Translation Service Containers for the European Language Grid

Provided by the University of Edinburgh

All services rely on a dockerized translation REST server based on the Marian Toolkit.

English to German and German to English

Available on dockerhub as edinburghnlp/mt4elg-de-en and edinburghnlp/mt4elg-en-de.

These services use “basic transformer” models trained on data provided for WMT19, plus a selection of data from Version 5.0 of Paracrawl.

• First, basic transformers were trained on high-quality WMT parallel data only (Europarl, News Commentary, Rapid).
• N-gram language models were trained on target-language monolingual news data
• The initial translation models and these languag models were used to select additinal parallel training data from Paracrawl v5.0.
• The pooled data was then used to train the final models.

No back-translation was used in the production of these models.

More information: ugermann at inf dot ed dot ac dot uk.

Czech to English

Available on dockerhub as edinburghnlp/mt4elg-cs-en.

This translation service is an ensemble of 3 “basic transformer” models trained for back-translation for UEDIN’s participation in WMT19.

More information: rgrundki at inf dot ed dot ac dot uk.