BIPNMT

Code for our paper "A Reinforcement Learning Approach to Interactive-Predictive Neural Machine Translation" link: https://arxiv.org/abs/1805.01553

The code is written based on Nguyen et.al, 2017's code: https://github.com/khanhptnk/bandit-nmt/

Requirements:

• Python 3.5
• PyTorch 0.3

1. Download and create the vocabulary

• Go to the ``scripts'' folder and run the script:
  • ./download_data.sh
  • ./make_data.sh fr en

2. To train the model; you can follow a sample script in ``scripts'' folder

• run_train.sh

3. To evaluate the trained models

• use the -eval flag (a sample script: run_eval.sh)

1. For BIP-NMT, it takes about 25 mins for training 1000 sentences under a Nvidia Tesla P40 - 24GB RAM GPU.