When Annotation Schemes Change Rules Help: A Configurable Approach to Coreference Resolution beyond OntoNotes

Abstract

This paper approaches the challenge of adapting coreference resolution to different coreference phenomena and mention-border definitions when there is no access to large training data in the desired target scheme. We take a configurable, rule-based approach centered on dependency syntax input, which we test by examining coreference types not covered in benchmark corpora such as OntoNotes. These include cataphora, compound modifier coreference, generic anaphors, predicate markables, i-within-i, and metonymy. We test our system, called xrenner, using different configurations on two very different datasets: Wall Street Journal material from OntoNotes and four types Wiki data from the GUM corpus. Our system compares favorably with two leading rule based and stochastic approaches in handling the different annotation formats.

Anaphora and coreference resolution: A review

“…Though most of the rule-based algorithms were knowledge rich, there were some (Baldwin, 1997;Harabagiu et al, 2001;Haghighi and Klein, 2009;Lee et al, 2013;Zeldes and Zhang, 2016) that aimed at reducing the level of dependency of rules on external knowledge. These were categorized as the "knowledge-poor algorithms".…”

“…The figure below shows the composition of different sieves used in this deterministic system. The shifting trend of CR research from rule-based systems to deep learning systems has come at the cost of loss of the ability of the CR systems to adapt to different coreference phenomenon and border definitions, when there is no access to large training data in the desired target scheme (Zeldes and Zhang, 2016). A recent rule-based algorithm (Zeldes and Zhang, 2016) also used dependency syntax as input.…”

“…The shifting trend of CR research from rule-based systems to deep learning systems has come at the cost of loss of the ability of the CR systems to adapt to different coreference phenomenon and border definitions, when there is no access to large training data in the desired target scheme (Zeldes and Zhang, 2016). A recent rule-based algorithm (Zeldes and Zhang, 2016) also used dependency syntax as input. It aimed at targeting the coreference types which were not annotated by the CoNLL 2012 shared task like cataphora, compound modifier, i-within-i, etc.…”

“…As discussed earlier the field of entity resolution is composed of different types of references. Some of these references are rare and some types are not labelled by the current CR datasets (Zeldes and Zhang, 2016). This has led to proliferation of research targeting specific types of references like multi-antecedent references (Vala et al, 2016), Abstract Anaphora (Marasović et al, 2017) and One Anaphora (Goldberg and Michaelis, 2017).…”

Coreference Resolution Using Neural MCDM and Fuzzy Weighting Technique

“…In general, coreference resolution methods are divided into rulebased methods, machine learning-based (statistical), and deep learning-based groups. In rule-based methods [21][22][23][24][25][26][27][28], a collection of rules are handwritten by experts. These rules are implemented in an orderly manner to specify co-referents in the text.…”