Status Updates From Statistical Machine Transla...

Now on the last chapter, and the question we've all been waiting for: when doing machine translation, would it help if the machine actually understood the text to some extent?

Well... maybe. There is some evidence to suggest that could be true. But don't get too excited.

When it comes to evaluating machine translation systems, we tend to put most trust in the judgement of human evaluators. This method, however, has one major disadvantage. It takes a lot of time, and, if the evaluators expect to get paid, also a lot of money. The typical statistical machine translation researcher, on the other hand, has no money and would like to carry out evaluations very frequently.

A good summary of the differences between word-based and phrase-based statistical translation. I somehow wasn't aware that phrase-based translation model for handling word-order was so primitive, but it explains a lot!

He starts by considering possible English translations of the German noun Haus, ranging from "house" (very common) to "shell" (very rare), and segues from this to a definition of IBM Model 1.

He makes the interesting point that research in machine translation has historically been driven largely by American worries about military and commercial threats. In the 1950s, Russian; in the 2000s, Arabic and Chinese.

The author is a member of our new project, so I thought I'd read his book. Everyone says it's excellent.