Which Factors Contributes to Resolving Coreference Chains with Bayesian Networks?

International audienceThis paper describes coreference chain resolution with Bayesian Networks. Several factors in the resolution of coreference chains may greatly affect the final performance. If the choice of machine learning algorithm and the features the learner relies on are largely addressed by the community, others factors implicated in the resolution, such as noisy features, anaphoricity resolution or the search windows, have been less studied, and their importance remains unclear. In this article, we describe a mention-pair resolver using Bayesian Networks, targeting coreference resolution in discharge summaries. We present a study of the contributions of comprehensive factors involved in the resolution using the 2011 i2b2/VA challenge data set. The results of our study indicate that, besides the use of noisy features for the resolution, anaphoricity resolution has the biggest effect on the coreference chain resolution performance