Neural Machine Translation Codebase in Matlab

Code to train state-of-the-art neural machine translation systems that can handle very complex languages like Czech using hybrid word-character models described in our ACL'16 paper.

This codebase can also train general attention-based models described in our EMNLP'15 paper and has all the functionalities of our previous nmt.matlab codebase.

Why Matlab? It was a great learning experience for me to be able to derive by hand all gradient formulations and implement everything from crash! Matlab supports GPU, so the code is also very fast.

Features:

• Train hybrid word-character as well as general attention-based models.
• Beam-search decoder that can ensembles models including hybrid ones.
• Code to compute source word representations and evaluate on the word similarity tasks or do tsne plots.
• Code to compute sentence representations and rerank scores.

Citations:

If you make use of this codebase in your research, please cite our paper

@inproceedings{luong2016acl_hybrid,
 author = {Luong, Minh-Thang  and  Manning, Christopher D.},
 title = {Achieving Open Vocabulary Neural Machine Translation with Hybrid Word-Character Models},
 booktitle = {Association for Computational Linguistics (ACL)},
 address = {Berlin, Germany},
 month = {August},
 year = {2016}
}

• Thang Luong [email protected], 2014, 2015, 2016

Files

README.md       - this file
code/           - main Matlab code
  trainLSTM.m: train models
  testLSTM.m: decode models
  computeSentRepresentations.m: compute encoder representations.
  computeRerankScores.m: compute decoding scores.
data/           - toy data
scripts/        - scripts

Core scripts

• We provide an one-for-all script that performs all the preprocessing steps & train a translation model

1-prepare_and_train.sh <trainPrefix> <validPrefix> <testPrefix> <srcLang> <tgtLang> <wordVocabSize> <charVocabSize> <outDataDir> <outModelDir> [options]
  trainPrefix   expect train files trainPrefix.(srcLang|tgtLang)
  validPrefix   expect valid files validPrefix.(srcLang|tgtLang)
  testPrefix    expect test files testPrefix.(srcLang|tgtLang)
  srcLang     Source languague
  tgtLang     Target languague
  wordVocabSize   Word vocab size.
  charVocabSize   Character vocab size. If 0, run word-based models.
  outDataDir    Output data directory where we save preprocessed data
  outModelDir   Output model directory that we save during training
  options     Options to trainLSTM

The script is smart enough to check if preprocessed data files have been created in so that we can reuse. When is greater than 0, we will train hybrid word-character models.

Examples

Training

• Process data & train a hybrid model:

./scripts/1-prepare_and_train.sh data/train.10k data/valid.100 data/test.100 de en 1000 50 data.hybrid.50 model.hybrid.w1000.c50

• We can also add options such as dropout (keep probability = 0.8) and use 2-layer character-level models as below:

./scripts/1-prepare_and_train.sh data/train.10k data/valid.100 data/test.100 de en 1000 50 data.hybrid.50 model.hybrid.w1000.c50.dropout0.8.charLayer2 "'dropout',0.8,'charNumLayers',2"

• To train regular attention-based sequence-to-sequence NMT:

./scripts/1-prepare_and_train.sh data/train.10k data/valid.100 data/test.100 de en 1000 0 data.1000 model.w1000

Testing (Decoding)

Miscellaneous

• Gradient checks:

./scripts/run_grad_checks.sh > output/grad_checks.txt 2>&1

Then compare with the provided grad check outputs data/grad_checks.txt. They should look similar.

• The Matlab code/ directory further divides into sub-directories:

  basic/: define basic functions like sigmoid, prime. It also has an efficient way to aggreate embeddings.
  layers/: we define various layers like attention, LSTM, etc. with forward and backprop code.
  misc/: things that we haven't categorized yet.
  preprocess/: deal with data.
  print/: print results, logs for debugging purposes.
  wordsim/: word similarity task