Awesome

Multimodal simultaneous NMT

This repository is a stripped down clone of the upstream nmtpytorch repository.

Contributors

• Julia Ive added all parts relating to reinforcement learning (RL) based simultaneous MT and MMT.
• As part of her MSc. thesis, Veneta Haralampieva contributed layers & models for Transformers support to simultaneous MT and MMT.

Installation

The installation should be straightforward using anaconda. The below command will install the toolkit in develop mode into a newly created simnmt environment. This will allow your changes to the GIT checkout folder to be instantaneously reflected to the imported modules and executable scripts.

conda env create -f environment.yml

Unsupervised reward in RL for MT

Code for the paper:

<b>Exploring Supervised and Unsupervised Rewards in Machine Translation</b>. Julia Ive, Zixu Wang, Marina Fomicheva, Lucia Specia (2021). To appear in the Proceedings of EACL.

1. Follow the guidelines above to install the main code

2. Pre-train the actor (modify the paths in the config):

$ nmtpy train -C ./configs/unsupRL/en_de-cgru-nmt-bidir-base.conf

3. Train SAC with the unsupervised reward (modify the paths in the config, pretrained_file indicates the location of the pre-trained Actor):

$ nmtpy train -C ./configs/unsupRL/en_de-cgru-nmt-bidir-diyan.conf

The implementation of the Soft Actor-Critic framework follows the architecture and style of the Deep-Reinforcement-Learning-Algorithms-with-PyTorch library, developed by Petros Christodoulou.