NMT decoding with recombination

Hi, this is an implementation of the paper "Exploring Recombination for Efficient Decoding of Neural Machine Translation" in EMNLP 2018.

Please refer to the paper for more details: paper bib

Usage

Train:

Fisrt, we need a NMT model, use train.py to get it. Please refer to examples/train.sh as an example for the training of En-De and Zh-En models.

Test:

Next, use the trained models for NMT decoding with test.py, see examples/test.sh as the reference.

Options:

Please refer to examples/options.md for descriptions about cmd options.

Requirements:

This code is implemented in python3 with the DyNet toolkit. Please use a relatively newer version of DyNet, we tested it with the version of this commit.