A Bayesian Network for the resolution of non-anaphoric pronoun it

The Bayesian Network (BN) formalism has been scarcely used to model NLP problems. However this formalism has real advantages to overcome the limits of traditional knowledge-based approaches in NLP. Since anaphora resolution is a traditional problem in NLP, we propose a system based on a BN for the recognition of non-anaphoric occurences of the pronoun it and show that this formalism takes advantage of existing knowledge, how heterogeneous, incomplete and noisy it may be. The identification is carried out on a corpus from the genomic domain. The first results are very encouraging compared with previous systems.