Outside the Box: Machine translation

By Chuck Smythe

I stumbled on to this piece from a December New York Times Magazine: “The Great A.I. Awakening” [Gideon Lewis-Krause, December 14, 2016]. It seems that artificial intelligence researchers have made a Great Leap Forward in machine language translation. The article is fascinating on many levels, among them its glimpse into the Google corporate culture.
    But the main thing is: this is Big News. Machine translation of language has long been one of the really hard problems in artificial intelligence. Back in my youth, it was commonly supposed that all you would need to supply was a dictionary and a grammar for each language. When this was tried, it didn’t work at all. In one famous incident, they translated “The spirit is willing, but the flesh is weak” into Russian and back. The return was “The vodka is strong, but the meat is rotten.”
    As recently as the 1990s I tried to use the available tools to translate libretti from some Italian arias, and found that at best I might hope to get the drift. Maybe.
    The reason for this fascinated me. It seems that natural language is full of puns and manners of speaking; many things just do not mean what they literally appear to mean. As a result, an adequate translation requires understanding Language A well enough to “understand” the actual intent of a passage, and Language B well enough to understand how it would conventionally express the intended meaning. It appeared that this would not happen short of giving a computer a practical understanding of the real world.

The genii at Google took a different approach: neural networks. The idea is that the brain doesn’t start with a codified map of the world, but with many, many neurons trying to match patterns. Perhaps artificial neurons could do the same? This idea was already around from the 1940s, and had been tried many times before. Results were unsatisfactory, and the approach was abandoned.
    In fact, it became disreputable, and the Google people had to fight a lot of prejudice to try it again. They argued that it hadn’t worked because it requires a very large scale. The human brain, after all, has something like ten billion neurons. So Google gave them the resources to try it on a large scale.
    The big thing about the neural-network approach is that it doesn’t generate rules. It only notices patterns, which are not necessarily describable – that is, the computer doesn’t actually know what it is doing. And neither do you.
    The results are impressive. With this approach, “The spirit is willing, but the flesh is weak” came back from Russian as “The spirit desires, but the flesh is weak.” A few passages from the Italian were less satisfactory, but still usable. I then lifted a passage from the article:

Uno no es lo que es por lo que escribe, sino por lo que ha leído. –Jorge Luis Borges

This came back as “You are not what you write, but what you have read.” I enjoyed this, as we now have a President who brags that he doesn’t read, and doesn’t need to.
    The article goes on to speculate about the impact of artificial intelligence on jobs, even complex, highly trained jobs. It reminds me of Thomas Friedman’s recent book Thank You for Being Late, which argues that computing power has reached a level such that the rules change faster than we can learn them. This may be a case in point. Enjoy your experiments!

Copyright © 2017 by Chuck Smythe