Unsupervised Machine Translation is a research repository that implements both phrase-based SMT and neural MT approaches for translation without parallel corpora. The neural component supports multiple architectures—seq2seq, biLSTM with attention, and Transformer—and allows extensive parameter sharing across languages to improve data efficiency. Training relies on denoising auto-encoding and back-translation, with on-the-fly, multithreaded generation of synthetic parallel data to continually refresh supervision signals. The project also provides scripts to fetch and preprocess monolingual data, learn BPE codes, and train cross-lingual embeddings that bootstrap unsupervised alignment between languages. Beyond the core EMNLP 2018 setup, the codebase exposes additional, optional capabilities such as multi-language training, language model pretraining with shared parameters, and adversarial training.
Features
- Neural MT with seq2seq, biLSTM+attention, and Transformer architectures
- Parameter sharing across encoders/decoders and embeddings for multiple languages
- Denoising auto-encoder training and back-translation with on-the-fly generation
- Utilities to download, tokenize, BPE, and binarize large monolingual corpora
- Cross-lingual embeddings via fastText or alignment methods to initialize models
- Unsupervised PBSMT pipeline with automated Moses training and evaluation
Project Samples
