Taking a closer look at machine translation’s bad rep

Taking a closer look at machine translation’s bad rep

October 2018

This translation is so bad; it looks like you just ran it through Google Translate!” If they’re being honest, any colleague from the translation and localisation industry will admit to recognising this particular one-liner.

To be sure, it is the number one insult disgruntled clients revert to when they feel the quality of your product is lacking. And this in fact perfectly sums up the public’s general perception of machine translation (MT). But is this bad rep really justified? After all, technology in virtually every sector is moving ahead at blinding speed, so surely MT must have made considerable headway as well. It has.

MT has come a long way

The image shows a translation machine as proposed by Georges Artsrouni in the 1930s.

Artsrouni’s translation machine

Machine translation a modern invention, you say? Well, actually the first “translating machines” were conceived … in the mid-1930s. The French-Armenian Georges Artsrouni and the Russian Petr Troyanskii separately applied for patents for a mechanized bilingual dictionary even before the computer had been invented. Troyanskii additionally proposed a scheme for coding interlingual grammatical roles and an outline for a method of analysis and synthesis. However, his inventions remained unknown to the public until the end of the 1950s.

Fast-forward to the 21st century. Google Translate was launched in April 2006, using transcripts from the UN and the European Parliament to gather linguistic data. It was a statistical MT service, generating translations based on models whose parameters were derived from analysing bilingual text corpora. Note the use of the past tense in the previous sentence. Indeed, two years ago next month Google made the switch to a neural MT engine. Its Neural Machine Translation (GNMT) now relies on a vast artificial neural network that employs the method of deep learning, an advanced form of artificial intelligence.

Close but no cigar

The appearance of the word ‘neural’ in the paragraph above indicates that this form of MT no longer uses a purely statistic approach. Instead, it is inspired by how biological brains work, thus providing a much more complex contextual framework for the interpretation of the source text and the subsequent translation into the target language. For its translation service, Google now combines highly sophisticated software and AI with a multilingual archive of mindboggling proportions. The improvement is immense and many sentences – even ones that contain tricky idioms – are now neatly translated in a grammatically correct manner, while making perfect sense.

Of course, despite this significant progress, when it comes to translating, computers can still take many a cue from their naturally conceived counterparts. For example, although Google Translate can now interpret the Dutch phrase Het weekend staat voor de deur correctly and translates it as “The weekend is coming”, when you replace het weekend with de winter, the result is the literal and rather nonsensical translation “The winter is on the doorstep” (instead of the ominous “Winter is coming” which you might have hoped for, should you be partial to a certain HBO series).

dhaxley’s stance

Although we are aware of the shortcomings of MT, we do acknowledge that text type plays a crucial role. Obviously, machines will have a lot of difficulty translating prose in a way that is acceptable to discerning human readers. Imagine, if you will, entering the famous words “A rose by any other name would smell as sweet” in a free translation tool. After translating it into six different languages and then performing a back-translation into English, the final result is “The fruits added to another island are very good.” It is highly doubtful that the Bard would be pleased with this mystifying output. As you may have guessed, this example is based on real events.

On the other hand, MT can be a viable tool for tackling highly technical documents, with sentences of limited syntactical complexity. And especially if a post-editing step by a qualified human translator is incorporated in the workflow, it may be an ideal way to improve delivery times and cut costs for the client. As has been indicated above, machine translation really has evolved significantly and implementing it as a tool for professional translation services is no longer a mere science fiction scenario.

However, choosing the right MT tool is no mean feat. Why is that? Have a look at the Venn diagram below. As is often the case in life, to obtain an ideal situation, various requirements need to be met. When considering MT, three key factors spring to mind: speed, cost and quality. Unfortunately, it seems that, in the current state of things, only two are ever able to converge.The image show a Venn diagram illustrating the difficulty of choosing the right machine translation tool.

This is why dhaxley Translations is still somewhat wary when it comes to MT. Nonetheless, we realise that it has great disruptive potential in the translation and localisation sector. We therefore intend to stay up to date on the subject and make sure we can deploy machine translation tools at that time when we truly see it as an opportunity to boost our services to an even higher level.