Your Very Own Machine Translation Journey

Your Very Own Machine Translation Journey


Understanding a new technology can be difficult, regardless of the industry where it is applied. Without understanding it, it can be hard to embrace and use it in an informed and efficient manner.

The majority of linguists will have a general idea of what machine translation (MT) is and have a basic understanding of how it works. But really exploring the mechanisms behind MT might not only be an interesting gimmick to play around with, but it is also a catalyst for becoming more apt at interacting with it as a tool in the professional context.

Have you ever been tempted to experiment with machine translation technology yourself? Now you can do so in an easy way and with no cost implications.

Tradumàtica is a free online platform with an intuitive user interface based on Moses. It was created with linguists in mind to allow them to create and customize their own statistical machine translation (SMT) engines.

It can also be used by small companies who are keen to embark on the machine translation journey.

With Tradumàtica, you can create and test your SMT engines to evaluate different translation outcomes in five simple steps:

1. File upload
2. Creation and management of monolingual texts
3. Language model building
4. Creation and management of bilingual texts
5. Training SMT models

The clear and simple interface guides you through each step, making it straightforward to effortlessly follow the end-to-end journey. Completing this round trip helps to understand the components needed to make an MT engine work and how each of them influences the final translation outcome.

Tradumàtica allows users to build statistical machine translation engines but SMT is not the only type of MT system that is out there. Let’s focus for a moment on the characteristics of statistical machine translation and how they differ to other machine translation types.

Scales on wooden floor

Main characteristics of SMT:

  • Returns good results with less training data
  • After the initial training, it can be manipulated and controlled with relative ease
  • Allows more control over terminology
  • It is easy to determine why a certain result has occurred and if it is not correct, to make suitable adjustments
  • Performs well on short phrases and sentences
  • There is a well-established community and knowledge base that covers SMT.

In comparison, a much newer type of MT framework – neural machine translation or NMT – has the following features:

  • It returns significantly more fluent target copy than SMT
  • NMT typically scores better than SMT in both automatic evaluation and human evaluation
  • Translations are produced faster due to the improvement in computing power versus SMT
  • Better handling of unknown words that did not feature in the training data
  • Growing research and translation community focus; switch of research focus from SMT towards NMT
  • Performs best on medium-length sentences.

Although both SMT and NMT offer strong benefits, neither of the methods are flawless and they both need careful consideration before being applied in the professional context – as well as an ongoing human moderation.

Written by Kasia Kosmaczewska
Kasia Kosmaczewska
Kasia Kosmaczewska is Machine Translation Programme Manager at TranslateMedia. She has extensive linguistic experience and a keen interest in machine learning. She spends her free time reading about socio-politics, practicing pilates and travelling.

Related posts