Don’t worry, we are not about to persuade you that machine translation is the work of the devil. However, what you will read here is why AI does such a good job with translation – and why the job is even better if humans are involved in the translation.
Language is so much more than just sequences of letters and words or chains of sounds and written characters. First and foremost, language has to do with culture, context, and social codes. To learn about this, we don’t even need to immerse ourselves in foreign languages with which we are unfamiliar. The English language will suffice. Have you ever made a suggestion to someone and received something like “We could look at that!” as a response? Although it might sound like the person is committed to taking a closer look at your suggestion, what the person really means is: “I’m not really interested in that.”
Machine translations and what they can do
Imagine a scenario where you failed to understand this code and then misinterpreted the person’s response. This is precisely what can happen to you if you rely exclusively on machine translation services – regardless of whether the service is free or paid. However, we are not trying to suggest that machine translation is nothing but a charade. On the contrary, we are hugely impressed by the technical capabilities and the development trajectory promised by online machine translation. So, let us dip our toes into the fascinating world of computer-generated text.
Where online translations happen
Let’s start by traveling to Iceland. Here, we find a vast server farm that runs the neuronal architecture of what is probably the best-known online translation service in this part of the world: DeepL. This farm can process 5.1 petaFLOPS of requests per second. Put another way: 5 100 000 000 000 000 operations per second. This unbelievable computing scale is made possible by neuronal networks, which represent a quantum leap in the quality of machine translations.
How AI is becoming a translator
In reality, neuronal translations represent an important breakthrough that takes them into the world of artificial intelligence (AI). Why? Because neuronal translations can learn from their own mistakes and adapt their language model to take account of even rarely used linguistic peculiarities. This is a major difference to static models. This is attributable to the fact that static models rely exclusively on probabilities. The process can be compared to music rehearsal: when a mistake is made, if the orchestra gets out of synch, or a wrong note is played, the piece is started again from the point at which the mistake was made. This method ensures that the “winner” is not simply the most common translation (which is the case in static models) but rather that linguistic idiosyncrasies or rarely used phrases or linguistic formulations can be taken into account.
Where machines create extra work for humans
There is no doubt that the use of neuronal networks is a major advance. However, numerous factors must come into play in order for machines to achieve what constitutes language and makes it useful: a mixture of precision and interpretation. In a study conducted back in 2016, the Finnish translation researcher Maarit Koponen attempted to discover how much post-editing, or human correction, machine translations require and gathered the experiences of translators for this purpose. The results indicate that the particular language pair being translated plays a factor when it comes to machine translation. If machine translations are confined to related languages such as French to Spanish, the quality of the machine translation is usually more impressive than for unrelated language pairs such as English and Estonian. In these cases, the amount of post-editing work required was significantly higher.
Don’t miss the key nuances of language
Today’s server-based translator is well able to handle straightforward translations and text types such as the agenda for a conference or information about a product. However, as mentioned about, language also has cultural relevance and its magic is revealed not just as a chain of sentences or paragraphs, but also within the structure of a text. English author Cyril Connolly once expressed it as follows: “Translating from one language to another is the most delicate of intellectual exercises; compared to translation, all other puzzles, from the bridge to crosswords, seem trivial and vulgar. To take a piece of Greek and put it in English without spilling a drop; what a nice skill!”
Using post-editing services to extract the best from the machine
So how can companies for whom accurate translations are business-critical ensure that not a single drop of language is spilled? The answer is a three-step system: selecting the right machine translation technology for the job, applying it correctly and then having human translators check and refine the result – post-editing.
This approach, therefore, combines machine translation with human experience and deep learning with profound linguistic understanding.
You may well ask at what point will machine translation be capable of producing accurate and culturally-nuanced translations without any human involvement?
Let us begin by traveling back to the year 1954. On January 8, 1954, the IT giant IBM issued a press release about the potential of machine translation. It quotes the language scholar Leon Dostert: Although he emphasized that it was not yet possible “to insert a Russian book at one end and come out with an English book at the other,” Doctor Dostert predicted that “five, perhaps three years hence, interlingual meaning conversion by electronic process in important functional areas of several languages may well be an accomplished fact.”
As we write this article in 2023, we can confirm that technological development has accelerated exponentially since the 1950s. Neuronal-based technologies have enabled enormous advances in quality.
But despite this: language is constantly changed, formed, and influenced by our everyday lives, our journeys, our exposure to and the combination of different cultures, and by real-time communication with other people. Translations therefore act as a mirror to this never-ending metamorphosis. Consequently, we feel confident that computers will not be able to simulate this network of experience in the immediate future.
