Towards Open World Event Knowledge Extraction with Weak Supervision скачать в хорошем качестве

Towards Open World Event Knowledge Extraction with Weak Supervision

Date Presented: 09/30/2022 Speaker: Lifu Huang, Virginia Tech Abstract: 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, and at the core is their heavy reliance on the human effort, in both creating large-scale manual annotations or defining the schematic templates for the target event types, limiting the extraction capability only to dominant domains and languages. In this talk, I will introduce our recent efforts towards open world event knowledge extraction by easing the reliance on human and making efficient use of the weak supervision from the guidance of the target event schema and large-scale unlabeled in-domain data. In particular, we aim to answer two research questions: (1) How to leverage the guidance from the target event schema? I will describe a new Query-and-Extract paradigm that takes event schemas as natural language queries to extract candidate events from any input text. This framework has shown the state-of-the-art performance under all the supervised, few-shot and zero-shot extraction settings, and be able to be incrementally updated with new event types for lifelong event extraction. (2) How to make more efficient use of the unlabeled in-domain data as self-supervision to improve the open world event extraction? I will talk about our recent self-training with gradient guidance framework that is particularly designed to characterize the noise introduced by conventional self-training and improve the learning of base event extraction model with pseudo labels. Speaker's Bio: Dr. Lifu Huang is an Assistant Professor in the Computer Science department at Virginia Tech. He obtained a PhD in Computer Science from University of Illinois at Urbana-Champaign in 2020. He has a wide range of research interests in natural language processing, including extracting structured knowledge with limited supervision, natural language understanding and reasoning with external knowledge and commonsense, natural language generation, representation learning for cross-lingual and cross-domain transfer, and multi-modality learning. He is a recipient of the 2019 AI2 Fellowship, 2021 Amazon Research Award.