POST-EDITING
Raw machine translation output is not "fit for consumption".
LOOKS CAN BE DECEIVING.
Don’t be misled by a text that “reads well”.
Neural machine translation engines have improved vastly in recent years, and they are now able to produce fluent syntax, even for longer sentences. It is that fluency in particular that makes it so easy for readers to "overlook" errors. Unidiomatic or “skewed” constructions are common.
This is why it is essential for post-editors to identify such inconsistencies and distortions of meaning to ensure that the target text correctly reflects the source text.
CONSISTENCY COUNTS!
The terminology whims of machine translation.
In legal and financial texts, defined terms and specialist terminology must be translated consistently. However, even well-trained engines are not able to note which term is to be used in a particular context. Words are translated in isolation, meaning that the same term may be rendered differently even within the same sentence. A prime example of this characteristic inconsistency is the following machine-generated text posted on the English website of a German major corporation: “Jane Doe is a member of the company's Managing Board. She was appointed to the Board of Management effective 1 January 2024. Her Executive Board responsibilities include HR and Compliance.“
As ubiquitous as such nonsensical translations are becoming, most would agree that they have no place in a professional environment. Raw machine translation output is not fit for consumption and must be consistently reworked by experts who know what they are doing.
LOOKING AT THE BIG PICTURE
Machine translations sometimes miss the point.
Human translators consider which wording makes sense in the context of each sentence. They infer and reflect, and draw on the wealth of their experience.
Machine translation engines lack this ability to see the big picture. They do not think for themselves, nor do they question whether the translation is plausible. In other words: They cannot fathom the reality that the language describes.
This can lead to the engine "hallucinating", i.e. freely inventing facts. Examples of such hallucinations include:
- repeated words
- information left out or added
- illogical connections
- twisted meaning
Our professional post-editors know what to look for and can help you side-step the pitfalls of machine translation.
THE “TELEPHONE EFFECT”
How important information can get “lost in translation.”
Machine translation is not always possible in direct language pairs. Because MT is based on corpora, i.e., collections of written and spoken material that the engine uses to extract results, it works better for some language pairs than it does for others, particularly for more widely-used languages. MT most often relies on English as a pivot language because the body of available data is vastly greater than with lesser-used languages. For more exotic language pairs, MT employs a process known as “relay translation”, which basically involves a chain of texts, ending with a translation made from another translation. For example, if a company has a Polish text it wants translated for its subsidiary in Italy, the Polish source text is translated first into English as the relay language and then from English into Italian.
Reminiscent of the childhood game of “telephone”, when the original message is passed on through multiple players, the meaning is often diluted or distorted when it comes out at the other end. It is quite literally “lost in translation”.
This is why the pre-translated MT output must undergo a critical review by a human post-editor to ensure that the source text is not incomplete, misleading or incorrect.