top of page

Machine or Human Translation?

Updated: Dec 11, 2022

Translation is a craft that obviously goes back a long way in history. The Rosetta Stone, which gave scholars the first clue to how to read hieroglyphic Egyptian texts, dates from BC 196, but Sumerian and Akkadian, considerably different languages that required translation between them, coexisted in Mesopotamia as much as 2000 to 3000 years before that, and translators may have been hard at work before then on languages that have been lost to history.

Today, translators do not scratch on clay tablets or scribble on parchment, of course (no one I know of, at least). Those materials have been replaced by the latest, shiniest computer technology – so-called “machine translation,” which is making more and more use of artificial intelligence.

Most of us are aware that artificial intelligence–which I would prefer to write as artificial “intelligence”–is currently being used to write prose and poetry, paint pictures, and compose pieces of music that are claimed to rival human works in quality. At least, art critics at the Kansas State Fair, among others, seem to think so. Some people fear that writers, artists, and musicians, among many other people, are in danger of soon losing their jobs en masse. But will that be true of translators in the near future?

Most translators currently think that, while computers are useful aids to translation in a number of ways, mass unemployment is not about to hit this profession. Why is this so?

As a translator who began work in the typewriter and snail-mail era, I have watched the development of translation-related technology since then, and have noticed that much of the work that we used to be asked to do has fallen off because computer-generated translations have become acceptable to many translation customers, even though the most accurate and well-written translations still need the human touch. There is a new specialty in the field: “post-editing,” which means revising machine translations — or what I think should be called machine “translations” — to correct their imperfections.

Tourists are using translation apps on their smartphones, and I suspect that this has completely destroyed the market for traditional phrasebooks. It is true that machine translation software does have a number of advantages over human-produced translation:

* It is much less expensive and faster for translation consumers. Once they buy a machine translation system, or hire someone to use it, very lengthy documents can be spit out at little to no expense, and in minutes or even fractions of a second.

* Machine translation systems can be integrated into other software, so that any documents that have been produced by statistical computer systems, for example, can be translated immediately into any number of languages automatically or at the touch of a finger.

* Machine translation systems can learn - or rather, “learn” in machine fashion - huge masses of vocabulary very quickly by roaming the internet or other sources of translations that have already been produced, so that they can be adapted to specialize in particular scientific fields, industries, sports and entertainment areas, and so on. (We humans know very well, of course, that "roaming the internet" will usually turn up a lot of trash unless we are very careful, but how could machines know that?)

The earliest machine translation systems, using the pioneering post-WW II computers, which were much slower and had much smaller memories that the ones we use today, were much less useful. The story of an early Defense Department experimental translation system which translated “time flies like an arrow” into a Russian sentence that meant something like “time flies, a species of insect, love to eat arrows” is well-known to collectors of machine translation lore. But those primitive translating systems used what is known as “rule-based translation,” in which grammatical rules and conventional dictionaries were laboriously entered by hand into the computers of that era. The “neural” systems which are now used more and more are supposedly designed to imitate the structure of the human brain (although brain researchers actually still understand extremely little about the brain). This is why enthusiastic champions of machine translation are convinced that flesh-and-blood translators will soon become extinct.

Why are human translators still in the game? Partly in order to do post-editing. But translations are still being done by people from scratch. I think that the main reason for this is that the activity of communicating by language is based on linguistic meaning, which computers simply cannot handle. Machine “learning” allows them to store vast collections of words in the form of strings of strings of characters making up phrases and sentences. They “learn” to analyze the texts which have been written by humans to digest the ways in which these strings are generated in everyday life, and thus they can often produce “translations” that more or less closely approximate what human translators would do.

But not being humans themselves, they do not live human lives, and cannot use language with an awareness of the meanings it conveys. It cannot be emphasized too often that using language is a human trait that arose, in some way that we still do not understand, a very long time ago, and while almost all of us become skilled speakers within a few years, computers need to ingest enormous amounts of materials - many millions or billions of words - from the internet or elsewhere to accomplish their machine learning. Anyone who has had children, or spent time with them, knows that they learn to speak much more quickly, with many fewer examples of speech to work with, than that, and can go on to produce original sentences they have never heard.

A computer aid to translation I often use, which operates with artificial intelligence, doesn’t get caught in the “time flies find arrows delicious” sort of trap. It almost instantly comes up with a Japanese sentence which means something like “time flows, or streams, the way an arrow does.” It doesn’t bother with any species of fly at all. But quite often it still comes up with nonsense, because grasping the meanings of the words is entirely out of its grasp.

AI, coupled with enormous stores of already translated words gleaned from the internet and other sources, can often fabricate pieces of prose (and poetry?) that are good enough for purposes such as giving a rough sense of what a text says to someone who doesn’t know its original language, or churning through massive quantities of materials, such as the series of documents that lawyers need to review to pick out the ones that need accurate translation. But these machine translations tend to include grammatical and other kinds of errors, and expressions which may be all right in some contexts but completely miss the tone and meaning of the text to which the machine translation system was applied. The AI-based machine translator I sometimes use has a habit of omitting words or whole sentences or, on the other hand, duplicating them. (Sometimes the mechanism spits out several alternative translations, presumably because it can’t understand the source text but wants to be helpful anyway). When fed very complicated texts, such as patents and other legal documents, it very often gets completely confused by the sentence structures (especially when going from Japanese to English, which have very different grammars).

Post-editing obviously works against the cheapness and rapidity of pure machine translation, since humans, who work more slowly than silicon chips, have to be hired and paid to do it. Sometimes post-editing is not considered necessary, because the person requesting the translation just wants a rough idea of what the original text said. And people who need to order translations because they don’t know the source languages at all obviously cannot compare the translations to the originals and are unaware of when a translation, whether done only by machine or post-edited, has small or serious defects. (Of course, this is also true of translations created the old-fashioned way by the best human brains, which can make mistakes.)

Many observers of the translation industry believe that before long, the only human workers needed in it will be post-editors, and as I said above, the jobs given to human translators are dropping off, at least in certain fields. But until machine translation reaches the near-human level of quality that its most enthusiastic fans are already claiming for it, we humans will not be completely thrown out into the pasture.

A good reference on artificial intelligence in machine translation:

Melanie Mitchell: Artificial Intelligence: A Guide for Thinking Humans (New York: Farrar, Straus and Giroux, 2019), Chapter 12

24 views0 comments

Recent Posts

See All

Looking at examples I have encountered of what these new developments, more technically called "large language models (LLMs)," have achieved so far, I must say that they are still far from replacing u

When many Americans descended on Japan after World War II to temporarily govern the country, many of them found it an attractive place to settle down for a number of years, or even for the rest of the

Although the computer is still far from using natural languages like the naturals (i.e., us Homo sapiens), the stumbles it falls into can give us some interesting perspective on why translators and ev

bottom of page