This repository provides the implementation for the paper "How to reward non-autoregressive models: Minimum risk training with string-based metrics in NMT", which was written as part of the "Natural Language Processing 2" course at the University of Amsterdam.

This repository is a simplified fairseq extension. For installation, please follow the instructions of fairseq.

The experiments were executed using the fairseq.ipynb notebook. Feel free to reach out in case of any questions regarding reproducibility or implementation details.