This fork of JoeyNMT is a term project by Jennifer Mell for the course "Introduction to Neural Networks and Sequence-to-Sequence Learning" at Heidelberg University. It implements an unsupervised neural machine translation architecture as described in Unsupervised Neural Machine Translation, Artetxe et al. (2017).
An example configuration for an unsupervised NMT model can be found in configs/unsupervised_small.yaml
.
Four scripts were added to the scripts folder:
preprocess.py
normalises and tokenises a corpus using the sacremoses language-specific normaliser and tokeniserfilter.py
filters sentences containing less than the minimum or more than the maximum number of tokens specified.noise.py
creates a noised parallel corpus for the denoising taskplot_vocab_coverage.py
plots the vocabulary coverage of a model
The training data used for the experiments was sampled from Leipzig Corpora Collection news corpora for German and English. The bilingual development and test data was sampled from WMT'14 News Commentary data.