In this project, you are going to implement a neural machine translation model, trained and tested on the IWSLT 2014 data set. To help you start, we have prepared some template (pseudo-) code in this repo. Note that you are not required to use this template code, and it may not be the best implementation, however you may find this a good reference.

File Structure

• nmt.py: contains the neural machine translation model and training/testing code.
• vocab.py: a script that extracts vocabulary from training data
• util.py: contains utility/helper functions

Dataset

The IWSLT 2014 dataset has 150K German-English training sentences. The data/ folder contains a copy of the public release of the dataset. Files with suffix *.wmixerprep are pre-processed versions of the dataset from Ranzato et al., 2015, with long sentences chopped and rared words replaced by a special <unk> token. You could use the pre-processed training files for training/developing (or come up with your own pre-processing strategy), but for testing you have to use the original version of testing files, ie., test.de-en.(de|en).

Environment

The (pseudo-) template code is written in Python 3.6 using some supporting third-party libraries. We provided a conda environment to install Python 3.6 with required libraries. Simply run

conda env create -f environment.yml

Usage

First, we extract a vocabulary file from the training data using the command:

python vocab.py --train-src=data/train.de-en.de.wmixerprep --train-tgt=data/train.de-en.en.wmixerprep data/vocab.bin

This generates a vocabulary file data/vocab.bin. The script also has options to control the cutoff frequency and the size of generated vocabulary, which you may play with.

For training and decoding/testing, you may refer to data/train.sh. Note that in the training script we set the values of some hyper parameters. They are not guaranteed to be the best hyper-parameters, and you are free to play with them. After training and decoding, we call the official evaluation script multi-bleu.perl to compute the corpus-level BLEU score of the decoding results against the gold-standard.