Zero-shot Label-Aware Event Trigger and Argument Classification

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1331–1340 August 1–6, 2021. ©2021 Association for Computational Linguistics 1331 Zero-shot Label-Aware Event Trigger and Argument Classification Hongming Zhang∗,1,2, Haoyu Wang2, Dan Roth2 Department of Computer Science and Engineering, HKUST Department of Computer and Information Science, UPenn hzhangal@cse.ust.hk, {why16gzl, danroth}@seas.upenn.edu Abstract Identifying events and mapping them to a pre-defined taxonomy of event types has long been an important NLP problem. Most previ- ous work has relied heavily on labor-intensive, domain-specific, annotation, ignoring the se- mantic meaning of the event types’ labels. Consequently, the learned models cannot ef- fectively generalize to new label taxonomies and domains. We propose a zero-shot event ex- traction approach, which first identifies events with existing tools (e.g., SRL) and then maps them to a given taxonomy of event types in a zero-shot manner. Specifically, we leverage la- bel representations induced by pre-trained lan- guage models, and map identified events to the target types via representation similarity. To semantically type the events’ arguments, we further use the definition of the events (e.g., ar- gument of type “Victim” appears as the argu- ment of event of type “Attack”) as global con- straints to regularize the prediction. The pro- posed approach is shown to be very effective on the ACE-2005 dataset, which has 33 trig- ger and 22 argument types. Without using any annotation, we successfully map 83% of the triggers and 54% of the arguments to the se- mantic correct types, almost doubling the per- formance of previous zero-shot approaches1. 1 Introduction Event extraction, the process of identifying events triggers and arguments and classifying them into a set of pre-defined types is an important part of natural language understanding, and a commonly studied NLP task (Grishman et al., 2005). Con- sider the example shown in Figure 1, where two events (i.e., “war” and “protesting”) are identified. ∗This work was done when the first author was visiting the University of Pennsylvania. 1Our code and models will be available at http:// cogcomp.org/page/publication_view/942. Figure 1: Event classification examples. Two events are highlighted with red and blue colors. Triggers and arguments are in bold and underline fonts, respectively. By mapping them to [Conflict:Attack] and [Con- flict:Demonstrate] and using the knowledge of “At- tack” might result in “Demonstrate”, we can infer that the war in Iraq is probably the cause of the protesting in Pakistan. Most existing event extraction work (Wadden et al., 2019; Lin et al., 2020) treats event iden- tification as a supervised sequence labeling task and event classification as a supervised classifica- tion problem, and relies on large amounts of event- specific annotated text. Take ACE-2005 (Grishman et al., 2005) as an example; the training set of ACE consists of 4,419 events, annotated and typed into 33 event types. Such large-scale and high-quality annotation requires significant expertise, and it fa- cilitates the success of supervised learning models. However, scaling these efforts to new domains and more event types is very costly and unrealistic. Indeed, there has already been some effort …

Content truncated to 500 words