Coherence relations have proved a useful tool in recent computational models of discourse. However, there is much disagreement between theorists about which set of relations to use, and also between analysts working within a particular theory as to how to analyse individual texts. These difficulties arise from the fact that present conceptions of relations lack a solid empirical basis. This paper describes a new methodology for determining a set of relations, whose starting point is a conception of relations as modelling psychological constructs. It is argued that evidence for these constructs can be sought in a study of the cue phrases that can be used to signal them in surface text.