Introduction

This repository contains Neural Machine Translation tools and models built at Softcatalà using OpenNMT-tf 2 and TensorFlow 2

Description of the directories

• data-processing-tools: set of data processing tools that convert for different formats to OpenNMT plain text input format
• serving: contains a microservice that provides a transtion API for web service and batch file processing.
• use-models-tools: contains tools to use the models to translate text files or PO files
• evaluate: set of tools and corpus to evaluatate diferent translation systems (including BLEU scores)
• training: scrips and configurations to train the models

Models

All the Softcatalà built models are available here: https://github.com/Softcatala/nmt-models

Serving

Requirements

You need Docker and Make for which there are different implementations depending on your operating system.

Serving the models in production

You can build and run the docker translation service:

• Build the solution make build-all
• Run make docker-run-translate-service
• Open in your browser http://localhost:8700/translate?langpair=en|ca&q=Hello!

To run exactly the system in production you also need docker-compose. You can execute it by running:

• make docker-run-all-services

Apertium API

One of the use cases for Machine Translation is to use it to speed up the work of translators.

In order to integrate easily with already existing translation tools we support the Apertium Web API. This means that you can use any tool that has support for Apertium.

We confirm that the following tools work using Apertium pluggins:

• Okapi Framework
• OmegaT translation plugin

Supported methods

Method	Verb
/translate	GET or POST
/listLanguageNames	GET
/listPairs	GET

Using the models in your machine

This is useful for example if you want to translate large volumes using our prebuild English - Catalan models using the same exact version that we have in production:

• Build command line tool make docker-build-use-models-tools

To test quickly that every works:

• echo "Hello World" > input.txt
• docker run -it -v "$(pwd)":/srv/files/ --env COMMAND_LINE="-f input.txt -t output.txt -m eng-cat" --rm use-models-tools --name use-models-tools
• more output.txt

To translate PO files:

• File ca.po is your current directory
• docker run -it -v "$(pwd)":/srv/files/ --env COMMAND_LINE="-f ca.po -m eng-cat" --env FILE_TYPE='po' --rm use-models-tools --name use-models-tools

The translated file will be ca.po-ca.po

To translate a text file from Catalan to English:

• echo "Hola món" > input.txt
• docker run -it -v "$(pwd)":/srv/files/ --env COMMAND_LINE="-f input.txt -t output.txt -m cat-eng" --rm use-models-tools --name use-models-tools
• more output.txt

Note: that the parameter -m cat-eng indicates the translation model to use.

Development

Performance test of the translation service

It is important to understand that there are no major performance regressions.

Install the wrk performance testing tool by using sudo apt-get install wrk

Follow these steps:

• Run make docker-run-all-services to run all the services for the performance test
• Use serving/perf-tests script to run the performance test

License

See license

Contact

Email address: Jordi Mas: [email protected]