Statistical Identification of Pleonastic Pronouns

This paper describes an algorithm to identify pleonastic pronouns using statistical techniques. The training step uses a coreference annotated corpus of English and focuses on a set of pronouns such as it. As far as we know, there is no corpus with a pleonastic annotation. The main idea of the algorithm was then to recast the definition of pleonastic pronouns as pronouns that never occur in a coreference chain. We integrated this algorithm in an existing coreference solver (Björkelund and Nugues, 2011) and we measured the overall performance gains brought by the pleonastic it removal. We observed an improvement of 0.42 from 59.15 of the CoNLL score. The complete system (Stamborg et al., 2012) participated in the CoNLL 2012 shared task (Pradhan et al., 2012), where it obtained the 4th rank.