Text and knowledge mining for coreference resolution

Evaluating Hybrid Versus Data-Driven Coreference Resolution

“…Others such as Ng and Cardie [7] and Harabagiu et al [10] also try to filter out less important or very easy positive instances to force the learning algorithm to specialize on the more difficult cases. Ng and Cardie [7] propose both negative sample selection (the reduction of the number of negative instances) and positive sample selection (the reduction of the number of positive instances), both under-sampling strategies aiming to create a better coreference resolution system.…”

“…Ng and Cardie [7] propose both negative sample selection (the reduction of the number of negative instances) and positive sample selection (the reduction of the number of positive instances), both under-sampling strategies aiming to create a better coreference resolution system. Given the observation that one antecedent is sufficient to resolve an anaphor, they present a corpus-based method for the selection of easy positive instances, which is inspired by the example selection algorithm introduced in [10]. The assumption is that the easiest types of coreference relationships to resolve are the ones that occur with high frequencies in the training data.…”

“…The assumption is that the easiest types of coreference relationships to resolve are the ones that occur with high frequencies in the training data. Harabagiu et al [10] mine by hand three sets of coreference rules for covering positive instances from the training data by finding the coreference knowledge satisfied by the largest number of anaphor-antecedent pairs. The high confidence coreference rules, for example, look for (i) repetitions of the same expression, (ii) appositions or arguments of the same copulative verb, (iii) name alias recognitions, (iv) anaphors and antecedents having the same head.…”

“…Whenever the conditions for a rule are satisfied, an antecedent for the anaphor is identified and all other pairs involving the same anaphor can be filtered out. Ng and Cardie [7] write an automatic positive sample selection algorithm that coarsely mimics the [10] algorithm by finding a confident antecedent for each anaphor. They show that system performance improves dramatically with positive sample selection.…”

“…In order to cope with these class imbalances, different instance selection techniques have been proposed to rebalance the corpus [8,9,7,[10][11][12]. In [12], rebalancing is done without any a priori linguistic knowledge about the class to be solved.…”