A Few Thoughts on the Inevitability of Exponential Progress of Machine Translation

MT propagandists always stress how futile it is to try to resist the progress of technology. Machine translation keeps getting better and it’s here to stay, they tell us. Change is inevitable and translators better get with the program if they know what’s good for them. Within a few years machine translation will be as good (or almost as good) as human translation and we’d better get used to the idea that most of us will become MT post-processors. Post-processing of the machine translation detritus is a tool that translators need to add to their arsenal of useful skills.

Technological progress is constant, not linear but exponential, it’s coming at us, poor little translators no matter what we do, etc., and so on. The only other option, for those stubborn enough not to join their more obedient colleagues, is to quit the profession and start doing something else.

Then they sometime show PowerPoint slides for better impact, complete with charts and curves illustrating the progress that machine translation has made over time. It all looks very scientific to most people – with the exception of translators who, unlike most people, actually understand what machine translation is about and how it works.

Technology keeps getting better, that is true. But that does not mean that a tool can replace the humans who are using this tool. If MT propagandists, who are usually sales people, were able to look at things as they are, they would have to recognize that the progress of machine translation has been only incremental, and very slight at that, in the last few years, in fact so slow that for most people, it is unrecognizable.

Especially when we are talking about translations between disparate languages such as English and Japanese or Chinese, or English and Slavic languages, where there has been very little progress.

One reason why machine translations of claims in Japanese patents on the websites of Japan Patent Office (JPO), World Intellectual Property Organization (WIPO), and European Patent Office (EPO) are nonsensical is that the machine translation software does not understand that the verb that belongs to the topic of the claim, called wadai (話題) in Japanese, which is similar to subject in European languages, but not quite the same thing, is hidden at the end of a long sentence, even if the sentence has several hundred words.

It would seem like a bug that would be relatively simple to fix – all a programmer would have to do is tell the software how to recognize a claim and that if a sentence is a claim, the software needs to look first for the verb at the end of the sentence and only then come back to the verbs preceding the last one. But so far, the MT geniuses working on the software package have not been able to figure out this simple fact and program the software correctly. How is that possible after at least 15 years of “exponential progress”? Don’t they understand how the Japanese language works at all?

In machine translations from English to Slavic languages, for example, the endings of nouns are often wrong in just about every sentence “translated” with machine translation.

Unlike in English, the system of declensions of nouns is quite complicated in Slavic languages. For example Czech has seven cases in singular and seven cases in plural with different endings for nouns. And because these different cases are combined with many different classes of nouns that are based on the grammatical gender and the type of the noun, a very complicated system is created in this manner. But it is still a finite system that can be programmed into software if you understand how the system works in the language.

I remember when I was studying Japanese in Prague in the seventies, Izuru-san, my Japanese friend from Kyoto who was studying art history, had the whole system of Czech nouns taped on the wall above his desk so that he could learn the damn system by looking at it every time when he was not sure about the correct ending. (Which inspired me to try and do the same thing with Japanese characters).

It took Izuru-san’s remarkable but relatively slow human brain about 2 years to learn the system of nouns in Czech by constantly looking at it. But for some reason, machine translation programmers have not been able to program a relatively simple finite system of noun endings in Slavic languages into their software in the last 20 years. And this even though it is a finite system of possibilities that can be clearly defined in mathematical terms that could be easily handled by software and hardware.

The fallacy of MT propagandists is that they seem to think that a tool can replace the humans who are using this tool. They don’t really seem to understand the difference between these two concepts – a tool is not the same thing as the end result of the work that can be done with this tool.

A powerful vacuum cleaner does not replace a janitor. Powerful weapons and other military equipment costing a million dollars do not replace the intelligence of a soldier on the front line (or the lack thereof). They’re just tools, tools that don’t even understand what it is they are used for.

If MT propagandists really were knowledgeable language and translation experts, and not mere propagandists and salesmen who are trying to sell us something, or ingratiate themselves to “the translation industry”, they would have to start every lecture by saying the following words: Machine translation is not translation and probably never really will be because it is just a tool, not the actual product. Real translation is based on the understanding of the meaning of the words, which is something that a software program will never be able to do. Machine translation is a very powerful tool that can be used both by non-translators and translators, within certain, very definite limits.

And that is all it is and ever will be.

But this is not what they say. At least I have never heard “an expert on machine translation” admit that the Achilles heel of machine translation is that it is not really translation and never will be because there is this thing called “meaning” in human languages that cannot be programmed into a software package.

Instead, they are trying to sell us their version of reality – namely what machine translation is supposed to be, a version that you know has nothing to do with reality if you know anything about translation.

To say that we will eliminate the boundaries between human languages with machine translation software is like saying that people will bridge over distances between us and the things we want to do by learning how to fly. But there is this thing called gravity, combined with the fact that unlike birds, we are too heavy to fly, even if we could learn how to grow wings. So we have no choice but to buy an expensive ticket, get on a plane and entrust our lives to a pilot who knows how to use a tool that can be used by humans for flying called airplane.

I find all this talk about the inevitability of constant and exponential progress kind of infantile because it ignores the very real possibility (probability, or intermittent certainty?) of something called regress.

If 150 years ago most people lived only to the age of about 60 and now they live well into their eighties (women to the age of 84 because they finally figured out how to beat patriarchy, and men only to the age of 77 because they are not as smart as women), does that mean that as a result of unrelenting, exponential progress, 150 years from now most people will live on average to the age of 200, and 300 years from now most people will be immortal?

Actually, the way things are going, it seems much more likely the result of exponential progress in everything will be that no people will be living on planet Earth within a few decades.

Let’s hope that this will not be the case. But even if people somehow do figure out how to survive the next three centuries on a planet full of stockpiles of thousands of nuclear weapons, a planet that is heating up more and more every year, it is clear to me that even 300 years from now, machine translation will still be nothing more than what it is today, namely a tool that does not replace the end product of the work that can be done with the tool, just like a vacuum cleaner does not replace a janitor, and a machine gun does not replace a soldier.

However, considering the steady, albeit incremental progress of machine translation in the last few years, it is possible that in about three centuries, the machine translation packages on the websites of the Japan Patent Office (JPO), the World Intellectual Property Organization (WIPO), and the European Patent Office (EPO) will be able to identify the verb that goes with the noun in the topic (話題) of the claims in Japanese patents and that some machine translation packages will be even able to match the right endings with the right case of nouns in Slavic languages.

After more than half a century of incremental rather than exponential progress, maybe machine translation programmers will figure out how to teach their software basic rules of Japanese and Czech grammar within the next three centuries or so. Although it does seem to be a tall order at this point, especially since even the assumption that there will be any human life left on this planet 300 years from is at this point already a somewhat unlikely proposition.