Integration

We want to plug these systems into a wide variety of MT systems, both for practical purposes and experimental ones. So the question is, how are we going to do that?

L3

For plugging in to L3, we need a really good story...

SCFGs

In any kind of SCFG system, whether it's got hand-tuned rules or a zillion extracted ones, at the end, you're likely going to have rules of the shape

[NP] ||| ich ||| i ||| LogProb=-0.6

... so the obvious thing to do here would be just to generate different rules, just before decoding. Say we think that a certain translation of a word or phrase is more likely than another one, due to the disambiguation. Well, then we just give it a higher probability, or maybe mark it with some other feature...