In this thesis, the generation of voice (active and passive) from the conceptual representation is described from the syntactic, semantic, and pragmatic viewpoint, especially in Korean. When compared with English and even Japanese, Korean has a more complex passive-verb system with heavy passive constraints such as verb constraints, object constraints, and meaning constraints; especially, Korean does not allow the indirect passive. Moreover, an adoption of language-independent conceptual representation as an interlingua brings about lack of surface-level information on the syntactic structure which could guide how to synthesize a target sentence.
To solve these difficulties, the following are proposed: (1) The scope of passive is extended to notional passives which are morpho-syntactically active but semantically passive, which makes it possible to cope with passive constraints on the indirect passive and others. (2) The selection of the passive form is based on the syntactic type of passive, the semantic feature of active verb, and the verb class in order to get appropriate passive verbs. (3) Information on exceptions to passivization is encoded into each lexical item so as to handle individual idiosyncracy. (4) The "autonomous" passive, one of the two semantic types of passives, is regarded as an intransitive in order to lexicalize it directly from the dictionary, not through passivization. (5) The decision on the target voice depends on both pragmatic and stylistic factors. Finally, a computational model for the voice generation under an interlingua approach is proposed, which makes it possible to overcome several passive gaps between languages.