SDL wb MachinevsMan Blog 620x300

Translation: Machine or Human?

Over the past couple of years, the fast evolution of Machine Translation (MT) took the localization world by storm and became a sweeping force that is radically changing multilingual content creation.

Machine Translation offers tremendous advantages in addressing today’s translation volume demands on the global enterprise. Raw Machine Translation (MT) offers immediate results at temptingly low costs. But traditional, full human translation (HT) services are still available. In between these extremes is the combination of both machine output and human efforts (i.e., Post-Edited Machine Translation or PEMT). So how do you choose among these three translation processes?

A gradual but sure change

Machine Translation is here to stay. Choosing to adopt it – when and how, not “if" – is a matter of trust and expectations. Although most global enterprises recognize that MT will surely to be an essential tool in their localization bag of tricks, some are still suspicious that it can deliver the desired output. But when MT is used to its full potential, it can not only deliver exactly the results you need, but also offer significant savings – all with iron-clad security.

Matching quality to purpose

To get the best from MT, you must first evaluate the source content for its specific business purpose in order to ultimately choose the best process, including any quality assurance (QA) steps. Meeting your specific quality requirements might take zero to the most stringent QA.

For content with a short shelf life (hotel reviews that help customers make choices, user forums or knowledge base articles that lead to solutions to simple problems) may not need perfect translation, so raw MT may be the fastest and most cost-effective choice.

Creative materials, such as Marketing content, and content that requires absolute perfection and preciseness, such as binding legal documents, lie on the other side of the spectrum. These use cases will always require the involvement of human translators, be it to perform the full translation process or to perfect the MT output with post-editing.

Choosing the right process

Raw MT: Volume and Speed

Being able to use straight-up raw MT is definitely the dream. Neural Machine Translation (NMT) promises to deliver superior quality, especially in terms of fluency and natural language. When trained with high-quality, clean and diverse corpora (that is, good data is better than a lot of data), NMT systems can deliver raw translations that are adequate for:

  • Gist and basic understanding
  • Sentiment analysis
  • eDiscovery
  • Instant chat
  • Customer support portals
  • User forums
  • Knowledge base

To learn more about NMT, see this presentation by Mihai Vlad, SDL’s VP of Machine Learning  “Demystifying AI and NMT."

PEMT: The Best of Both Worlds

Post-Edited Machine Translation has been used successfully for decades with Statistical MT and Adaptive MT. Adaptive MT is especially effective in a PEMT scenario since it learns from the post-edits and applies them to subsequent translation segments.

When used in conjunction with an efficient Translation Management System (TMS), PEMT yields multilingual content that can be easily leveraged and reused for continuously growing ROI since the degree of post-editing effort required reduces over time. This proven method, therefore, not only brings the benefit of higher quality, but also has a high impact on cost and productivity.

Depending on the use case, varying degrees of PEMT can be applied.

Light PEMT

When the main goal is just to ensure accurate meaning, but fluency and style are not important, the post-editor improves the MT output by ensuring all words and phrases are translated correctly. This level of post-editing is appropriate for simpler, straight-forward materials such as manuals, for example.

In a light post-editing job, the editor:

  • Preserves the raw MT output as much as possible.
  • Focuses on semantically correct translation and accurate spelling.
  • Corrects any instances where text/concepts were added or omitted.
  • Eliminates offensive or culturally inappropriate content.


This level of post-editing requires a greater time investment, but when done right, it can achieve human-level quality while still benefiting from the productivity boost provided by a well-trained MT engine. Since the post-editor can leverage the translation suggested by the MT engine in addition to his/her own knowledge and instincts, the final translation can potentially be better than a completely human translation.

Note that use cases that require full PEMT are best handled by an editor who is also an SME. While ensuring translation accuracy, the editor:

  • Changes as much of the raw MT output as necessary.
  • Addresses style, terminology and grammar issues.
  • Strives for near-human translation results.
  • Corrects semantics, syntax and grammar.
  • Ensures terminology is used correctly and consistently.
  • Corrects any instances where text/concepts were added or omitted.
  • Eliminates offensive or culturally inappropriate content.
  • Addresses any formatting issues.

An additional and crucial benefit: since results from this process are of such high quality, post-edited content can – and should – also be used to train your MT engine further.

Human translation

Although NMT provides much more fluent translation than previous MT systems, it still has limited capabilities to handle creative, subjective content. This includes play on words, humor, irony, figures of speech, hyperbole, metaphors and any ambiguous language use, commonly found in Marketing content and literature.

Since the creative use of language breaks language patterns (and MT systems are based on patterns), only a human translator can decipher the source content to produce high-quality translation.

Warning: garbage in, garbage out

It’s important to ensure your MT engine will consistently produce the highest possible quality output to reduce post-editing efforts over time. The best results will come if everything that your system learns (i.e., training data) and processes (i.e., source material) is also of high quality. These are the essential requirements:

  • Train your system appropriately from the start by using accurate and complete dictionaries and glossaries, as well as a clean, accurate, domain-specific corpus.
  • Ensure the source text is well written and free of spelling, punctuation and ambiguity issues.
  • Use only qualified post-editors.
  • Retrain your engine with fully post-edited translated materials.

As with everything in life, it’s important to have realistic expectations of this amazing technology. The quality delivered by Machine Translation has improved exponentially in the recent past, and the truly cutting-edge MT technologies will not disappoint. For the cases where human translation is the best choice, make sure to work with an experienced and reliable ISO-certified localization partner for top-notch quality.

To find out more about SDL’s MT solutions, please visit here. If a Language Services partner is what you need, please see SDL’s full ecosystem of technologies and services.