Deriving rhetorical relationships from semantic content

摘要

This paper investigates to what extent rhetorical relations can be assigned purely on the basis of propositional content, without any reference to speaker goals or other pragmatic information. This task confronts any NLG system designed to generate coherent text from a set of formally represented statements; we consider it here in the context of an ontology verbaliser, for which the input is a set of axioms encoded in the web ontology language OWL. A simple set-theoretic model of the possible semantic relationships between two statements is proposed; this model allows 46 logically consistent relationships, of which we hypothesise that 11 are rhetorically coherent. This hypothesis is tested through an empirical survey which also provides some evidence on how the coherent patterns are expressed linguistically.