The possibility of automatic translation has been around for a long time. In recent years machine translations have become increasingly sophisticated and popular. – But will machine translations ever be of the same quality as those of a qualified human translator? Please read our previous article onmachine translations versus human translators here.
As early as 1629 René Descartes proposed a universal language. He proposed that the same concepts in different languages would be expressed by the same symbols. This could be called the birth of the idea that one day it will be possible to convey a written text in one language in another without the aid of a human translator.
The first mention of our modern concept of machine translation occurred in 1949; the first formal research started in 1951.
Early attempts of machine translation services
Systran and Babelfish were the earliest providers of web-based translations. Starting in 1997. They were met with enthusiasm by users and derision by linguists. Many did not believe that machines would ever be able to understand and translate complex and intuitive texts as well as human translators.
Examples of ridiculously inappropriate translations abounded. By the time Google Translate came along in 2002 machine translation had advanced considerably. So that many translators now offer post‑editing services to polish machine translated texts. The quality of machine translated texts is constantly improving.
So, will human translators eventually be out of work? Will machines translate text A into text B as easily as a driverless car can take us from point A to point B?
Some people believe that it is only a matter of time until translators become obsolete. Can this be true?
1 The enormity of the task
A universal machine translation program would need a huge scope. There are some 7000 languages in the world. Even if machine translation covered only the few hundred major ones, there are major obstacles to overcome. The basic environment would not only have to be programmable with the grammatical structures, rules and vocabulary of each one of the languages involved. It would also need to contain the translation rules for translating each one of them into every other one.
In addition, it would have to contain every single word and concept in each of those languages, ranging from day-to-day conversations to nuclear physics. It would have to encompass politics, medicine, law, cutting-edge scientific research, technology and every other conceivable context. The list is sheer endless.
2 Choosing the correct register
The software program would also have to know how to recognise what register was correct in every case. This is to say what the correct level of language would be.
When would more colloquial words be right and when are formal or scientific translations needed? When should the program opt for “heart attack” and when for “acute myocardial infarction”? The rules for this can change from language to language. So it would not be enough to rely on the register of the source text.
In the case of translating from a language with little formality, such as English, into one of much more complex formality, the translator has to make very fine judgments. A human translator has years of experience to fall back on. How can a software program accumulate the same knowledge?
English addresses everyone as “you”. Many other languages have familiar and formal addresses (German and French for example). English people often address each other by their first names. Many other countries are much more formal and use Mr/Ms + name for most interactions. In a business context, hybrid forms of addresses have developed, whereby people use first names plus the formal version of “you”.
A machine translation program would have to know exactly what combinations are possible in the target language. It needs to know which combinations to use in any given type of text – informal letter, formal letter, business proposal, scientific paper, abstract, speech, presentation. It would also need to take into account the type of readership or audience. Furthermore, when it comes to letters, it would have to know the precise personal or business relationship of the writer and the intended reader. In an advertising context the software would have to have the ability to apply the correct level of formality appropriate in the target country.
3 Evolution of language: new and changing words and concepts
Language changes all the time. We now have concepts and words that didn’t exist 10 or 20 years ago. Facebook, selfies, mobiles, OMG….
Other words change their meaning and are no longer used in the old context. Once, gay meant “jolly” and “jolly” was used more often than “happy”, “lively” and “cheerful”.
Our language grows with each new product, service and each new scientific discovery. The Global Language Monitor estimates that a new English language word is created every 98 minutes. It further states that, as of 14 June 2016, the total number of English words stood at 1,035,877.3 (one wonders how they arrived at .3 of a word).
4 Evolution of language: changing grammar
It’s not only words that change, our grammar, too, is forever evolving.
It is now almost acceptable to speak of a “large amount” of people. Not many people even notice constructions such as “5 items or less” any more. Wordings that were unacceptable 20 years ago are now commonplace.
Who still bats an eyelid at “He invited my husband and I.” or “Me and my colleagues are looking forward to seeing you soon.”?
It’s now accepted practice and considered perfectly correct in many instances to use contractions. No longer are English language students taught that the written form is “do not” but the spoken form is “don’t”.
5 Evolution of language: knowing the history of terminology
To start with, machine translation tools would have to be programmed with the current meaning and usage of terms. Additionally, the programs need to “know” the complete history of terminology and its usage. Last but not least, the program would need to have the capability of recognising the era a text is from.
- It would have to recognise whether the word “gay” in a text referred to “gaily gadding about teddy bears at a picnic” or to sexual orientation.
- The software would have to know that the Internet was once, briefly, called the Information Superhighway, and it would have to know the historically correct equivalent in all other languages.
- It would have to know what a phrase like “be there or be square” once meant, or at least how to research it. And it would have to know whether there was an equivalent square satellite dish to the British Squarial in the target country of the translation, or whether the translator needs to explain the sentence as well as translate it.
- Knowing that Smart was once a word processing package competing with Word and WordPerfect for supremacy would also be a requirement.
- Finally, it would have to decide whether the text it is translating targets people who also know all of this or whether it targets readers who need further explanations. It would have to know whether it can simply translate Smart as Smart or whether it would have to say “Smart, one of the early word processing packages.”
6 Keeping the software updated
It is clear that even if it was theoretically possible to write a 100% perfect machine translation program, this would depend on constant updating. This huge task could only be accomplished by a team of programmers, translators and experts. They would spend all their working lives updating the program with language rules, grammar rules, terminology, usage standards and the translation of all of these from one language into all the others.
It’s impossible to write a program that, without any further input, can translate all topics under the sun in every detail and for every circumstance.
7 Like a driverless car without a defined road network
Machine translations are getting pretty good at translating engineering manuals where A = B throughout and where no other judgments are necessary. The equivalent to this would be the driverless car on a normal road that takes passengers safely from A to B, following all the traffic rules.
Driverless cars are relatively easy, though. There are clearly defined roads and clearly defined rules. As human error accounts for most, if not all accidents, it has always been conceivable that one day, machines would be able to do the job better than humans.
Let’s say that language equates to the roads, that it is the pathways by which meaning gets transported from A to B.
Unlike roads, language does not have clearly defined parameters. As we saw, language is a very flexible, purely human construct. It has multi-faceted meanings and emotional undertones, underpinned by so many cultural assumptions, and it is ever-changing.
Texts do not just convey information but also emotions, humor, puns and poetry. How ill machine translations ever be able to transfer these fittingly into another language?
What works in one language does not necessarily work in another.
The Fiat car “Nova” was a complete flop in Spain where “no va” means “it’s not working”.
“To use something like water” means “to use it carefully” in some languages and “to use it liberally” in others.
Or take the iconic election campaign poster “Britain isn’t working”. It relies on “not working” meaning “not running effectively” as well as “not having a job”. This dual meaning does not exist in many other languages.
How would a machine translation program decide which version to drop? How would is decide on a compromise or whether to re-write a translated term completely or use entirely different words that might convey the same underlying ideas?
Very often, the original text is a mere guideline for a translator to start again. It is the basis for finding another way of conveying the same meaning in the source language.
That means that a translation starting with source text A doesn’t have a pre-defined arrival point B, as the many possible translations of the same text show. Each chosen word has its limitations and its strengths. Often, the translation depends on how you interpret text A. This means that even your starting point isn’t always clear.
Translating is a bit like asking a car to leave from an undefined parking bay in a multi-storey car park (point A) and drive to an undefined point B on roads with a rudimentary track system but with road-signs that look like “choose 10, 15, 30 or 70 mph hour here, or be inventive if you think that’s better”.