Year: 2024
Type: dissertation
Abstract: Events are essentially a representation of change, whether it is the change in the physical state of an object, the social status of a person, or the regime of a society.In this dissertation, we study events through their textual representation: we recover the key knowledge elements through written reports of events, identify the relationship between multiple events, and distill this knowledge into event schema graphs, which we then utilize for current event monitoring and future event simulation.Event extraction is a well-established task in natural language processing, however, the definition leans towards local event instances that fall under a pre-defined ontology.This has largely limited the applicability of current event extraction systems: local event instances often do not contain complete information and fixed ontologies mean that trained models can only be applied to a single domain.Our research seeks to mitigate these issues by making event extraction more global and open-domain and eventually more applicable and helpful.How do we extract the event knowledge elements from text?In Chapter 2 and 3, we revisit two core tasks in event extraction: event trigger detection and event argument extraction.For event detection, we explore the possibility of extending a system to new event types with minimal supervision: a few keywords or a few annotated examples.In TAPKEY, we show that high-quality keywords can be used to create pseudo-labeled data, which in turn can bootstrap event detection models.In P4E, we decompose the event detection task into two stages of identification and localization and demonstrate how we can take advantage of pretrained language models to train event detection models with a handful of examples.In GLEN, we tackle the ambitious goal of building a general-purpose event detection system covering 3k+ event types.For event argument extraction, we consider document-level context and propose the new task of informative argument extraction.We further introduce a new formulation of event argument extraction as a templated-based generation task, which allows us to flexibly extend the context length and also utilizes the reasoning ability of generative language models.Which events should we extract?In Chapter 4, we take a step back and re-examine the underlying assumption of event extraction: we are only interested in event types that belong to the pre-defined ontology.We introduce the task of type discovery, which aims to automatically induce new types and extract them when presented with an unlabeled corpus.To achieve this, we design a co-training framework that learns to transfer Mohit Bansal, Eduard Hovy, Clare Voss, Shih-Fu Chang, Kyunghyun Cho etc.They provided me with new perspectives on projects and often one collaboration sprung into a series of new ideas.I am blessed to have called UIUC my home for PhD and so grateful to be part of the tight-knitted DMG family across my master's and PhD years, working with
Author: Li, Sha
Institution: University of Illinois Urbana-Champaign
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220814512
