An expectation maximization approach to pronoun resolution

Abstract

We propose an unsupervised Expectation Maximization approach to pronoun resolution. The system learns from a fixed list of potential antecedents for each pronoun. We show that unsupervised learning is possible in this context, as the performance of our system is comparable to supervised methods. Our results indicate that a probabilistic gender/number model, determined automatically from unlabeled text, is a powerful feature for this task.

NADA: A Robust System for Non-referential Pronoun Detection

“…1 Yang et al [33] only evaluate on "pronouns with non-empty candidate sets." Systems that do detect non-referentials as part of a fully-automatic pronoun resolution system include [24,8,7].…”

EM works for pronoun anaphora resolution

“…Probably the closest approach to our own is Cherry and Bergsma (2005), which also presents an EM approach to pronoun resolution, and obtains quite successful results. Our work improves upon theirs in several dimensions.…”

“…We evaluate all programs according to Mitkov's "resolution etiquette" scoring metric (also used in Cherry and Bergsma (2005)), which is defined as follows: if N is the number of non-anaphoric pronouns correctly identified, A the number of anaphoric pronouns correctly linked to their antecedent, and P the total number of pronouns, then a pronoun-anaphora program's percentage correct is N +A P . Most papers dealing with pronoun coreference use this simple ratio, or the variant that ignores non-anaphoric pronouns.…”

Chinese Zero Pronoun Resolution: A Joint Unsupervised Discourse-Aware Model Rivaling State-of-the-Art Resolvers

“…Motivated in part by previous work on English overt pronoun resolution (e.g., Cherry and Bergsma (2005) and Charniak and Elsner (2009)), we estimate the model parameters using the Expectation-Maximization algorithm (Dempster et al, 1977). Specifically, we use EM to iteratively (1) estimate the model parameters from data in which each ZP is labeled with the probability that it corefers with each of its candidate antecedents, and (2) apply the resulting model to re-label each ZP with the probability that it corefers with each of its candidate antecedents.…”