Choosing a Set of Coherence Relations for Text Generation: A Data-Driven Approach

An active research programme in Natural Language Gener ation has grown up around the notion of coherence relations Relations are being used in a variety of roles in generation systems in planning the structure of messages and in guiding their linguistic realisation However a confusing diversity exists between systems with respect to the partic ular relations they use This paper outlines a systematic methodology for justifying a set of re lations on the basis of the cue phrases that can be used to signal them explicitly in text A large taxonomy of cue phrases has been gath ered from which a conception of relations as feature based constructs is motivated By considering small extracts from the taxonomy individ ual features of relations can then be investigated in isolation The paper concludes by noting some advantages of a feature theoretic treatment of relations for NLG as well as some problems with a reliance on cue phrases in a theory of text coherence