Reference Resolution and Views Working Note 25

A common phenomenon in text understanding is to refer to an entity which was not explicitly introduced earlier in the text, but rather whose existence is indirectly implied (or suggested) from earlier facts. For example, in “He dropped a cup. The handle broke off.”, the handle referred to is not explicitly introduced, rather the reader infers that this (probably) means the handle of the cup. This working note gives a simple account for this phenomenon, based on selectively adding background knowledge to the “scene” the text is describing to make sense of it. We describe how this can be implemented in the representation language KM. 1 The Problem A common phenomenon in text understanding is to refer to an entity which was not explicitly introduced earlier in the text, but rather whose existence is indirectly implied (or suggested) from earlier facts. For example, in “He dropped a cup. The handle broke off.”, the handle referred to is not explicitly introduced, rather the reader infers that this (probably) means the handle of the cup. This working note gives a simple account for this phenomenon, based on selectively adding background knowledge to the “scene” the text is describing to make sense of it. Consider interpreting the sentence (about an airplane’s hydraulic system): “The hydraulic system supplies power to the thrust reversers.” The issue of interest here is how to resolve the reference to “the thrust reversers”. Clearly (to a person) this refers to the thrust reversers on the airplane being talked about. How can we model this recognition computationally? To understand a paragraph, we assume the computer is gradually building up a “scene” describing the paragraph’s contents, where each sentence augments (or modifies) this scene, e.g., in the style of [Zadrozny and Jensen, 1991]. In our representation language KM, a scene is represented by a set of instances and relations, stored in working memory (a KM “context” [Clark and Porter, 1999]). In this particular case, the context will already contain an instance denoting the airplane (introduced in some earlier sentence), but will not contain any instance denoting the thrust reverser. So to resolve the