CAREER: Learning to Extract Consistent Event Graphs from Long and Complex Documents

University Of Texas At Dallas, Richardson TX

Investigators

Abstract

Documents about real-world events are published daily. The large number of such documents makes it very hard for people to read and absorb them all, a phenomenon known as “information overload". Applying computer algorithms that can automatically extract events is a promising solution because they can transform large amounts of text into smaller summaries in the form of structured event knowledge graphs that reveal the relationships between the people, places, and times in the events. Current deep learning-based event extraction techniques mainly focus on extracting event knowledge at the level of individual sentences and are unable to extract a knowledge graph spanning multiple sentences with sufficient accuracy or efficiency. For example, existing techniques would struggle with events described in a long document having multiple sections. Moreover, these extraction techniques do not capture accurate information regarding real-life events because they typically include nuanced attributes such as causes and effects. The research goal of this CAREER award is to build information extraction (IE) methods with natural language processing methods, using the latest deep learning-based techniques, to construct an event knowledge graph for storing knowledge and improving the ability of people to track rapidly evolving event information. In the short term, the project will improve the quality and comprehensiveness of event knowledge graphs. In the long run, the project will entirely transform people's experiences and habits in acquiring event knowledge from various sources. The system to be developed through this award will better support numerous event-oriented tasks that people need to perform, such as future event prediction, event factuality verification, and risk event prevention, all of which have profound impacts on society. Moreover, our work would make fundamental contributions to a wide range of interdisciplinary applications such as statutory reasoning based on legal documents, prediction of disease outbreaks, and biomedical document understanding, all of which currently rely on extremely slow and high-cost methods. The general technical goal of this project is to address the knowledge gap of event extraction from long and complex documents (as compared to the traditional sentence-level extraction) and to do so in an efficient manner. The general goal is divided into three sub-research goals. First, to extract the entirety of event attributes, which is not possible for current models trained on a dataset with a predefined schema, the project introduces a new question-answer generation paradigm that enables a novel representation of events from clusters of documents discussing the same events. The project will leverage document hierarchy information for extracting events, which enforces the validity and broad coverage of event information. Motivated by the fact that current event knowledge construction is inefficient and is impaired by pairwise event-event relation predictions, the second research goal is to develop novel techniques enabling the construction of the event knowledge graph. For this purpose, the investigators propose interleaving targeted retrieval and joint modeling of event arguments and entity-entity relations. This not only enables efficient updating of graphs, but also ensures its global consistency. Finally, the third goal is to adapt to individual information-seeking needs, which is not considered by current methods. The project will study schema induction strategies and schema matching algorithms for adapting the event knowledge graph to user preferences. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.