Heng Ji from UIUC will be visiting NUS on 19 July to give this talk on schema induction. The talk is open for both physical and virtual attendance. Please mark your calendars to attend this sure-to-be-wonderful talk on an important NLP topic.
Event schemas encode knowledge of stereotypical structures of events and their connections. As events unfold, schemas are crucial to act as a scaffolding. Previous work on event schema induction focuses either on atomic events or linear temporal event sequences, ignoring the interplay between events via arguments and argument relations. We propose to use ``models as schemas,'' since the probabilistic schema induction models can be probed on demand for event prediction. Furthermore, We aim to learn the hierarchical structure of event complexes. Instead of processing complex events as a whole, humans segment events into chunks, or even nested hierarchies. Such an organization also appears in human memory and correct segmentation has shown to improve recollection of events. We introduce a new concept of Hierarchical Complex Event Schema: a hierarchical graph-based schema representation that encompasses events, arguments, temporal and hierarchical connections and argument relations, and event salience. We design a hierarchical graph schema model to unify these new aspects. Our model is a probabilistic graph generation model capable of predicting events, entities, relations, even entire episodes of events and their logical relations. Our model is able to outperform text-based and graph-based baselines on discovering episode structure and the inferred episode structure can further improve event prediction. Intrinsic evaluations by schema matching and instance graph perplexity, prove the superior quality of our probabilistic graph schema library compared to linear representations. Extrinsic evaluation on schema-guided future event and episode prediction, as well as event graph completion, further demonstrates the predictive power of our event graph model, significantly outperforming human induced schemas and baselines. This work is done by two wonderful woman PhD students at UIUC, Manling Li and Sha Li, and based on collaborations with Prof. Kyunghyun Cho at NYU.
Heng Ji is a professor at the Computer Science Department, and an affiliated faculty member at the Electrical and Computer Engineering Department of University of Illinois at Urbana-Champaign. She is an Amazon Scholar. She received her B.A. and M. A. in Computational Linguistics from Tsinghua University, and her M.S. and Ph.D. in Computer Science from New York University. Her research interests focus on Natural Language Processing, especially on Multimedia Multilingual Information Extraction, Knowledge Base Population and Knowledge-guided Generation. She was selected as "Young Scientist" and a member of the Global Future Council on the Future of Computing by the World Economic Forum in 2016 and 2017. The awards she received include "AI's 10 to Watch" Award by IEEE Intelligent Systems in 2013, NSF CAREER award in 2009, Google Research Award in 2009 and 2014, IBM Watson Faculty Award in 2012 and 2014, Bosch Research Award in 2014-2018, Amazon AWS Faculty Award in 2021, Best-of-ICDM2013 Paper, Best-of-SDM2013 Paper, ACL2020 Best Demo Paper Award, and NAACL2021 Best Demo Paper Award. She was invited by the Secretary of the Air Force and AFRL to join Air Force Data Analytics Expert Panel to inform the Air Force Strategy 2030. She was elected as the North American Chapter of the Association for Computational Linguistics (NAACL) secretary 2020-2021. She has served as the Program Committee Co-Chair of many conferences including NAACL-HLT2018 and AACL-IJCNLP2022, and she has been the coordinator for the NIST TAC Knowledge Base Population track since 2010.