Here the methods to improve the efficiency of the Typed-Feature-Structure-Directed Generation, a unification-based generation mechanism which is developed for dialogue translation, are described. Unification is a time-consuming process and, in systems that use unification as their basic mechanism, most of the computing time is consumed by unification. Better algorithms for unification and/or reducing the amount of unification can improve the efficiency of such systems. We have adopted the latter approach, experimenting with several methods from both the mechanism side and the grammar side. For the mechanism, delaying surface lexical selection and eliminating disjunctive feature structures in the derivation tree can reduce the generation time up to one-third in some cases. Modification of the grammar to reduce nondeterminism is so effective that it can increase the efficiency up to 10 times.