A Conceptual Reasoning Approach to Textual Ellipsis

the centering model [6] to which we adhere is not considered a semantic theory but rather a discourse processing model which lends itself to an easy integration into a text understanding framework. Still, it does not provide for well-developed methods for textual ellipsis resolution. Grosz et al. rather sketchily point to the difference between the relations directly realizes and realizes whose precise definition they suggest depends on the semantic theory one adopts [6, p.209]. We have shown, however, that there are a lot of constraints at the conceptual level which cannot reasonably be accounted for by semantic theories. Only few NLP systems exist which deal with textual ellipsis in a dedicated way. For example, the PUNDIT system [17] provides a fairly restricted solution in that only direct conceptual links between the concept denoted by the antecedentand the elliptical expression are considered (" plausible " paths of length 1, in our terminology). The approach reported in this paper also extends our own previous work on textual ellipses [7] by the incorporation of an elaborated model of functional preferences on Cf elements which constrains the set of possible antecedents according to information structure criteria. 7 Conclusions In this paper, we have outlined a model of textual ellipsis resolution. It considers conceptual criteria to be of primary importance and provides conceptual well-formedness and strength criteria for role chains in a terminological knowledge base in order to assess the plausibil-ity of various possible antecedents as proper bridges [3] to elliptical expressions. Functional constraints based on the utterances' information structure contribute further restrictions on proper elliptical antecedents and require a basic revision of the centering model. The two principled difficulties inherent in every network-based symbolic knowledge representation approach are its dependency on hand-crafted and often domain-specific knowledge and the exorbitant costs for unconstrained search (often limiting the scalability of the approach). We cope with the first of these problems by postulating two formal criteria, viz. the non-cyclicity and the inclusion condition, which only depend on features present in almost any knowledge representation language, i.e., isa links, domains and ranges of relations and inverse relations. No reference to a specific network structure or a specific domain is made. The definition of patterns relies mainly on structural properties of semantic relations, which are entirely domain-independent. However, this heuristic criterion is only effective when the underlying knowledge base is built on top of a clear taxonomy of relations …