An Iterative Model for Discovering Person Coreferences Using Name Frequency Estimates

Abstract

In this paper we present an approach to person coreference in a large collection of news, based on two main hypothesis: first, coreference is an iterative process, where the easy cases are addressed first and are then made available as an incrementally enriched resource for resolving more difficult cases. Second, at each iteration coreference among two person names is established according to a probabilistic model, where a number of features (e.g. frequency of first and last names) are taken into account. The approach does not assume any prior knowledge about persons mentioned in the collection and requires basic linguistic processing (named entity recognition) and resources (a dictionary of person names). The system parameters have been estimated on a Italian news corpus of 5K and then experimented on a collection containing more than 7 millions person names. Evaluation, over a sample of four days news, shows that the error rate of the system (1.4%) is above a baseline (5.4%) for the task. Finally, we discuss open issues for evaluation.