Over at Due Dilligence, they’ve posted a nice an overview of the current state of machine translation. Not surprisingly, it doesn’t seem to have improved much since my cognitive science days in the early 90s. As with other problems in AI, our effortless use of language makes it seem like an easy problem; whereas the truth is that we’re dealing with a problem that our brain is highly specialized to solve. In fact, that problem our brain is specialized for isn’t the obvious problem (ie communicating information) but something more subtle (e.g. reading the goals and intentions of others from their actions, being able to predict how our actions will affect the thoughts and actions of others, etc).
In chess, they’ve been able to match and even beat the best human players — not by solving the problem the way people would solve it, but by using brute force combined with specialized optimizations in various domains to trim the scope of the brute force problem down. It sounds like MT is headed in a similar direction. I have less hope this approach will succeed for MT as it has for chess. The domain of language is much more subtle and complex than chess.