Neural machine translation: learning by doing

Neural machine translation (NMT) has revolutionised the language service industry over the past decade. Able to translate large volumes of text rapidly for a fraction of the price of a from-scratch human translation, NMT is the go-to choice for businesses on a budget with short deadlines.

That said, it does have its limitations: when context, language style, specialist terminology and consistency are taken into consideration, a machine translation is far less accurate than a human being, meaning it is only suitable for informal in-house use. In addition, free online NMT tools rarely offer the level of data privacy a business needs. This is why it is always recommended to have machine translations performed by an accredited language service provider.

It may sound like something from a William Gibson novel, but neural machine translation is quickly becoming an integral part of language services. NMT, as it is known for short, is the latest innovation in an industry which has seen plenty of major advancements over the past couple of decades. Having first appeared in academic articles in 2013, developers of machine translation software around the world have spent more than eight years evolving and refining NMT to make it the powerful tool it is today – not to mention a secret weapon for those who need translations fast.

But what is neural machine translation? And how does it differ from other types of machine translation?

The difference is relatively simple: the previous generation of machine translation relied on statistical probability to generate a text (hence its full name of ‘statistical machine translation’, or SMT). For example, if an expert translated the sentence “To be, or not to be, that is the question” as “Sein oder Nichtsein; das ist hier die Frage” seven times out of ten, the SMT tool would understand that sentence to be the best solution for every text. As it states in Artificial Intelligence: A Modern Approach: “This approach […] does not need handcrafted grammars of the source and target languages […] all it needs is data—sample translations from which a model can be learned”.

“This approach […] does not need handcrafted grammars of the source and target languages […] all it needs is data—sample translations from which a model can be learned”.

Sound impressive? SMT sadly comes with many drawbacks, not least the need for a huge archive of data to draw on, the difficulty in identifying and remedying specific errors, the superficiality of ‘understanding’ between the source and target languages, and the relative inability of SMT to translate into a language with different word orders.

Mimicking the human brain

By contrast, NMT uses an artificial neural network to predict the sequence of words that will appear in a block of text. Where SMT uses separate source and target models, NMT consists of a single integrated model that goes through the text word by word based on methods of deep learning and representational learning. This is cutting-edge technology which is evolving all the time, and will only become faster and more accurate as developers improve the algorithms behind the tools. Some readers, for example, may have noticed how much Google Translate (an NMT tool) has improved in recent years. This isn’t a coincidence: in 2015, researchers at the Montreal Institute of Learning Algorithms developed new AI techniques that made Google Translate a truly ‘thinking’ tool, and it has evolved from there.

The artificial intelligence that powers an NMT tool is designed to learn and expand in the same way as a human mind. Over time, this allows it to adapt to the fluid, infinitely malleable nature of human language, pick up on the nuances inherent to certain words, phrases and contexts, overcome ambiguities, and ‘fill in the blanks’ as required. If a translation memory is linked to the tool, so much the better: the algorithm can draw on specific terminology and align the text according to a certain corporate house style or industry.

Machine intelligence has its limits

Despite the advent of NMT, human translators aren’t looking for new lines of work just yet. In terms of concrete advantages, NMT has three main ones: speed, insight and scope. Able to work through large amounts of texts at an exponentially faster rate than a human being, NMT is ideal if a translation is needed at short notice. This is also useful if a text is set to become outdated shortly after it is published: if, say, multiple French authors were working on a draft of a book about startup culture and one of the authors wanted to send it to a monolingual business expert in London, they could have this draft translated via NMT to give the expert a rough overview of the project. The ease with which this can be done makes it more likely that texts that would otherwise go untranslated can be run through the software with minimal time, effort and cost required. And there is always the option of taking the machine translation and having it post-edited [link to previous article] by a professional translator later on if required.

However, the drawbacks of NMT cannot—or at least should not—be ignored. Cutting edge though the technology may be, it isn’t perfect. As stated, NMT approximates a human mind, and human minds make mistakes from time to time. It isn’t advisable to simply run a text through an NMT tool and put it out as a finished product in the world. Whether in terms of style, semantics, syntax, consistency or all four, the quality simply cannot be guaranteed. Then there is the issue of data privacy: feed a text into an online NMT and you may be putting data directly into the hands of people who won’t think twice about misusing it. It simply isn’t worth the risk.

E-discovery: exploring new legal frontiers

Some might think that the COVID-19 pandemic has put the brakes on globalisation, but an article in the Financial Times argues the opposite: “As individuals and companies move online, national borders become less relevant […] this increased digital connectivity facilitates the rapid flow of ideas, the most influential dimension of globalisation.”

As the world of business becomes increasingly interconnected, so the number of cross-border, multi-jurisdictional legal cases involving multiple actors has increased. This has led to the rise of ‘e-discovery’, or the procedure that takes place before a trial where evidence is obtained by each party requesting information in electronic format.  

This is an area where neural machine translation can be of huge benefit. When sent to a language service provider, this information can be translated into as many languages as required, regardless of volume. It is quick, accurate and secure, as the data is processed from a single, central location—especially important when dealing with evidence that could end up being presented in court. This is much more affordable than having a human translator translate all the documents (many of which may end up being of no relevance to the case). Of course, if one or more of the documents prove to be of critical importance to the legal team’s preparations, a professional human translator could then be called in to polish the NMT text, while an attorney for that legal system can verify the content.

From the DAX to the FTSE in minutes

The markets move fast. According to FIRNA, a self-regulatory organisation for the financial industry, up to 75 billion market events take place around the world each day. That is a lot of data. Much of it becomes obsolete within hours (or even minutes), and most professionals simply do not have the time to wade through it anyway. Still, if they do need to consult this information, they will want it in their mother tongue – and they require it ASAP.

This is where NMT comes into play once again: say a bank had a tool that automatically published dozens of stock reports for listed companies every day. The majority are archived without anyone setting eyes on them, but there is always a small contingent of relationship managers and clients that wishes to consult some of the reports (in French, German or Italian). An NMT tool can put these reports into their hands without delay – in the language with which they are most comfortable.

How professionally sourced NMT can help your business

The message is clear: for large volumes of data that need to be translated rapidly for internal use only, neural machine translation is the innovation the world has been waiting for. When sourced through an accredited language service provider, the process is quick, affordable and secure, and can lead to greater insights across projects and industries. Keep in mind that the end product will not be a polished text that is ready for prime time—for that, a human post-editor needs to be part of the picture.

Thankfully for our customers, SwissGlobal offers both neural machine translation and post-editing. This may be a relatively new field, but we take it seriously: from May 2021, we will be certified according to ISO 18587, the international standard for post-editing. All of our machine translation partners are carefully vetted to ensure that they, too, meet the highest standards. For all NMT projects, we work with the customer to develop a bespoke translation memory based on company and industry-specific terminology, so that every machine-translated text offers the right balance in terms of tone, consistency of language and quality.

Legal, financial, business or industry: neural machine translation could help speed up your day to day. Get in touch: we’d be pleased to show you what our NMT tools are capable of—and take you through the other quality translation and localisation services we provide.