How to Deal with Wicked Anaphora? Dan Cristea 1,2 and Oana-Diana Postolache 1 1 Al.I.Cuza University, Faculty of Computer Science 2 Romanian Academy, Institute for Theoretical Computer Science This paper revises a framework (called AR-engine) capable of easily defining and operating models of anaphora resolution. The proposed engine envisages the linguistic and semantic entities involved in the cognitive process of anaphora resolution as represented in three layers: the referential expressions layer, the projected layer of referential expression’s features and the semantic layer of discourse entities. Within this framework, cases of anaphora resolution usually considered difficult to be tackled are investigated and solutions are proposed. Among them, one finds relations triggered by syntactic constraints, lemma and number disagreement, and bridging anaphora. The investigation uses a contiguous text from the belletrist register. The research is motivated by the view that interpretation of free language in modern applications, especially those related to the semantic web, requires more and more sophisticated tools. 1 Introduction Although it is generally accepted that semantic features are essential for anaphora resolution, due to the difficulty and complexity of achieving a correct semantic approach, authors of automatic systems mainly preferred to avoid the extensive use of semantic information (Lappin & Leass, 1994; Mitkov, 1997; Kameyama, 1997). It is well known that anaphora studies reveal a psychological threshold around the value of 80% precision and recall that seems to resist to any attempt to be surmounted by present systems (Mitkov, 2002). It is our belief that one of the causes for the current impasse of devising an anaphora resolution (AR) system with a very high degree of confidence should be searched also in the choice for a sub-semantic limitation. Drawn mainly on strict matching criteria, in which morphological and syntactic features are of great value, these systems disregard resolution decisions based on more subtle strategies that would allow lemma and number mismatch, gender variation, split antecedents, bridging anaphora or cataphora resolution. Moreover, types of anaphora different than strict coreference, like type/token, subset/superset, is-element-of/has-as-element, is-part-of/has-as-part, etc. often impose more complex types of decision-making, which could get down to the semantic level as well. Our study makes use of the AR framework defined by Cristea and Dima (2001), and Cristea et al. (2002a) (called AR-engine) with the aim of applying it to the treatment of cases of anaphora resolution usually considered to be