Detecting reference chains in Norwegian

This article describes a memory-based mechanism for anaphora resolution and coreference. A program labels the words in the text with part-ofspeech tags, functional roles, and lemma forms. This information is used for generating a representation of each anaphor and antecedent candidate. Each potential anaphorantecedent pair has a match vector calculated, which is used to select similar cases from a database of labeled cases. We are currently testing many feature combinations to find an optimal set for the task. The most recent results show an overall F-measure (combined precision and recall) of 62, with an F-measure of 40 for those cases where anaphor and antecedent are non-identical, and 81 for identical ones. The coreference chains are restricted so that an anaphor is only allowed to link to the last item in a chain.