rbtsbg/diamat

Machine Translation Diagnostics Tool

DiaMaT

An MT diagnostics tool. DiaMaT learns from corpora containing millions of translations but offers explanations on sentence level.

Code accompanying the paper:

@inproceedings{Schwarzenberg_tse_2019,
  title = {Train, Sort, Explain: Learning to Diagnose Translation Models},
  booktitle = {Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT): Demonstrations.},
  author = {Schwarzenberg, Robert and Harbecke, David and Macketanz, Vivien and Avramidis, Eleftherios and M\"oller, Sebastian},
  location = {Minneapolis, Minnesota, USA},
  year = {2019}
  }

Info

DiaMaT deploys the iNNvestigate toolbox.

To facilitate the replication of experiments, if this repo is cloned, 500 MB of data will be directly downloaded from the GitHub LFS server.

Installation

Clone repo w/ lfs:

git lfs clone https://github.com/DFKI-NLP/diamat.git

If the lfs download fails, download diamat/data/output.zip (and other files) from here.

Download embeddings:

wget https://dl.fbaipublicfiles.com/fasttext/vectors-wiki/wiki.de.vec
wget https://dl.fbaipublicfiles.com/fasttext/vectors-wiki/wiki.en.vec

Link wiki.de.vec and wiki.en.vec in ./data/input/word_vectors (see ./data/input/config.INI)

Use Python 3.6, e.g.:

conda create --name diamat python=3.6

Activate environment:

source activate diamat

Install requirements (if CUDA GPU is available install tensorflow-gpu==1.7.0):

pip install -r requirements.txt

Install SpaCy language models

python -m spacy download de
python -m spacy download en

Run / Replicate Experiments

Please, first validate data/input/config.INI.

Unzip ./data/output.zip.

Run train.py and then explain.py, afterwards link data/output/explain.jsonl in server/static/input, e.g.:

CUDA_VISIBLE_DEVICES=0 python3 train.py 2>&1 | tee -a ./data/output/fit_gpu_deeplee_quadro.log &&
CUDA_VISIBLE_DEVICES=0 python3 explain.py 2>&1 | tee -a ./data/output/fit_gpu_deeplee_quadro.log &&
ln -s ../../../data/output/explain.jsonl ./server/static/input/explain.jsonl 2>&1 | tee -a ./data/output/fit_local.log

Then start the server:

cd server &&
sh run_flask.sh

Visit

firefox localhost:5000

Data

All preprocessed data needed to replicate the experiments is contained in ./data/output

• ./data/output/train.jsonl > 1M JSON lines to train the text classifier
• ./data/output/val.jsonl > 100k JSON lines to validate the classifier during training
• ./data/output/xtrain.jsonl > 100k JSON lines to train the explainability method (if needed)
• ./data/output/explain.jsonl > ca 20k JSON lines to test DiaMaT on (drawn from the official WMT test sets / excluding WMT13)

./data/output/explain.jsonl contains contributions (from the experiments) which are overwritten by explain.py

How to use your own data:

• Prepare three parallel text files: source.txt, machine.txt and human.txt
• Then run:

python preprocess.py

Afterwards ./data/output/bundle.jsonl contains the bundled texts, tokens, indices etc. Rename bundle.jsonl, replace source.txt, machine.txt, human.txt and repeat the step if you would like to preprocess more data.

Update ./data/input/config.INI and run train.py and explain.py again.

Update server/static/input/explain.jsonl w/ the new explanations and run the server again.