I often feel like machine translation gets a bad press. Early attempts at machine translation, with first used rules-based translation and then statistical translation, certainly had their issues. However, neural machine translation, which is one of the latest trends in machine translation, has taken this field of work to the next level. Using machine learning for translation has elevated both the quality of the documents produced and the time is taken to produce them.
In light of this, I wanted to take a quick dive into what neural machine translation is, how it works and what it means for professional translators.
It’s possible to talk in huge depth about how neural machine translation (NMT) works. Indeed, people have written entire books on the subject. However, in simple terms, NMT works by using neural networks (an interconnected set of algorithms, based on the workings of the human brain, that recognize patterns) to translate text from one language to another. It uses machine learning, with an encoder network and a decoder network, to tackle translation by looking at entire sentences at once (rather than individual words or phrases).
Neural networks can deliver outstanding translation results compared to other forms of machine translation. This form of machine learning for translation results in faster, higher quality results, which is why the likes of Google Translate and Baidu Translate have embraced it so fully.
Neural translation of this nature is driven by mathematics. Bidirectional recurrent neural networks are used to convert the language to numbers and back again. An end-to-end machine learning translation approach helps to drive up the accuracy of the results.
Neural network translation is a very different approach to former attempts at machine translation. Up until the 1990s, rules-based machine translation was the norm, until statistical machine translation became popular.
Each of these machine translation models – rules-based and statistical – tackled the translation of individual words and phrases. They took a very literal translation approach, which did little to allow for the subtleties and ambiguities that are such a core part of the way we use language.
Neural machine translation is different. The language translation machine learning model can look at much larger chunks of text when translating. In addition, all parts of the model are trained jointly. This end-to-end machine learning results in a far more contextualized form of translation than previous attempts at machine translation were able to deliver. It is why NMT so quickly became the preferred model of leaders in the machine translation field.
There are practical benefits to using neural translation as well. It requires substantially less memory than statistical translation uses. It also needs fewer engineering and design choices. All this means a slicker, more cost-effective process for those designing neural network translation models, compared to those providing statistical translation services.
Neural machine translation has a wide range of uses. Many individuals looking for a quick translation of the occasional word of phrase rely on Google Translate and the like. Many businesses also use machine translation to connect with overseas customers, suppliers, retailers, and more. This trend of using NMT for business purposes means that working with the output of machine translation is nothing new to many translators.
While some businesses are happy to use the output of machine translation without further work, those that want to assure the accuracy of their content are opting to use machine translation post-editing (MTPE). This is where human linguists work with the machine translation results to ensure that the text flows accurately. Part of the process involves learning machine translation nuances and quirks to look out for.
As the quality of neural machine translation continues to advance, so too do its uses. Businesses are relying on it to translate everything from individual emails and blog posts to entire websites. If the goal is simply to aid understanding, they will often use the raw NMT output. If more finesse is required, that’s where machine translation post-editing comes into its own.
Of course, not every document is suited to neural translation. Specialist medical or legal documents, for example, will likely fare better in the hands of a human translator with appropriate subject matter expertise. The same is true of marketing copy, where slight shifts of nuance can deliver markedly different results.
Likewise, not all languages are suited to neural machine translation. NMT companies have tended to focus on the world’s most commonly spoken tongues. That said, as the machine learning translation model matures, further languages are being added. In early 2022, for example, Google Translate announced the addition of a further 24 languages, focused mainly on those from Africa, Asia, and the Americas.
What is neural machine translation able to deliver in terms of benefits? Plenty, actually. That’s why it has seen such growth around the globe in recent years.
As I mentioned above, NMT delivers far more accurate results than other forms of machine translation. If you want to understand a document and you don’t speak the language, using a neural translation service can usually provide you with a decent enough translation of it to deliver sufficient understanding.
Using neural machine translation can also prove to be very cost-effective. Services such as Google Translate are free to use. As such, if the required outcome is simply a basic understanding of a document, the cost of professional human translation can be eliminated.
Even where machine translation post-editing is required, using NMT can still be cost-effective, with the cost of the NMT plus the MTPE often still being less than if a human translator was used for the entire process.
For large projects, neural network translation delivers the benefit of scalability. With human translators, the translation process is limited by the time that it takes the translator to work on the document. The same is true with NMT, but the difference is that NMT can deliver results in seconds. As such, a business with 100 documents can scale its translation efforts far more easily using NMT than it can using a single translator (or a team of translators, all of whom have to be found and onboarded to the project).
Neural machine translation also delivers plenty of flexibility when it comes to overcoming language barriers. Because it is fast and delivers sufficient accuracy to aid understanding, businesses can use it for everything from translating a quick social media message or email to the translation of entire brochures – again, nearly instantaneously. In the commercial world, being able to move fast can make the difference between winning and losing, and businesses have been quick to understand this when it comes to the benefits of neural machine translation.
Neural machine translation has made instant, cost-effective translation widely accessible. It can deliver excellent results (depending on the content being translated). When combined with machine translation post-editing, NMT can deliver both cost and speed efficiencies compared to traditional human translation. As such, NMT is an important part of many businesses’ translation strategies – and will continue to be over the months and years ahead.