Please refer to a clean version of improved NMT, which incorporates coverage, context gates, and reconstruction models.
In this version, we introduce a coverage mechanism (NMT-Coverage) to indicate whether a source word is translated or not, which proves to alleviate over-translation and under-translation. If you use the code, please cite our paper:
@InProceedings{Tu:2016:ACL,
author = {Tu, Zhaopeng and Lu, Zhengdong and Liu, Yang and Liu, Xiaohua and Li, Hang},
title = {Modeling Coverage for Neural Machine Translation},
booktitle = {Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics},
year = {2016},
}
For any comments or questions, please email the first author.
NMT-Coverage is developed by Zhaopeng Tu, which is on top of lisa-groudhog (https://github.com/lisa-groundhog/GroundHog). It requires Theano0.8 or above version (for the module "scan" used in the trainer).
To install NMT-Coverage in a multi-user setting
python setup.py develop --user
For general installation, simply use
python setup.py develop
NOTE: This will install the development version of Theano, if Theano is not currently installed.
See experiments/nmt/README.md