In many cases, RST (Mann & Thomson 1988) is interpreted or used in such a way that discourse relations hold only between sentences or clauses, i.e. the leaves of the rhetorical tree structure of a text correspond to sentences or clauses. This interpretation of RST has already been criticized (Smedt et al. 1996) and my talk goes in this line. In general, a discourse relation can be realized in many different ways. In addition to traditional ways such as a text made up of two sentences with or without a cue phrase (Ted hammered the metal. (Therefore), it is flat), there exist other ways such as a single sentence, e.g. a "resultative construction" (Ted hammered the metal flat) or an "operator verb construction" Ted's hammering the metal caused it to be flat). The question arises: given the numerous and heterogeneous set of possible linguistic realizations, how can these possible realizations be expressed in a linguistically motivated manner so that a natural language processing system can make use of this knowledge? I will present a uniform framework for expressing the linguistic knowledge needed to relate discourse relations to linguistic realizations. In addition, I will sketch how such a representation can be used during the linguistic planning stage of text generation and presents a generation system. This system is based on works on the lexicon which "is now recognized as an essential component of the grammar in all major theoretical frameworks" (Stowell & Wehrli 1992, p.1). I adopt this view on the lexicon and adapt it for both generation and text issues (the linguistic works are directed towards analysis of isolated sentences with lexicalized grammars). The goal is to design a lexicalized text generation system. In such a system, lexical entries give semantic and syntactic information in a well defined structure. Therefore, the generation process for linguistic choices relies mainly upon a unique operation: lexicalization, i.e. choice of a lexical item with its semantic and syntactic structure to express a concept. Finally, I will present a formalim to write a lexicalized text generation system, namely G-TAG, which is inspired from TAG.