Applying inductive decision trees in co-reference resolution of definite NPs

In this paper we present a system for resolving definite descriptions co-reference in English written texts. The system uses a set of heuristics that identifies different types of definite descriptions. The heuristics development was based on an empirical analysis of definite description use in newswire texts. We developed algorithms for the automatic analysis of several features of each definite description and then generated a list of descriptions with their features and the corresponding classification taken from a manually classified corpus. This list is given as input to an implementation of Quinlan's ID3 learning algorithm (Quinlan, 1993) and an algorithm that integrates the set of heuristics is produced automatically.