Consequences for the Modelling of Coherence Relations

We have shown in this note how various general mechanisms of discourse interpretation can play a role in the recognition of coherence. In the abduction model, the availability of those general processes allows the specification of the coherence relations themselves to be minimized, thus reducing the need to encode elaborate conditions in the definitions of coherence relations and to distinguish many specific relation types or application cases. A rather limited set such as, for instance, that proposed by Hobbs (1990) would seem quite sufficient. In terms of Redeker's (1990, in press) model, this could be seen as a terse description of the rhetorical and sequential components (the pragmatic structure), with conditions formulated mostly in terms of the ideational component (the semantic structure).

To illustrate how the integration of coherence processing with more general processes may obviate the need to introduce specific subclasses of coherence relations, let us consider again our analysis of example (3) as metonymy (Figure ). The analysis still contains the Explanation relation, although the final interpretation is equivalent to a coherence relation that could be defined as follows:

(36)	$cause(e_{2},e_{1}) \, \supset \,CoRel(e_{1},e_{2},e_{2})$

where the assertion of the composite segment is e₂, not e₁ as in Axiom (1). This relation might be interpreted as a special case in the class of Hobbs's (1990) Evaluation relations with the evaluation satellite preceding the evaluandum, as in

Something wonderful happened. Ann got a promotion.

The pragmatic properties of example (3), however, are quite different from those of ordinary evaluations. The implicit presence of an Explanation relation in our analysis of (3) provides one way of describing this difference. It would not seem necessary or even useful, therefore, to introduce the special kind of relation defined in (36).

A final remark on coherence relations is in order. It is sometimes believed that to adopt a framework that includes coherence relations is to commit oneself to a claim that the hearer is consciously aware of the occurrence of the coherence relations. This is not a necessary consequence, as consciousness and interpretation are largely orthogonal issues, just as consciousness and syntax are. But the approach to coherence we have formulated in this note is not even compatible with consciousness of the specific coherence relation. There is no mention of ``explanation'' in the Explanation pattern (1). To make this explicit we would have to rewrite (1) as

(37)	$cause(e_{2},e_{1}) \, \supset \,Explanation(e_{1},e_{2})$
(38)	$Explanation(e_{1},e_{2}) \, \supset \,CoRel(e_{1},e_{2},e_{1})$

In our view, speakers are sometimes aware of the coherence relation; more often they are not. In any case, from a formal point of view, the difference is slight, whatever its importance on a more intuitive level.

In summary, we have shown that an analysis of the coherence of a text is necessary, even though much of what is discovered thereby will often have been forced from coreference considerations alone. The requirement to establish coherence is similar to analogous interpretation requirements in syntax and local pragmatics. The interpretive moves, such as metonymic interpretation, that are available in local pragmatics are also available in discourse processing, and that fact makes possible a significant economy in the specification of what constitutes a coherent discourse.