Whilst Machine Translation has usually been seen as a tool with great potential but poor results, the perception has now evolved since the revolution of Neural Machine Translation in 2015. The new NMT models and further developments have produced great results in certain language pairs and are now used as another translator productivity tool in the language services industry. Research and experiments have shown that translators were able to increase their daily word output by working with machine translated content for suitable texts which had a direct impact on capacity, prices and even global trade.
This has changed the way translators work and we can see 2 new work models emerge: Machine Translation + Post Editing (MTPE) and Interactive Translation Prediction (ITP). These are 2 ways in which we can see human-machine collaboration on producing translated content.
Post-Edited Machine Translation
MTPE is currently the most widely adopted model within the localisation industry. In the MTPE model, the translator works with content that is pre-translated and pre-populated by machine translation whilst also looking at the source text. He can then go through the machine translated content to review, correct errors, makes necessary edits, adapt content to the target market or local culture. In brief, anything that the machine cannot do and that needs human input.
Example of PEMT model
More and more translators are required to carry out post-editing jobs and post-editing or enhancing MT content is a skill of its own. Whilst MTPE is a great productivity boost for translators it can also have the reverse effect. Machine Translation does not perform the same for all language pairs and content types and sometimes it is faster for the translator to translate from scratch rather using the MTPE model.
We do not recommend MTPE for marketing content nor any type of creative translation.
Interactive Translation Prediction
ITP is another model that is less spoken about than MTPE but that has proven to increase translator’s productivity as well. In this model, the machine acts as an assistant of the translator leaving him in charge of the translation when in MTPE the translator has to work with what he gets as pre-translated content.
Example of ITP model
When MTPE pre-populates the placeholders where the translator is supposed to add the translated text to be reviewed and edited, ITP is more dynamic and acts as an auto-complete feature that proposes target translations as the linguist type his translation and he can accept the proposition or writes his own translation. On top of that, ITP takes into account the translator’s choices in proposed translations, his own translations and adapts to suggest better translations for the rest of the segment.
Which model works best?
Both MTPE and ITP have their pros and cons. It depends on various factors which model is the best in increasing the translator’s productivity such as what MT engine is used, the translator’s ability to work with the machine, how much specific terminology is involved and more.
Whilst the ITP has proven to be more dynamic, it also can develop an over-reliance of the translator in the Machine Translation suggestions which can lead to mistakes. MTPE in essence focuses on reviewing and correcting mistakes leaving less margin for errors. On the other hand, the translator has potentially more control over the translation using ITP rather than working with MTPE.
Both ITP and MTPE have different effects on processing time, technical effort and final quality.