Given the fact that machine translation has been studied for over 30 years now, it is surprising to find so little research on its central questions: in what ways can one implement a semantic or pragmatic feature, and how does one choose between these possibilities given a specific context? Of course, there have been many systems that have implemented one scheme or another. For the most part, these systems simplified the problem by assuming a one-to-one mapping between semantics and surface structures while effectively ignoring pragmatic considerations. To find any discussions on the subject in the past, one had to move into the field of human translation. The Summer Institute of Linguistics (SIL) has studied the problem of discourse relations and their possible implementations for over twenty years. [Larson, 1984] and [Beekman et al., 1981] will be reviewed below in relation to this research. More recently, Rhetorical Structure Theory (RST) [Mann, 1987] has become a focus of attention in the text generation community. A number of articles will be reviewed below that seek to identify the realizations of RST relations, especially in the popular field of instructional texts. Finally the work of [Hovy, 1988] in the field of pragmatics will be discussed.