Towards Open World Event Knowledge Extraction with Weak Supervision

Explore a comprehensive lecture on open world event knowledge extraction using weak supervision techniques. Delve into the challenges of current research paradigms and discover innovative approaches to reduce reliance on human effort in event extraction. Learn about the Query-and-Extract paradigm, which leverages event schemas as natural language queries to extract candidate events from text. Examine the self-training with gradient guidance framework, designed to improve event extraction models using pseudo labels from unlabeled data. Gain insights into supervised, few-shot, and zero-shot extraction settings, as well as strategies for incrementally updating models with new event types. Understand the application of these techniques in intelligence analysis and their potential for advancing natural language processing in open world scenarios.

Intro
What is event extraction?
Application - Intelligence Analysis
Traditional Event Extraction
How to solve the limitations?
Our First Solution: Query-and-Extract
Query and Extract: Type-oriented Binary Decoding
Approach Details - Event Trigger Detection
Approach Details - Event Argument Extraction
Experiments - Supervised Event Extraction
Experiments-Zero-Shot Event Extraction
Event Detection based on Type Specific Prompts
Few-Shot Event Detection
Can the model be continuously updated?
Episodic Memory Prompting
How well the model retrains the capability on old types?
Pros and Cons of Query-and-Extraction Paradigm
Our Second Solution: Self-Training
Why Traditional Self-Training Does Not Work?
Self-Supervised Event Extraction with Gradient Guidance
Self-Training with Gradient Guidance
Evaluation of the Scoring Model
How Effective is the STGG to Event Extraction?
Future Directions: Open-Environment Event Extraction