Constrained Multi-Task Learning for Event Coreference Resolution

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4504–4514 June 6–11, 2021. ©2021 Association for Computational Linguistics 4504 Constrained Multi-Task Learning for Event Coreference Resolution 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 We propose a neural event coreference model in which event coreference is jointly trained with five tasks: trigger detection, entity coref- erence, anaphoricity determination, realis de- tection, and argument extraction. To guide the learning of this complex model, we incorpo- rate cross-task consistency constraints into the learning process as soft constraints via design- ing penalty functions. In addition, we propose the novel idea of viewing entity coreference and event coreference as a single coreference task, which we believe is a step towards a uni- fied model of coreference resolution. The re- sulting model achieves state-of-the-art results on the KBP 2017 event coreference dataset. 1 Introduction Event coreference resolution is the task of determin- ing whether two event mentions in a document refer to the same real-world event. For two event men- tions to be coreferent, their triggers (i.e., the words realizing the occurrence of the events) should have the same subtype and their corresponding argu- ments (e.g., the times, places, and people involved) have to be entity-coreferent. However, identifying potential arguments (which is performed by an en- tity extraction system), linking arguments to their event mentions (which is also performed by an event extraction system), and determining whether two event arguments are coreferent (which is the job of an entity coreference resolver), are all non- trivial tasks. Hence, a key challenge in designing an event coreference resolver involves determining how to integrate these noisy components. One of the most common approaches to event coreference resolution is pipelined approaches, where a trigger detection component, which iden- tifies triggers and assigns event subtypes to them, is followed by an event coreference component, which clusters coreferent event mentions. It should therefore not be surprising that errors propagate from the trigger detection component to the event coreference component. To avoid aggravating this error propagation problem, knowledge provided by other information extraction (IE) components (e.g., entity coreference, event arguments) is typically employed as features for training event coreference models (Chen et al., 2009; McConky et al., 2012; Cybulska and Vossen, 2013; Araki et al., 2014; Liu et al., 2014; Peng et al., 2016; Krause et al., 2016; Choubey and Huang, 2017). Oftentimes, these fea- tures provide limited improvements to event coref- erence models as they are too noisy to be useful. Though less popular than pipelined approaches, bootstrapping approaches have been used for event coreference resolution, where an event coreference model is bootstrapped with models trained for one or more related IE tasks. For instance, Lee et al. (2012) incrementally build clusters of coreferent event and entity mentions by iteratively bootstrap- ping event coreference output using entity corefer- ence output and vice versa. While in pipelined ap- proaches …

Content truncated to 500 words