Reinforcement Learning for Non-Autoregressive Neural Machine Translation

Requirements

This repository is an example of fairseq extension for NLP2 course. Please follow the installation instruction of fairseq

0. Get access to project drive (ask your TA)
1. Read project description pdf NLP2_ET_2023.pdf and suggested papers
2. Follow GroupC-fairseq notebook in order to get familiar with running model training and evaluation
   1. data for training is in iwslt14 folder
   2. you have pretrained checkpoint at checkpoint_best.pt
   3. if you are not familiar with NMT you can read https://evgeniia.tokarch.uk/blog/neural-machine-translation/
   4. Some notes on fairseq extension https://evgeniia.tokarch.uk/blog/extending-fairseq-incomplete-guide/
3. Objective implementation:
   1. check rl_criterion.py in criterion folder it gives you a hint how to start working on your objective
   2. Nice explanation of RL for NMT https://www.cl.uni-heidelberg.de/statnlpgroup/blog/rl4nmt/
   3. You can pick any metric and import any library of your choice
4. Run the training with your new objective function:
   1. You can strat fine-tuning to get better/faster results, find checkpoint_best.pt in the drive
   2. set criterion._name to the name of your implemented criterion
   3. It's enough to fine-tune for <1k steps