next up previous
Next: Why Recognize Coherence? Up: A Note on Coherence Previous: A Note on Coherence

Background

Hobbs et al. (1993; henceforth, IA) elaborated a theory of discourse in which discourse is interpreted by the pragmatic strengthening of its explicit content as provided by weighted abduction. In this approach, the interpretation of a sentence is the minimal abductive proof of its logical form. In Section 6.3 of IA and in Hobbs (1996a) an account was given of the structure of discourse in which relations between successive segments of discourse are seen as pragmatic strengthenings of the relation conveyed by the adjacency of the segments. The most common of these pragmatic strengthenings have been called ``coherence relations''.

The coherence relations are defined in terms of ``what is said'' by the segments, the assertions or summaries of the segments, or the principal information conveyed by the segments. For example, a definition of the Explanation coherence relation similar to that given in IA is1

(1)

$ cause(e_{2},e_{1}) \, \supset \,CoRel(e_{1},e_{2},e_{1}) $
If a segment which asserts the eventuality e1 is followed by a segment which asserts the eventuality e2, where e2 causes e1, then there is a coherence relation between the two segments. The predicate CoRel expresses the vague relation conveyed by the adjacency of the segments, and the predicate cause is one of its possible pragmatic strengthenings.

For sentences, the assertion is generally the proposition conveyed by the main verb or modal or a top-level adverbial, although this can be altered by high stress or, as we will see below, by other factors.

When two segments are related by a coherence relation, the concatenation of the two itself constitutes a segment, which can in turn enter into coherence relations with other segments. By this recursion, a tree-like structure can be built up for an entire discourse. But in order for this to go through, we must specify the ``assertion'' of segments of text longer than a single sentence. This is the role of the third argument of CoRel. In (1) the assertion of the composed segment is the eventuality e1. The fact that the assertion of the composed segment is identical to the assertion of one of the component segments is the formal way of specifying in this framework that the relation is ``hypotactic''. In terms of Rhetorical Structure Theory (Mann and Thompson, 1988), the first segment is the ``nucleus''. For paratactic relations, like Elaboration, the assertion of the composite arises equally from the assertions of the components.

(2)

$ imply(e_{1},e) \& imply(e_{2},e) \, \supset \,CoRel(e_{1},e_{2},e) $

If the same eventuality e is derivable from the assertions of the two component segments, there is a coherence relation between the segments and the assertion of the composite is e.

Redeker (in press) raises the question of why coherence relations are needed at all in discourse processing, since the inferences they induce can and must generally be derived by other means. She also puts forward a counterexample to the specific coherence relation rule (1). Suppose Bill calls John and says,

(3)

Ann is so happy! She just got a promotion!

This matches the Explanation pattern (1). The promotion is the cause of Ann's being so happy. But the principal message of the composite segment is the promotion, not the happiness. The assertion of the combined segment, then, is not the explanandum as Axiom (1) specifies, but the explanans.

In this note, we address these two difficulties in the context of this specific example, formalizing to the extent required to elucidate the issues involved. Axiomatizations are simplified where a more elaborate account would only obfuscate. Defeasibility has been simplified out, but could easily be added using the technique of IA, Section 4.2.

Every aspect of discourse structure we discuss is shown to have a close analogue in local pragmatics, in particular, in the problems of coreference resolution, the pragmatic strengthening of vague predicates, and the expansion of metonymies. In this note, we in turn describe the IA solution to a local pragmatics problem and then illustrate an analogous phenomenon in the interpretation of discourse.

Section 2 addresses the question of why coherence relations are needed. We illustrate how the requirements of coreference often force the same inferences that are forced by the requirements of coherence, but argue that the establishment of coherence is necessary nevertheless. Section 3 explores the issue of what is asserted by the composite segment in an explanation relation. Specifically, we show how consideration of the speaker's goals can sometimes alter the main point of a segment of discourse, by a process that is at least analogous to metonymy resolution. In Section 4, we examine some consequences of the approach to coherence sketched here.


next up previous
Next: Why Recognize Coherence? Up: A Note on Coherence Previous: A Note on Coherence
Jerry Hobbs
2000-08-23