simultaneous-nmt

Simultaneous neural machine translation that uses prediction on the source side. This code is a Theano implementation of EMNLP18 paper Prediction Improves Simultaneous Neural Machine Translation. Our implementations are based on dl4mt-simul-trans repository developed by Gu et al.

Dataset

• We have used WMT'15 corpora as our dataset for pretraining and training our agent parameters.
• Newstest 2013 for validation and testset.
• The data should be tokenized and Byte Pair Encoded.

Pretraining

The first step of training the model starts with pretraining Environment. The parameters of the uni-directional LSTM can be changed using the function pretrain_config() in config.py. After setting up configuration, pretrain can be started:

$ export THEANO_FLAGS=device=gpu,floatX=float32
$ python pretrain_uni.py

Training the Agent

Like pretraining the settings of the Model can be configured in config.py. Then training the Agent can be started using sh run_train.sh.