What happens before and after: Multi-Event Commonsense in Event Coreference Resolution

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1708–1724 May 2-6, 2023 ©2023 Association for Computational Linguistics What happens before and after: Multi-Event Commonsense in Event Coreference Resolution Sahithya Ravi1,2 Chris Tanner3,4 Raymond Ng1 Vered Shwartz1,2 1 University of British Columbia 2 Vector Institute for AI 3 Massachusetts Institute of Technology 4 Kensho Technologies {sahiravi, rng, vshwartz}@cs.ubc.ca, cwt@mit.edu Abstract Event coreference models cluster event men- tions pertaining to the same real-world event. Recent models rely on contextualized represen- tations to recognize coreference among lexi- cally or contextually similar mentions. How- ever, models typically fail to leverage common- sense inferences, which is particularly limiting for resolving lexically-divergent mentions. We propose a model that extends event mentions with temporal commonsense inferences. Given a complex sentence with multiple events, e.g., “The man killed his wife and got arrested”, with the target event “arrested”, our model generates plausible events that happen before the target event – such as “the police arrived”, and af- ter it, such as “he was sentenced”. We show that incorporating such inferences into an exist- ing event coreference model improves its per- formance, and we analyze the coreferences in which such temporal knowledge is required. 1 Introduction The goal of cross-document event coreference res- olution is to determine if various event mentions (e.g. shot, gunshot), across one or more documents, refer to the same event. Existing systems represent each mention within its context using a language model (Cattan et al., 2021a; Allaway et al., 2021), and train a scorer to predict if two mentions corefer, based on their lexical and contextual similarity. While many coreferring mention pairs in event coreference datasets such as ECB+ (Cybulska and Vossen, 2014) are lexically and contextually sim- ilar, or even share the same lemma (Wolfe et al., 2015), the difficulty arises for dissimilar corefer- ring mentions. For example, in Figure 1, spent and hospitalized are coreferring. These mentions are not lexically similar, and are not often used in simi- lar contexts. In this paper, we improve the ability of existing cross-document event coreference systems to resolve such challenging coreferring mentions, by providing additional context in the form of com- monsense knowledge. We focus on two temporal commonsense relations — before and after — per- taining to typical events that happen before and after the target event. For instance, in Figure 1, we may infer that before Dalton was shot, a shooter loaded their gun. Similarly, we may infer that Dal- ton was hurt prior to his hospitalization and got discharged afterward. Our first contribution is the development of a commonsense reasoning engine that can reason about these two temporal relations. Existing com- monsense models (Gabriel et al., 2021a; Hwang et al., 2021) may generate such inferences for sim- ple sentences with a single event, such as “Bryant Dalton was shot”, but they do not support com- plex sentences with multiple events of interest (e.g. shot, hospitalized). Further, they may conflate the inferences for different events. We …

Content truncated to 500 words