Event Schema Induction with Double Graph Autoencoders

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2013 - 2025 July 10-15, 2022 ©2022 Association for Computational Linguistics Event Schema Induction with Double Graph Autoencoders Xiaomeng Jin, Manling Li, Heng Ji University of Illinois Urbana-Champaign {xjin17, manling2, hengji}@illinois.edu Abstract Event schema depicts the typical structure of complex events, serving as a scaffolding to ef- fectively analyze, predict, and possibly inter- vene in the ongoing events. To induce event schemas from historical events, previous work uses an event-by-event scheme, ignoring the global structure of the entire schema graph. We propose a new event schema induction frame- work using double graph autoencoders, which captures the global dependencies among nodes in event graphs. Specifically, we first extract the event skeleton from an event graph and de- sign a variational directed acyclic graph (DAG) autoencoder to learn its global structure. Then we further fill in the event arguments for the skeleton, and use another Graph Convolutional Network (GCN) based autoencoder to recon- struct entity-entity relations as well as to detect coreferential entities. By performing this two- stage induction decomposition, the model can avoid reconstructing the entire graph in one step, allowing it to focus on learning global structures between events. Experimental re- sults on three event graph datasets demonstrate that our method achieves state-of-the-art perfor- mance and induces high-quality event schemas with global consistency. 1 1 Introduction Event Schema (Chambers and Jurafsky, 2008, 2009; Balasubramanian et al., 2013; Nguyen et al., 2015; Modi et al., 2016; Li et al., 2021) is induced from historical events to describe the common or stereo- typical evolution pattern of complex events. Figure 1b shows an example schema of complex event “IED-bombing” (Improvised Explosive Device), where multiple events are inter-connected via tem- poral links (e.g., TRANSPORT happens after AS- SEMBLE) and arguments (e.g., the transporting EN- TITY is the weapon that is being assembled; the 1The programs, data and resources are made publicly available for research purpose in https://github.com/ tracyjin/DoubleGAE.git transporting DESTINATION is controlled by the as- sembler, i.e., the AGENT). It enables a descriptive analysis of inter-event structures, as well as the pre- diction of future events over temporal-based and argument-based structures. A number of methods have been proposed for learning event schemas from instance event graphs, called event schema induction, which fall into three categories. The set-based (Cham- bers, 2013; Sha et al., 2016; Huang et al., 2016) and sequence-based (Granroth-Wilding and Clark, 2016; Rudinger et al., 2015) methods treat a com- plex event as a set or a linear sequence of atomic events, respectively. However, they fail to cap- ture the multi-dimensional evolution of real-world complex events, i.e., multiple events may precede or follow one event. Graph-based methods (Li et al., 2020, 2021), instead, adopt graphs to for- mulate event schemas. A graph model is usually learned from instance event graphs through gener- ating the schema event by event. Although graph- based methods are theoretically superior to the first two categories, existing graph-based …

Content truncated to 500 words