One of the major drawbacks of modu- larized task-completion dialogue systems is that each module is trained individu- ally, which presents several challenges. For example, downstream modules are af- fected by earlier modules, and the per- formance of
