This paper describes the architecture of a sentence generation module that maps a language-neutral 'deep' representation to a language-specific sentence-semantic specification, which is then handed over to a conventional front-end generator. Lexicalization is seen as the main task in the mapping step, and the article specifically examines the role of verb semantics in the process. By separating the various kinds of knowledge involved, for related languages (such as English and German) the task of multilingual sentence generation can be treated as a variant of the monolingual paraphrasing problem.