| Anaphora resolution is a challenging and important problem in Natural Language Processing. We use machine learning from both labelled and unlabelled corpora to gather probabilistic information for reference resolution. Our unsupervised, unlabelled textual extraction approaches are a form of "bootstrapping" for information extraction. By assuming coreference links in unlabelled text, we can infer statistically meaningful information to assist coreference determination. This includes information on a noun's gender and number, its frequency as an antecedent, and the likelihood of coreference occurring between entities along a given syntactic relationship. These new sources of information are combined with well known constraints and preferences by inducing classifiers using supervised learning. |
