The notion that a text is coherent in virtue of the “relations” that hold between the elements of that text has become fairly common currency, both in the study of discourse coherence and in the field of text generation. The set of relations proposed in Rhetorical Structure Theory (Mann & Thompson, 1987) has had particular influence in both of these fields. But the widespread adoption of “relational” terminology belies a certain amount of confusion about the relational constructs themselves: No two theorists use exactly the same set of relations, and often there seems no motivation for introducing a new relation beyond considerations of descriptive adequacy or engineering expedience. To alleviate this confusion, it is useful to think of relations not just as constructs with descriptive or operational utility, but as constructs with psychological reality, modelling real cognitive processes in readers and writers. This conception of coherence relations suggests a methodology for delineating a set of relations to work with. Evidence that a relation is actually used by speakers of a language can be obtained by looking at the language itself—in particular by looking at the range of cue phrases the language provides for signalling relations. It is to be expected that simple methods will have evolved for signalling the relations we find most useful. This article presents a bottom-up methodology for determining a set of relations on the basis of the cue phrases that can be used to mark them in text. This methodology has the advantage of starting from concrete linguistic data, rather than from controversial assumptions about notions like “intention” and “semantics.