Evaluating the Impact of a Hierarchical Discourse Representation on Entity Coreference Resolution Performance

Abstract

Recent work on entity coreference resolution (CR) follows current trends in Deep Learning applied to embeddings and relatively simple task-related features. SOTA models do not make use of hierarchical representations of discourse structure. In this work, we leverage automatically constructed discourse parse trees within a neural approach and demonstrate a significant improvement on two benchmark entity coreference-resolution datasets. We explore how the impact varies depending upon the type of mention.

Light Coreference Resolution for Russian with Hierarchical Discourse Features

“…Closest to our work are (Khosla et al, 2021) implementing various features over an RST tree produced with a parser for English. However, their main concern is how general is the lowest common ancestor of ²https://catalog.ldc.upenn.edu/LDC2002T07 Chistova E. V., Smirnov I. V.…”

“…• Rhetorical Distance (𝐷𝐷 𝑅𝑅𝑅 ) is a number of nuclear EDUs occurring between two spans in a hierarchical rhetorical tree. We also adopt a feature estimating the amount of generality required to have two mentions in the same discourse subtree (Khosla et al, 2021)…”

Incorporating Centering Theory into Neural Coreference Resolution

“…In the deep learning period, only a few researchers incorporated discourse information into a coreference model to our knowledge. Recently, Khosla et al (2021) use rhetorical structure theory (RST) (Mann and Thompson, 1988) to capture the hierarchical discourse structure of documents, from which they encode three distance features for the candidate and query mentions on different levels (i.e., word-level, discourse-unit-level and discourse subtree). Held et al (2021) apply discourse coherence (Grosz, 1977(Grosz, , 1978Grosz and Sidner, 1986) to cross-document coreference resolution.…”