NMTVis

NMTVis is a system that supports users in correcting translations generated with neural machine translation. Our system supports translation with an LSTM-based and the Transformer architecture.

It provides a web interface for visualizations and to support interactive correction.

You can find an introduction to our system here. If you want to reproduce our evaluation you can follow the steps here.

Starting NMTVis

In order to run the system you need Node.js, AngularJS, Python 3, and several Python modules defined in requirements.txt (see below for a more detailed description to prepare everything).

You start the server from the nmtvis-server directory with

python server.py -m [seq|trafo] -sl [de|en] -tl [de|en]

-m: model with options LSTM (seq) and Transformer (trafo)
-sl: source langugage with options German (de) and English (en)
-tl: target langugage with options German (de) and English (en)

You start the client from the nmtvis-client directory with

ng serve

Next, start the application in a browser: localhost:4200/documents.

Now, create a user (or use user = admin, password = admin for the default account) and upload documents for translation and correction. Each document may contain multiple source sentences, each sentence has to be given in a separate line. After uploading a new document you have to refresh the page.

Preparing NMTVis

In the following, we describe in more detail how to set everything up to run the system.

Translation Models

We prepared some trained models: https://doi.org/10.18419/darus-1850. They are for German to English translation and vice versa. You have to place these models in the directories nmtvis-server/.data/models/ under transformer for the Transformer models and seq2seq for the LSTM-based models. Additionally, you have to place the vocabulary files in nmtvis-server/.data/vocab/. Alternatively, you can train the models yourself, however, training will take several days.

For the server:

First, install all python requirements (we used Python 3.8 and 3.9) for the server (in the nmtvis-server/ directory ):

pip3 install -r requirements.txt

For the client:

First, you need Node.js:

On Linux (we used Xubuntu 20.04), you can get it like this:

sudo apt install curl
curl -sL https://deb.nodesource.com/setup_12.x | sudo -E bash -
sudo apt-get install -y nodejs

On Windows 10, download and install Node.js.

On MacOS, we recommend using homebrew:

brew install node

Next, install angular cli:

npm install -g @angular/cli

In case you have permission problems, you can follow these instructions.

Change to the directory nmtvis-client and install AngularJS:

npm install angular

Preliminary Work

This project started as master thesis by Paul Kuznecov (his NMT model and server code) and was later extended by an additional translation model and some additional features.

License

Our project is licensed under the MIT License.

Citation

When referencing our work, please cite the journal paper Visualization-Based Improvement of Neural Machine Translation or conference paper Visual-Interactive Neural Machine Translation:

T. Munz, D. Väth, P. Kuznecov, N. T. Vu, and D. Weiskopf. Visualization-based improvement of neural machine translation. Computers Graphics, 2021.

@article{nmt2021_2,
  title = {Visualization-based improvement of neural machine translation},
  journal = {Computers & Graphics},
  year = {2021},
  doi = {10.1016/j.cag.2021.12.003},
  author = {Tanja Munz and Dirk Väth and Paul Kuznecov and Ngoc Thang Vu and Daniel Weiskopf},}

T. Munz, D. Väth, P. Kuznecov, N. T. Vu, and D. Weiskopf. Visual-interactive neural machine translation. In Proceedings of Graphics Interface 2021, GI 2021, pages 265 – 274. Canadian Information Processing Society, 2021.

@inproceedings{nmt2021,
  author = {Munz, Tanja and Väth, Dirk and Kuznecov, Paul and Vu, Ngoc Thang and Weiskopf, Daniel},
  title = {Visual-Interactive Neural Machine Translation},
  booktitle = {Proceedings of Graphics Interface 2021},
  series = {GI 2021},
  year = {2021},
  isbn = {978-0-9947868-6-9},
  location = {Virtual Event},
  pages = {265 -- 274},
  numpages = {10},
  doi = {10.20380/GI2021.30},
  publisher = {Canadian Information Processing Society},
}