1. Why and how does the Centre use neural machine translation?
NMT has now been widely adopted by the translation industry as it outperforms, in a disruptive way, the earlier MT systems based on statistical models. It uses deep neural networks that are able to learn complex linguistic patterns and offer more accurate and fluent translations.
At the Centre, we have built about 100 engines in-house, using our past translations (translation memories) as well as other EU data. They are called ‘Neural Integrated Custom Engines’ (NICE) and they can be used to produce general and domain-specific translations in 23 EU languages.
With the integration of these engines, we implemented a customised workflow to make the most of NMT in combination with our own translation memories.
2. Which CdT services are based on machine translation?
Automatic translation and light post-editing are the two services based on machine translation.
Standard translations and so-called paste ‘n’ go translations (short texts for which formatting does not play a role), carried out by human translators, can include in specific cases the full post-editing of machine-translated output. This means that translation requests that meet the right conditions for using the Centre’s custom machine translation engines are automatically pre‑processed using these engines. Such pre-translated documents undergo full post-editing by the Centre’s linguists to make sure that the output is entirely comparable with the result obtained through traditional translation (i.e. human translation using a translation memory tool combined with machine translation).
3. What is the ‘automatic translation’ service provided by the Centre?
This service allows our clients to request the automatic translation of documents and other types of texts based on a combination of the Centre’s translation memories and custom machine translation engines. It does not include any quality control by linguists, nor any preparation of the source documents for technical processing or formatting. Automatic translation requests are processed in a very short timeframe.
When do we advise our clients to use automatic translation? Automatic translation can be used to quickly translate a text from a language one does not understand in order to have an idea of its content. This kind of translation is meant for internal use and should not be published.
4. What is the added value of the Centre’s automatic translation service?
When ordering automatic translation on the CdT Client Portal, clients benefit from:
- Customised technology - our NMT engines are uniquely tailored to our clients’ translation needs.
- Integration into the eCdT translation workflow - this service works seamlessly for clients and leverages our state-of-the-art IT technology.
- Privacy – our clients’ translation requests are processed 100% on-site and will never be sent to the cloud.
5. What is light post-editing at the Centre?
Light post-editing is a translation service in which our linguists improve the machine translation output and translation memory matches. They make sure that the text is understandable and complete by checking that:
- nothing has been mistranslated;
- no important information has been left out;
- no information has been added;
- the spelling is correct;
They are not expected to make any stylistic corrections unless these affect the meaning of the text.
When do we advise our clients to use light post-editing? This service is particularly suitable where the highest level of quality is not required in the final text. Whether or not to publish lightly post-edited texts is a decision that we advise our clients to take for themselves, as each use case is different.
Tip: For more insight into this service, take a look at our video on light post-editing.
6. How does the Centre monitor and optimise the quality of NICE?
The Centre’s MT team continuously analyses the MT production data to quantify and assess the modifications made by post-editors by comparing the edited content with the original MT output. They also monitor the impact of technical improvements over time (e.g. retraining of MT engines or the implementation of advanced features).
In addition, the Centre’s linguists provide valuable feedback to the MT team on the performance of the MT engines based on their post‑editing work, thus contributing to the creation of a more robust and customised system over time.
Finally, feedback on MT output received from our clients is also taken into account to train the engines.
7. Can authors have an impact on the quality of MT output?
Nowadays, documents drafted in a multilingual environment are more likely to be translated by a machine. Authors should be aware of this trend and know how to draft ‘machine-friendly’ texts.
Generally speaking, one of the first tips to bear in mind is to follow the principles of clear writing, which help authors produce clearer and more effective messages. In addition, authors need to be aware of some of the pitfalls that can ‘confuse’ modern machine translation systems. These can be related to language, content or layout.
The booklet ‘Writing for machine translation’, available on the Centre’s website, shows how the source text can affect the quality of machine translation output through genuine examples in various EU languages. It shows, for example, what happens when sentences are overly complex or lack important context, when terminology is inconsistent or when the punctuation is not correct. We believe that this booklet will be useful to anyone drafting content intended for machine translation.