In the past few years, neural machine translation (NMT) has advanced by leaps and bounds, enough so that MT with or without post-editing has become a viable, cost-efficient alternative to human-translated texts.

Next-level translation technology

Digital technologies have always been central to our business. For example, we use Computer Assisted Translation (CAT) tools to maintain translation memories. Translation memories are databases of past translations used to isolate changes and ensure consistency. Each sentence of a new source document is compared against prior translations. 100% matches require less attention from the translator and almost-perfect matches alert the translator that a small but significant change has been made. Segments without existing matches need to be translated in full, and these become part of the memory for the next iteration.

Do you need large-volume, fast-turnaround technical translation?
Find out how we can help.

Machine translation based on neural processing can do this and much more. In addition to identifying matches with previously translated texts, it uses deep learning methods to translate completely new source materials. To succeed, an NMT system must be trained for a particular language pair and subject matter. In early stages of training, the output needs to be reviewed closely by professional translators who can check the source against the target and feed the corrections back into the system. This process requires an investment of time and resources.

Even after the system has been trained, post-editing is always necessary to ensure accuracy.

What impacts MT quality?

The accuracy of machine translation output depends on several factors. The first is the quality of the source. Unambiguous, plainly written, and terminologically consistent source texts are best suited for translation by an automated system. Using standardized style guides, term glossaries, and controlled language such as simplified English will result in better translations.

In addition, NMT works better for some language pairs than for others.  Kamusi Project International published research on neural machine translations out of English into 100 target languages. Afrikaans, German, Portuguese, Spanish, and Polish ranked highest for accuracy, with scores over 70 on a scale of 1-100. The top score for accuracy in a high-demand target language, however, was only 80 (for German), confirming that post-editing is essential when accuracy is required. The worst scorers included Southeast Asian languages and several Indian languages, with Bengali at the bottom with an accuracy score of zero. The project also measured the fluency of the translations, but the top fluency score out of 100 (after Afrikaans) was German with a score of 60 on a scale of 1-100.  Latin American Spanish scored 55 for fluency and French scored 45. These translation outputs “read” like machine translations, but could be (mostly) understood.

That said, one shouldn’t be surprised if the language of a technical manual is wooden and somewhat artificial-sounding. Theoretically, NMT should produce the best results when trained to operate within a particular industry, with standardized term bases and controlled language source texts. However, in this situation, the opposite problem arises: if a technical translation is “too” good, inaccuracies can be masked by fluency, making it more difficult for post-editors to catch a small error that can make a big difference.

What is MT good for?

In order to decide whether MT makes sense for any particular purpose, take these factors into account:

How large is the volume?  If you need millions of words translated in a short period of time, human translation will not be feasible.

Is the material repetitive and frequently updated?  Technical manuals and software documentation are candidates for MT; one-off publications are not.

Is the content durable or is it ephemeral?  Customer feedback, emails, knowledge bases, FAQs, and reviews see a lot of “churn:” they are constantly being updated or replaced. In addition, the quality expectations for user-generated content are much lower than for professionally published content. The quality of the translation also matters less if the translations are to be used for in-house review or research.

For these types of content, the ROI of machine translation may be higher than human translation.

What are the drawbacks of MT?

No matter how good it gets, MT-generated content can’t match professional translation in its ability to persuade, reassure, or delight the reader. Marketing materials, websites, and any other customer-facing assets need to be translated by professionally trained, native-language translators with subject matter knowledge. MT might be used for technical instructions, but when absolute accuracy is crucial, these also require thorough post-editing. If the content is complex, and the MT system is untrained, post-editing can be as labor-intensive (and expensive) as straight translation. What took only 15 minutes to “translate” could take an additional 48 hours of labor to correct.

What about SEO penalties for auto-generated content?

In 2016, Google confirmed that unedited machine translation, even when generated by its own online tool, would be considered auto-generated content, impacting site ranking accordingly. However, when a website is translated “on the fly” either when the user clicks “translate this page” or through a widget installed on the site, there is no penalty because the page is only ranking for its English content anyway. A foreign-language domain, subdomain, or subdirectory created using human translation or at least intensive post-editing and sound international SEO is the only way to rank on foreign-language search engines.

Should I use Neural Machine Translation?

If you are interested in using machine translation, you should do so in consultation with language partners familiar with recent developments with neural and an appreciation for the efforts it takes to customize an MT engine for a given situation. In some ways, neural machine translation has much in common with the cost-saving translation-management technologies we’ve been using for years. As the demand for translated content increases across the globe, NMT is an exciting development. At this point, however, human translators and editors are still necessary for most business documents, and MT should be used only in particular circumstances, for particular types of documents.

Can your solution help users worldwide?
Ask us about localization.

Related posts: