Incremental Centering and Center Ambiguity

of verbs in Japanese, given that the number of candidate antecedents exceeds that of the zero anaphors. Their use of functional notions, viz. topicality and empathy, naturally extend the role preferences of the underlying centering model, but unlike our case, no structural ambiguities are involved. In contrast, our approach applies to the representation of both types of ambiguities. We motivated the treatment of (structural) ambiguities within the centering framework as a consequence of assuming an incremental mode of anaphor resolution, a topic that has not been raised in the centering literature so far. This is surprising insofar as even psycholinguistic studies on centering (Gordon et al., 1993; Brennan, 1995) do not touch upon this issue, though the immediacy of anaphor resolution is a common theme in cognitive text processing studies (Just & Carpenter, 1987; Sanford & Garrod, 1989). Our proposal, based on a dependency-style grammar model (Hahn et al., 1994), claims to integrate both the sentence-level as well as text-level of anaphora analysis. Furthermore, it is also fully integrated with terminological reasoning facilities as needed for in-depth text understanding , and is based on an incremental, single-pass procedure. Thus, it is superior to the work on binding theory as developed within the GB framework (Chomsky, 1981) that is restricted to the sentence-level of analysis; just recently, however , Merlo (1993) has proposed an incremental procedure for computing intrasentential coreferences based on binding theory constraints. Also Haddock (1987) considers an incre-mental mode of anaphora resolution which boils down to a variable binding, i.e., a constraint satisfaction problem in the context of a Combinatory Categorial Grammar. Any of these approaches neglects the important aspect of a preference scaling for properly selecting among several candidate discourse units as antecedents. This drawback in the same way applies to the framework of DRT (Kamp & Reyle, 1993), which is also non-incremental. Conclusions Our approach to anaphora resolution extends the original centering model by embedding the centering approach into an incremental, single-pass processing model, by providing data structures for the centering algorithm which allow for the treatment of local and global (parsing) ambiguities, and by homogeneously integrating the resolution of sentence-level (intrasentential) as well as text-level (intersentential) anaphora based on the strict requirements set up by the binding criteria (adapted to a dependency grammar framework). The anaphora resolution module has been realized as part of a dependency parser for the German language. The parser has been …