Semantically-informed Hierarchical Event Modeling

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 353–369 July 13-14, 2023 ©2023 Association for Computational Linguistics Semantically-informed Hierarchical Event Modeling Shubhashis Roy Dipta, Mehdi Rezaee, Francis Ferraro Department of Computer Science and Electrical Engineering University of Maryland Baltimore County Baltimore, MD 21250 USA {sroydip1,rezaee1,ferraro}@umbc.edu Abstract Prior work has shown that coupling sequen- tial latent variable models with semantic on- tological knowledge can improve the repre- sentational capabilities of event modeling ap- proaches. In this work, we present a novel, doubly hierarchical, semi-supervised event modeling framework that provides structural hierarchy while also accounting for ontolog- ical hierarchy. Our approach consists of multiple layers of structured latent variables, where each successive layer compresses and abstracts the previous layers. We guide this compression through the injection of struc- tured ontological knowledge that is defined at the type level of events: importantly, our model allows for partial injection of seman- tic knowledge and it does not depend on ob- serving instances at any particular level of the semantic ontology. Across two different datasets and four different evaluation metrics, we demonstrate that our approach is able to out-perform the previous state-of-the-art ap- proaches by up to 8.5%, demonstrating the benefits of structured and semantic hierarchi- cal knowledge for event modeling. 1 Introduction Intuitively, there is a hierarchical nature to complex events: e.g., on Fig. 1, there are two events, one involves going to the hospital and another one is getting treatment. Even if important portions may differ, but these two situations have one abstract concept in common: Cure (of a disease). Clearly, there is a connection among the events reported in a situation and they all contribute to a bigger goal (“Cure” in this case). The main purpose of our work is to exploit this nature of connection to improve event modeling. However, much like linguistic structure, this event structure is generally not directly observed, making it difficult to learn event models that reflect this hierarchical nature. For high-level inspiration, we look to past ap- proaches in syntactic modeling (Collins, 1997; Medical_Intervention Medical_Condition Cure Semantic Knowledge Bill went to hospital Doctors started treatment Figure 1: Complex events can be hierarchical. The pur- ple boxes represent the events themselves (as would be reported in a news story). Blue dashed boxes are annotated semantic frames & the orange dashed box is the more abstract, general frame connecting the “Medical_Condition” and “Medical_Intervention” events. Events and frames are sequentially connected. Klein and Manning, 2003; Petrov et al., 2006): we can approach hierarchical event modeling through structured learning, or through richer (semantic) data. A structural approach accounts for the hi- erarchy as part of the model itself, such as with hierarchical random variables (Cheung et al., 2013; Ferraro and Van Durme, 2016; Weber et al., 2018; Huang and Ji, 2020; Gao et al., 2022). Richer se- mantic data provides hierarchical knowledge, such as event inheritance or composition, as part of the data made available to the model and learning algo- rithm (Botschen …

Content truncated to 500 words