An implementation of a neural machine translation system with a LSTM encoder-decoder plus attention architecture.

The authors of this project are Gabriel Kihembo Enemark-Broholm, Christoffer Øhrstrøm and Oldouz Majidi.

We got a BLEU score of 30.37 using local attention versus 16.62 with no attention mechanism.

To recreate and visualize the results run the Jupiter notebook test.ipynb

To train the model using the settings we used run the program with configuration final.json.

python main.py --config final.json

The following arguments may then be used to configure the model:

• --config
  • Path to model configuration (defaults to 'configs/default.json')
• --debug
  • Use a debug dataset.
• --dummy_fixed_length
  • Use a dummy dataset of fixed length sentences
• --dummy_variable_length
  • Use a dummy dataset of variable length sentences
• --iwslt
  • Use the IWSLT dataset
• --name
  • Name used when writing to tensorboard (visualiation)

The default dataset is the Multi30K dataset.

• Python 3.6.5
• Run script hpc/install_requirements.sh