser_test_eval.xxxx.txt can be obtained from the process ID of the decoding process (this is NOT the process ID of the training process).
In contrast, the parameter file (following the -m in the decoding command above) can be obtained from log.txt file that logs the training process (e.g. there will be a line near the top of log.txt that for instance will say PARAMETER FILE: ntparse_0_2_32_128_16_128-pid138938.params. In that case, the decoding option should specify -m ntparse_0_2_32_128_16_128-pid138938.params)
If training was done using pre-trained word embedding (by specifying the -w and --pretrained_dim options) or POS tags (-P option), then decoding must alo use the exact same options used for training.
The generative model achieved state of the art results, and decoding is done using sampled trees from the trained discriminative model
For the best results the generative model takes about 7 days to converge.
nohup build/nt-parser/nt-parser-gen -x -T [training_oracle_generative] -d [dev_oracle_generative] -t --clusters clusters-train-berk.txt --input_dim 256 --lstm_input_dim 256 --hidden_dim 256 -D 0.3 > log_gen.txt
The training log is printed to log_gen.txt, including information on where the parameters of the model is saved to, which is used for decoding later.
Decoding with the generative model requires sample trees from the trained discriminative model
build/nt-parser/nt-parser --cnn-mem 1700 -x -T [training_oracle_file] -p [test_oracle_file] -C [original_test_file (PTB bracketed format, see sample_input_english.txt)] -P --pretrained_dim [dimension of pre-trained word embedding] -w [pre-trained word embedding] --lstm_input_dim 128 --hidden_dim 128 -m [parameter file of trained discriminative model] --alpha 0.8 -s 100 > test-samples.props
important parameters
- s = # of samples (all reported results used 100)
- alpha = posterior scaling (since this is a proposal, a higher entropy distribution is probably good, so a value < 1 is sensible. All reported results used 0.8)
utils/cut-corpus.pl 3 test-samples.props > test-samples.trees
build/nt-parser/nt-parser-gen -x -T [training_oracle_generative] --clusters clusters-train-berk.txt --input_dim 256 --lstm_input_dim 256 --hidden_dim 256 -p test-samples.trees -m [parameters file from the trained generative model, see log_gen.txt] > test-samples.likelihoods
utils/is-estimate-marginal-llh.pl 2416 100 test-samples.props test-samples.likelihoods > llh.txt 2> rescored.trees
- 100 = # of samples
- 2416 = # of sentences in test set
rescored.treeswill contain the samples reranked by p(x,y)
The file llh.txt would contain the final language modeling perplexity after marginalization (see the last lines of the file)
Compute generative model parsing accuracy (final step to get parsing accuracy from the generative model)
utils/add-fake-preterms-for-eval.pl rescored.trees > rescored.preterm.trees
utils/replace-unks-in-trees.pl [Discriminative oracle for the test file] rescored.preterm.trees > hyp.trees
utils/remove_dev_unk.py [gold trees on the test set (same format as sample_input_english.txt)] hyp.trees > hyp_final.trees
EVALB/evalb -p COLLINS.prm [gold trees on the test set (same format as sample_input_english.txt)] hyp_final.trees > parsing_result.txt
The file parsing_result.txt contains the final parsing accuracy using EVALB
As the generative model takes a while to train, a pretrained model is available here: https://drive.google.com/open?id=0Bz1ZN2dBHG1dNTRYbmEwNUtoa0E
Since CNN/DyNet relies on Boost to serialize and save/load the model, using this pretrained model requires using boost version 1.60.0 to compile the system. It is important to specify the same training set and clusters in order to load the model (otherwise the model cannot be loaded). Here is the command for training the pretrained model:
build/nt-parser/nt-parser-gen -x -T /usr1/home/akuncoro/rnng-dataset/english/train-gen.oracle -d /usr1/home/akuncoro/rnng-dataset/english/dev-gen.oracle -t --clusters /usr1/home/akuncoro/rnng-dataset/english/clusters-train-berk.txt --input_dim 256 --lstm_input_dim 256 --hidden_dim 256 -D 0.3
The clusters that we used is the "clusters-train-berk.txt", and please contact us to get access to the oracle due to PTB licensing issues.
To get tree samples from the discriminative models (in order to use the generative model to rescore these samples), please find "test-samples.props" here: https://drive.google.com/open?id=0Bz1ZN2dBHG1dTTZXc2FBOXZObm8
We thank Daniel Fried for discovering a bug in the discriminative model.
If there are any issues, please let us know at adhiguna.kuncoro [ AT SYMBOL ] gmail.com, miguel.ballesteros [AT SYMBOL] ibm.com, and cdyer [AT SYMBOL] cs.cmu.edu
This software is released under the terms of the [Apache License, Version 2.0] (http://www.apache.org/licenses/LICENSE-2.0)