I would model them as feature functions over phrases. You might imagine that you can exploit vector similarity to do smoothing.
Good luck Miles
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
