Evolution of Machine Translation
Over the past few years, Machine Translation (MT) has evolved at an enormous pace. The introduction of (Deep) Neural Machine Translation has presented the industry with ever-increasing output quality.
The emergence of new technology has compelled more suppliers of Machine Translation to become (publicly) available. Think of companies such as Amazon, Microsoft or DeepL, but also players such as Google Translate have made the switch to Neural Machine Translation. Users now have access to these various suppliers with engines in all language pairs imaginable. Translation has never been so easy: copy-paste a text or upload a file, and you can generate Machine Translation output in a matter of seconds.
On top of that, companies can now create their so-called ‘Custom Machine Translation engines’. In this process, multilingual data of a company is fed to an artificial intelligence system. That way, the MT engine is trained specifically to deal with a company’s terminology, style and domain. CrossLang can help companies in this journey. Click here for more information.
Because of these developments, our approach to Machine Translation, and more specifically the way we use Machine Translation in localization, requires a rethink.
Machine Translation Gateway
The primary challenge for anyone aiming to apply Machine Translation in localization projects, is not to use just any engine or supplier, but the one most suited for the project. The ‘right’ engine applied to a specific domain and language pair typically increases output quality, thereby lowering costs and human effort.
Finding the best fit is therefore of key importance. That is where tools such as the CrossLang Machine Translation Gateway come into play. Based upon language pair, domain and a test file, the Gateway offers users an automated advisory which selects the supplier that fits the localization project best.
Localization Productivity
When Machine Translation is used in localization projects, the most common way is to apply a post-editing step. A human translator will correct the Machine Translation output where necessary. On average, localization productivity is 50% higher compared to a process without Machine Translation.
The exact productivity increase will depend on various factors:
- MT quality. The better the output quality of the engine, the less post-editing corrections need to be made;
- Domain. Machine Translation remains more cut-out for certain fields or domains. Manuals, product information, etc. tend to be more suitable, whereas projects which require more creativity (e.g. literature or website copy) do not guarantee the same productivity increase;
- Language Pair. High-resource languages are typically better suited for Machine Translation compared to low-resource languages. The more data is available in any given language pair, the better the Machine Translation engine will become. There can also be quality differences among suppliers.
The bottom line is that it is time to rethink our approach to Machine Translation. The increase in productivity (and subsequent time and cost decrease) is too high to ignore Machine Translation any longer. The time in which Machine Translation was regarded as a threat to the translation industry has come and gone. We must start to see it more as a welcome tool to assist both businesses and translators.
