Graph Refinement for Coreference Resolution

Findings of the Association for Computational Linguistics: ACL 2022, pages 2732 - 2742 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Graph Refinement for Coreference Resolution Lesly Miculicich ∗ Microsoft leslym@microsoft.com James Henderson Idiap Research Institute james.henderson@idiap.ch Abstract The state-of-the-art models for coreference res- olution are based on independent mention pair- wise decisions. We propose a modelling ap- proach that learns coreference at the document- level and takes global decisions. For this pur- pose, we model coreference links in a graph structure where the nodes are tokens in the text, and the edges represent the relationship be- tween them. Our model predicts the graph in a non-autoregressive manner, then iteratively re- fines it based on previous predictions, allowing global dependencies between decisions. The experimental results show improvements over various baselines, reinforcing the hypothesis that document-level information improves con- ference resolution. 1 Introduction Current state-of-the-art (SOTA) solutions for coref- erence resolution such as (Toshniwal et al., 2020; Xu and Choi, 2020; Wu et al., 2020) formulate the problem in an end-to-end manner where the models jointly learn to detect mentions and link coreferent mentions. The objective is to predict the antecedent of each mention-span in a document, so the model performs pair-wise decisions of all mentions. After having the model predictions, related mentions are grouped into clusters. Under this scenario, each decision (i.e., whether two mentions are related to the same entity or not) is independent. Lee et al. (2018) proposed an iterative method to update the representation of a mention with information of its probable antecedents. However, the final decisions are still made locally. We propose a modeling approach that learns coreference at the document-level and takes global decisions. We propose to model mentions and coreference links in a graph structure where the nodes are tokens in the text, and the edges represent the relationships between them. Figures 1 and 2 ∗ Work done as a PhD student at EPFL/Idiap Figure 1: Example of a graph structure for coreference. Mention spans are shown in bold, and colors represent entity clusters. The mention heads are underlined. Figure 2: Example of a graph in matrix representation. The connection types are encoded as, 0: no links, 1: mention links, 2: coreference links. show a short example taken from the CoNLL 2012 dataset (Pradhan et al., 2012) showing the graph in two perspectives. Figure 1 shows how the token nodes in a text are connected with edges drawn with arrows. We differentiate the connections be- tween words in a coreference mention, ‘mention links’, and the ones among mentions in a cluster, ‘coreference links’ (see Sec. 4). Figure 2 shows the same graph in a matrix representation, where the number in a cell indicates the type of relation be- tween the row and the column. Our model receives a document as input then predicts and iteratively refines the graph of mentions and coreference links. We follow a similar approach to the Graph-to- Graph Transformer (G2GT) proposed in (Moham- madshahi and Henderson, 2021, 2020) for syntactic …

Content truncated to 500 words