Coreference Resolution with Entity Equalization

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 673–677 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 673 Coreference Resolution with Entity Equalization Ben Kantor Tel Aviv University benkantor@mail.tau.ac.il Amir Globerson Tel Aviv University amir.globerson@gmail.com Abstract A key challenge in coreference resolution is to capture properties of entity clusters, and use those in the resolution process. Here we provide a simple and effective approach for achieving this, via an “Entity Equalization” mechanism. The Equalization approach rep- resents each mention in a cluster via an ap- proximation of the sum of all mentions in the cluster. We show how this can be done in a fully differentiable end-to-end manner, thus enabling high-order inferences in the resolu- tion process. Our approach, which also em- ploys BERT embeddings, results in new state- of-the-art results on the CoNLL-2012 corefer- ence resolution task, improving average F1 by 3.6%.1 1 Introduction Coreference resolution is the task of grouping mentions into entities. A key challenge in this task is that information about an entity is spread across multiple mentions. Thus, deciding whether to as- sign a given mention to a candidate entity could require entity-level information that needs to be aggregated from all mentions. Most coreference resolution systems rely on pairwise scoring of entity mentions (Clark and Manning, 2016; Lee et al., 2017; Denis and Baldridge, 2008; Rahman and Ng, 2009; Durrett et al., 2013; Chang et al., 2013; Wiseman et al., 2016; Martschat and Strube, 2015). As such they are prone to missing global entity information. The problem of entity-level representation (also referred to as high-order coreference models) has attracted considerable interest recently, with meth- ods ranging from imitation learning (Clark and Manning, 2015) to iterative refinement (Lee et al., 2018). Specifically, Lee et al. (2018) tackled this 1Our code is publicly available at https://github. com/kkjawz/coref-ee problem by iteratively averaging the antecedents of each mention to create mention representations that are more “global” (i.e., reflect information about the entity to which the mention refers). Here we propose an approach that provides an entity-level representation in a simple and intuitive manner, and also facilitates end-to-end optimiza- tion. Our “Entity Equalization” approach posits that each entity should be represented via the sum of its corresponding mention representations. It is not immediately obvious how to perform this equalization, which relies on the entity-to-mention mapping, but we provide a natural smoothed rep- resentation of this mapping, and demonstrate how to use it for equalization. Now that each mention contains information about all its corresponding entities, we can use a standard pairwise scoring model, and this model will be able to use global entity-level information. Similar to recent coreference models, our ap- proach uses contextual embeddings as input men- tion representations. While previous approaches employed the ELMo model (Lee et al., 2018), we propose to use BERT embeddings (Devlin et al., 2018), motivated by the impressive empirical per- formance of BERT on other tasks. It is challenging to apply BERT to …

Content truncated to 500 words