CAREER: Towards Open World Event Knowledge Extraction with Weak Supervision

About This Grant

Understanding events, such as who did what to whom, when and where, is one of the fundamental human activities to learn about the changing world. The answers to these questions underpin the key information conveyed in the overwhelming majority, if not all, of language-based communication. However, current research paradigm suffers from several shortcomings in extracting event knowledge from the open world scenarios. In these scenarios, knowledge extraction from data is limited to a few large domains (e.g., news or biomedical) or common languages (e.g., English, Spanish and Chinese), because of the heavy reliance on the human effort to contextualize data. This includes creating large-scale manual annotations or defining the schematic templates for a few target event types. This project aims to lay the foundation and establish new paradigms for open world event knowledge extraction by developing new and more efficient algorithms to extend the extraction capability to the wide range of scenario, while requiring minimal human effort. This foundation should provide extensive coverage of different event types and be easily adapted to emerging scenarios. The success of this project will directly benefit users of the intelligent information access systems. For applications that analyze emerging and trending topics and events, such as natural disasters, protests and disease outbreak, success of the proposed research will not only provide an accurate and abstractive summary and easy access of each topic for humans, but also allow analysts to better discover the participants of the events, the cause, effects and temporal orders among them, and help discover more insights. The technical aims of the project are divided into three thrusts. Thrust 1 develops schema-guided event extraction approaches. This is done by leveraging the knowledge from the complex target event schema, such as the event type structures (i.e., type name and argument roles), hierarchy and temporal/causal/part-whole relations among the event types, which provide valuable guidance, especially when there is few to no annotations available. While event annotations for most of the domains and scenarios are not existing and extremely expensive and time-consuming to obtain, the large-scale unlabeled in-domain data are usually accessible. Thus, Thrust 2 will further develops a suite of more efficient and novel self-training strategies to make use of the large-scale unlabeled data through self-supervision. In practice, there is even no event type schema available to most of the domains and scenarios, such as natural disaster or disease outbreak. Manually defining an event schema with high coverage is extremely challenging and time consuming as it requires background knowledge in both linguistics and the target domain, and humans need to manually examine a large amount of in-domain data to determine the salient event types. Considering these challenges, Thrust 3 further explores novel solutions to automatically deduce the target event schema, including event types, the roles of their participants, as well as their relations from the raw text and extract their event mentions accordingly. 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.