Representation, Coherence and Inference

Approaches to story comprehension within several fields (computational linguistics, cognitive psychology, and artificial intelligence) are compared. Central to this comparison is an overview of much recent research in cognitive psychology, which is often not incorporated into simulations of comprehension (particularly in artificial intelligence). The theoretical core of this experimental work is the establishment of coherence via inference-making. The definitions of coherence and inference-making in this paper incorporate some of this work in cognitive psychology. Three major research methodologies are examined in the light of these definitions: scripts, spreading activation, and abduction. This analysis highlights several deficiencies in current models of comprehension. One deficiency of concern is the `one-track' behaviour of current systems, which pursue a monostratal representation of each story. In contrast, this paper emphasises a view of adaptive comprehension which produces a `variable-depth' representation. A representation is pursued to the extent specified by the comprehender's goals; these goals determine the amount of coherence sought by the system, and hence the `depth' of its representation. Coherence is generated incrementally via inferences which explain the co-occurrence of story elements.