m3yrin/nar-latent-alignment

Unofficial implementation of "Non-Autoregressive Machine Translation with Latent Alignments" https://arxiv.org/abs/2004.07437

nar-latent-alignment

Unofficial implementation of "Non-Autoregressive Machine Translation with Latent Alignments" https://arxiv.org/abs/2004.07437

• Built on AllenNLP (0.9.0)
• Dataset (En-Ja)
  • A small parallel corpus for English-Japanese translation task provided by @odashi
  • see detail at https://github.com/odashi/small_parallel_enja

Memo

• CTC model only
  • Imputer model is under construction.
• The CTC model is inspired heavily by Libovicky and Helcl (2018)
  • Jindrich Libovicky and Jindrich Helcl. 2018. End-toEnd Non-Autoregressive Neural Machine Translation with Connectionist Temporal Classification. In EMNLP.
  • https://arxiv.org/abs/1811.04719
• Distillation is not tested.

TODO

• Add comments to source codes..
• Understand and build the Imputer model.

Notebook for Google Collaboratory

https://colab.research.google.com/github/m3yrin/nar-latent-alignment/blob/master/colab/experiment.ipynb

Examples (CTC)

0. Download datasets

$ cd datasets
$ git clone https://github.com/odashi/small_parallel_enja.git

1. training

$ allennlp train -f --include-package src -s tmp configs/ctc.jsonnet

2. evaluation

$ allennlp evaluate --output-file tmp/output_test.json --include-package src tmp/model.tar.gz datasets/small_parallel_enja/test

3. prediction
   1. make json inputs

      $ python datasets/make_json.py -I datasets/small_parallel_enja/test.ja -O datasets/test.ja.json
      

   2. predict

      $ allennlp predict --output-file tmp/output_pred.json --include-package src --predictor small_parallel_enja_predictor  tmp/model.tar.gz datasets/test.ja.json