Pairwise Representation Learning for Event Coreference

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 69 - 78 July 14-15, 2022 ©2022 Association for Computational Linguistics Pairwise Representation Learning for Event Coreference Xiaodong Yu1 Wenpeng Yin2 Dan Roth1 1University of Pennsylvania 2Temple University {xdyu, danroth}@seas.upenn.edu wenpeng.yin@temple.edu Abstract Natural Language Processing tasks such as re- solving the coreference of events require under- standing the relations between two text snip- pets. These tasks are typically formulated as (binary) classification problems over indepen- dently induced representations of the text snip- pets. In this work, we develop a Pairwise Rep- resentation Learning (PAIRWISERL) scheme for the event mention pairs, in which we jointly encode a pair of text snippets so that the rep- resentation of each mention in the pair is in- duced in the context of the other one. Fur- thermore, our representation supports a finer, structured representation of the text snippet to facilitate encoding events and their arguments. We show that PAIRWISERL, despite its simplic- ity, outperforms the prior state-of-the-art event coreference systems on both cross-document and within-document event coreference bench- marks. We also conduct in-depth analysis in terms of the improvement and the limitation of pairwise representation so as to provide in- sights for future work. 1 1 Introduction In this work, we study the event coreference resolution problem. Event coreference resolution is commonly modeled as a binary classification prob- lem over independently induced representations on the text snippets of each event mention (Lee et al., 2012; Barhom et al., 2019).2 Understanding the relations between two text snippets is the essential part in the tasks. In this work, we argue that the representations of prior work are not expressive enough to learn the pairwise relations due to the following two reasons: (i) Counterpart Unawareness. The relationship between two mentions can be different in different 1Our code is available at http://cogcomp.org/ page/publication_view/979 2Some work maps the two mentions into a single matching score, e.g., (Barhom et al., 2019); this can be treated as a special case of binary classification. contexts. To address different scenarios, it is better for each mention to ensure that its representation is aware of what its counterpart’s representation. However, most early work induces mention rep- resentations independently by extracting features only from the sentence that contains the mention, without using the context of the other mention (Barhom et al., 2019; Huang et al., 2019). Some more recent work attempts to encode the whole document to represent each mention (Lee et al., 2017; Cattan et al., 2020). This is beneficial for short documents, since the representation of each mention will also include information from the context of the other candidate mention. However, this is not sufficient for cross-document settings, when the comparison is, for example, between two event mentions that appear in separate documents. In this case even encoding large pieces of text leave the candidate mention representations independent of each other. (ii) Unstructured representation learning. An event mention consists of multiple arguments that describe the event: who, when, where, etc. When …

Content truncated to 500 words