This is the code for paper "Chunk-Based Bi-Scaled Decoder for Neural Machine Translation".

The chunk-based neural machine translation system is on basis of session 2 of dl4mt-tutorial, which is a attention based encoder-decoder machine translation model.

The main difference between our proposed model and dl4mt is that we use a bi-scaled decoder to leverage the target-side phrase information for better translation, and propose the phrase attention for phrase level soft alignments.

• Python 2.7
• Theano

export THEANO_FLAGS=device=gpu2,floatX=float32
python ./train_nmt_zh2en.py

export THEANO_FLAGS=device=gpu2,floatX=float32
datadir=/home/zhouh/Data/nmt
modeldir=./

python ./translate_gpu.py -n -jointProb \
	$modeldir/model_hal.iter.npz  \
	$modeldir/model_hal.npz.pkl  \
    $datadir/hms.ch.filter.pkl \
	$datadir/hms.en.filter.chunked.pkl \
    $datadir/devntest/MT0${i}/MT0${i}.src \
	./test.result.chunk.${i} 

[1]: Hao Zhou, Zhaopeng Tu, Shujian Huang, Xiaohua Liu, Hang Li and Jiajun Chen. Chunk-based Bi-Scale Decoder for Neural Machine Translation. In Proceeding of ACL 2017, short paper.