Awesome

Introduction

PR4NMT provides a general framework to incorporate multiple, arbitrary prior knowledge into Neural Machine Translation. Please refer to the following paper for details:

Jiacheng Zhang, Yang Liu, Huanbo Luan, Jingfang Xu and Maosong Sun. 2017. Prior Knowledge Integration for Neural Machine Translation using Posterior Regularization. In Proceedings of ACL 2017, Vancouver, Canada, July.

Installation

PR4NMT is built on top of THUMT. It requires THEANO 0.8.2 or above version (0.8.2 is recommended)

pip install theano==0.8.2

Preparation

Firstly, modify THUMT.config in the config directory to specify features and hyper-parameters. if bilingual dictionary (or phrase table) feature is selected, use cPickle to stringify a bilingual dictionary (or phrase table) in the following format:

import cPickle

word_table = [[source word 1, target word 1] , [source word 2, target word 2], ...] 
cPickle.dump(word_table, open('word_table', 'w'))

Training

The trainer.py script in the scripts folder is used for training NMT models. We recommend initializing MRT with the best model output by MLE using the --init-model-file option. The command for running PR is given by:

python /Users/Jack/THUMT/scripts/trainer.py --config-file /Users/Jack/THUMT/config/THUMT.config --trn-src-file /Users/Jack/THUMT/data/train.src --trn-trg-file /Users/Jack/THUMT/data/train.trg --vld-src-file /Users/Jack/THUMT/data/valid.src --vld-trg-file /Users/Jack/THUMT/data/valid.trg --training-criterion 3 --word-table-file /Users/Jack/THUMT/data/word_table --phrase-table-file /Users/Jack/THUMT/data/phrase_table --init-model-file model_mle_best.npz --device gpu0 

Noting that the effect of PR4NMT is sensitive to MLE weight and PR weight.

Decoding

Given a trained model model_best.npz, please run the following command to translate the test set without evaluation:

python /Users/Jack/THUMT/scripts/test.py --model-file model_best.npz --test-src-file /Users/Jack/THUMT/data/test.src --test-trg-file test.trans --device gpu0

License

The source code is dual licensed. Open source licensing is under the BSD-3-Clause, which allows free use for research purposes. For commercial licensing, please email thumt17@gmail.com.