One of our largest clients operates in the manufacturing sector and requires localisation of content such as user manuals and technical materials – ideal for machine translation. For a number of years, we have been building translation engines unique to their brand, using the statistical machine translation approach.
Our translation work is quality assessed by an independent party and year-on-year, we have achieved high satisfaction ratings of 90%.
Whilst our existing approach and technology brings great results, there is no advantage to standing still, especially in the translation industry where reams of content need to be translated in a short time, on lower budgets, and still be of a high quality.
In 2017, we decided to investigate the use of neural machine translation, an approach built on neural networks and deep learning, to discover if this approach could bring further efficiencies whilst maintaining quality scores.
After an extensive testing period, on 8 language pairs, using real-life data, the results were informative, yet unexpected. In summary, it isn’t a case of ‘neural is better than statistical’, nor vice versa – the choice of technology will be driven by the results gained, and we hope to continue using a blended approach to maintain client satisfaction.
This article was written by language technology specialists, Mark Unitt and Jie Jiang, and published in MultiLingual Magazine in January 2018.