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

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1645–1651 June 6–11, 2021. ©2021 Association for Computational Linguistics 1645 Evaluating the Impact of a Hierarchical Discourse Representation on Entity Coreference Resolution Performance Sopan Khosla James Fiacco Carolyn Rosé Language Technologies Institute Carnegie Mellon University, USA {sopank,jfiacco,cprose}@cs.cmu.edu 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 dis- course structure. In this work, we leverage au- tomatically constructed discourse parse trees within a neural approach and demonstrate a significant improvement on two benchmark en- tity coreference-resolution datasets. We ex- plore how the impact varies depending upon the type of mention. 1 Introduction Historically, theories of discourse coher- ence (Chafe, 1976; Hobbs, 1979; Grosz and Sidner, 1986; Clark and Brennan, 1991) have of- fered elaborate expositions on how the patterns of anaphoric references in discourse are constrained by limitations in human capacity to manage attention and resolve ambiguity. Hobbs (1979) acknowledges that these human limitations have meant that coreference resolution in natural text can be achieved with relatively high accuracy using a combination of recency and simple semantic constraints. State-of-the-art neural approaches for coreference resolution (Lee et al., 2017; Joshi et al., 2019, 2020) have therefore not surprisingly shown strong performance relying on surface-level fea- tures and local-context (i.e., extracted from a small text window around the mention). Traditional approaches, on the other hand, make an attempt to formally model the process of managing attention, for example, the stack in Grosz and Sidner (1986)’s model. Their stack-based model suggests specific places where recency might fail while a more explicit model of discourse structure might make a correct prediction, for example, where an anaphor and a nearby potential (but incorrect) antecedent are in adjacent but separate discourse segments. Because of the potential existence of such cases, we hypothesize that formally incorporating a representation of discourse structure would have a small but non-random positive impact on the ability to correctly resolve anaphoric references. This effect might vary depending upon the semantic informativeness of alternative types of anaphoric expressions, since they impose different constraints on where their antecedent can be located within a hierarchical discourse structure. There is also a danger that the level of accuracy with which the hierarchical structure of discourse can be obtained in practice might reduce the positive impact still further. The contribution of this paper is an empirical in- vestigation of the impact of including a representa- tion of the hierarchical structure of discourse within a neural entity coreference approach. To this end, we leverage a state-of-the-art RST discourse-parser to convert a flat document into a tree-like structure from which we can derive features that model the structural constraints. We embed this representa- tion within an architecture that is enabled to learn to use this information deferentially depending upon the type of mention. The results demonstrate that …

Content truncated to 500 words