CAREER: Discourse Level Event-Event Relation Identification

点击翻译按钮获取中文摘要

Understanding events (protests, elections, disease outbreaks, natural disasters) from natural language text is key to important analytic tasks like predicting future events, detecting fake news and other attempts to validate events, managing extreme events, answering complex questions and generating concise text summaries for analysis. Existing event extraction systems focus on identifying isolated events, but have rarely considered relations between events. Consequently, the extracted events are merely facts describing who did what, but it is hard to interpret how and why those events happened. Indeed, events tend to be described in a complex relationship with other events, for example, news articles are incomplete if they report an assassination event without mentioning how the event was conducted, or if they describe a protest event without information on why it was launched. This Faculty Early Career Development project aims to generate document-level event graphs that capture rich relations between events mentioned anywhere in a document, which will enable us to contextualize events, transform event extraction from simply extracting individual event facts to extracting informative context-rich event interpretations, and better support various event-oriented applications. The project will integrate research with education, train and prepare future researchers with advanced information extraction views and methods, as well as expose a large number of diverse undergraduate students and high school students to computer science and natural language processing research with a focus on significantly broadening participation of minorities and underrepresented groups. Building document-level event graphs requires identifying relations between two events even when they are sentences away, which presents multiple technical challenges. This project will lay the foundation for discourse-aware event-event relation identification, and study correlations between event-event relations and different dimensions of discourse structures. The research is motivated by the observation that events are major materials in forming a cohesive story and the presence of events is tightly correlated with the overall discourse structure of a document. The project develops both supervised and unsupervised learning methods to build effective discourse level event-event relation recognizers. Specifically, the project develops discourse guided approaches to identify two important types of event-event relations, coreference and temporal ordering, which are fundamental for building meaningful event graphs. Then, guided by event discourse correlations obtained via supervised learning, unsupervised learning methods are developed that can effectively make use of large volumes of unlabeled data, deal with lexical diversity issues and improve robustness of systems for event-event relation identification.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.