Conundrums in Event Coreference Resolution: Making Sense of the State of the Art

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1368–1380 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 1368 Conundrums in Event Coreference Resolution: Making Sense of the State of the Art Jing Lu and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {ljwinnie,vince}@hlt.utdallas.edu Abstract Despite recent promising results achieved by span-based approaches to event coreference resolution, there is a lack of understanding of what has been improved. We present an empir- ical analysis of our state-of-the-art span-based event coreference resolver (Lu and Ng, 2021) with the goal of providing the general NLP au- dience with a better understanding of the state of the art and coreference researchers with di- rections for future research. 1 Introduction Recent years have seen the successful application of span-based neural models to key entity-based in- formation extraction (IE) tasks such as entity coref- erence resolution (CR) (Lee et al., 2017, 2018) and relation extraction (Luan et al., 2019). Unlike many non-span-based neural models, which typi- cally learn task-specific contextualized word repre- sentations (Peters et al., 2018), span-based models are designed to learn task-specific representations of text spans. This potentially allows span-based models to create better representations of the en- tity mentions involved in entity-based IE tasks than their non-span-based counterparts as many entity mentions are multi-word expressions. Can the successes of span-based models be ex- tended to event CR? The vast majority of existing event coreference resolvers were developed in the pre-neural NLP era, focusing primarily on feature engineering (Ahn, 2006; Chen et al., 2009; Chen and Ji, 2009; McConky et al., 2012; Araki et al., 2014; Bejan and Harabagiu, 2010, 2014; Chen and Ng, 2014, 2015, 2016; Cybulska and Vossen, 2015a,b; Yang et al., 2015; Krause et al., 2016; Lu et al., 2016) and adapting the models originally developed for entity coreference to event corefer- ence (Liu et al., 2014; Peng et al., 2016; Lu and Ng, 2016, 2017, 2020). Neural event coreference models (Choubey and Huang, 2017, 2018, 2021; Huang et al., 2019) are few and far between, let alone span-based neural event coreference models. Recently, Lu et al. (2020) designed the first span- based model that has achieved state-of-the-art re- sults on a standard event coreference dataset. De- spite these promising results, the use of span-based models in event-based IE tasks such as event coref- erence is still in its infancy. In particular, there is little understanding of what has been improved. In light of the above discussion, we present an empirical analysis of our state-of-the-art span- based event coreference resolver (Lu and Ng, 2021) with the goal of gaining insights into its behavior. We believe that our analysis will not only provide the general NLP audience with a better understand- ing of the strengths and weaknesses of span-based event coreference models, but also provide corefer- ence researchers with directions for future work.1 2 Tasks and Definitions In this section, we define the six tasks to be learned by our …

Content truncated to 500 words