Why I Think Universal Translators Are Impossible

A universal translator is a device, featured in science fiction, which instantly translates languages.  Some commentators think that, given improving translation technology, that people won't have to learn foreign languages in the future; when meeting a foreign language speaker, one will be able to use their mobile device to communicate with them.  I am much more skeptical.


Links

• Here is the Wikipedia article on universal translators.
• Here is an article about translation software.
• Here is Google's mobile translation app.  There is a "conversation mode"--as of now, it's only available in English-Spanish--which is designed to allow two people speaking different languages to have a verbal conversation; the app uses its voice recognition technology to pick up what each person is saying and then translates it to the other person's language.

Statistics

The naive approach to programing translation software would be to program in all of the different grammar rules and then have the program fill in the blanks with words.  That is what people have tried in the past, and it didn't work.  Language is too complex and quirky to boil down to simple formulas.  As mentioned in my second link, computers would end up spitting out a lot of mangled phrases.

What has been more successful is using statistical analysis to figure out how things most likely should be translated.  What has been done is that documents which have been translated by hand from one language to another have been analyzed to see how different phrases most commonly are translated from one language to another.

Current Translation Software Is Useful, but not Perfect

There are uses for translation software.  It gives a rough estimate of what a document in a foreign language is saying.  It is the only option for translating something virtually instantly.  When it is translating something like a menu from a restaurant, it does well because there is usually a one to one relationship between those kinds of specific nouns.

On the other hand, if one wants a perfect translation, it has to be done by a person.  I remember while on my mission in Boston seeing a notice on an apartment building which had obviously been translated via an instant computer translation.  The apartment was presumably managed by an English-speaking American, but the apartment had a lot of Portuguese-speaking Brazilian immigrants living there.  There was a notice which tried to explain why the residents should not park in front of the dumpster.  It was obviously typed out in English, inputed into a translation program, and then printed out.  I remember it being funny how many blatant errors were in the document.  One I remember was that the document said, "não parque em frente do dumpster." For those who don't speak Portuguese, here's why that phrase is wrong: first, the word "parque" means a park as in the noun for a place; it never means the verb to park, which is estacionar; second dumpster is not even a word in Portuguese; the program obviously didn't recognize the word dumpster and so it just treated it as a proper noun.  Having a sentence like that would be like saying in English, "no park in front of the hoobaloo."

Artificial Intelligence

On a deeper level, to truly understand what a phrase means requires human intelligence.  There are many factors which influence the meaning of a phrase.  

The context of a conversation can change the meaning of a sentence.  For example, if I am reading a math paper, terms like "open", "ring", "annihilate", etc. have technical meanings referring to specific mathematic principles.  One then says, "OK, perhaps you tell the program that this is a mathematical paper."  Well, there may be times in that paper where I want to use the word "open" or "ring" in a conventional sense.

In a verbal conversation, intonation can affect what a phrase means.  Think of how many ways one can change the meaning of a simple phrase like "I love you": "I love you; I love you; I love you; I love you?"

The human has little difficulty sorting out subtlety.  A computer has a ton of trouble sorting out any kind of subtlety to the point that it is currently impossible to do with any degree of success approaching that of a human.  In my view, the issue is one of the possibility or impossibility of creating artificial intelligence or not.  Can a computer be created which can handle the same subtleties that a human can?

Why I Say No and Other References

I think it is impossible.  (Disclaimer: my knowledge on this subject is very limited.)  A big reason I do is because of the book Godel, Escher, Bach.  To give a (poor and woefully incomplete) summary of what I am trying to say, mathematically it is impossible to design a complete system of instructions for a computer to decipher subtlety.  This is, at least to my incomplete understanding, why a statistical method for translating something works better than trying to program a computer to translate via rules.

That Being Said, Even Human Translators Aren't Perfect


Languages are not simply copies of each other with different words in place of other words.  There are actually concepts which some languages have which cannot be adequately expressed in other languages.  To really understand what someone is saying in another language, especially in an informal setting, one needs to know the language.

Take for example sarcasm.  The concept of sarcasm is at least in part to say something where the true meaning is intentionally obscured.  Even for a human to try to translate sarcasm is difficult.  Plus, in some cultures sarcasm doesn't really exist like it does in the United States, for example.

There are just certain things which are lost in translation.  If not even an expert human can translate them, how can I expect a computer, no matter how expertly designed, could automatically do it.