Zero-Shot Event Extraction With Context-Definition Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing event extraction models struggle with zero-shot learning, as they are limited to pre-defined ontologies and suffer from domain shifting and high computational complexity, making them inefficient for real-time applications.
Innovation Solution
A three-stage event representation learning framework that includes offline training with context-definition alignment, query-specific warming, and inference, using cosine similarity to classify event mentions based on learned definitions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional supervised event extraction models are used, then classification accuracy is improved, but adaptability to new event types deteriorates
Solution Approach 1:
The patent uses definition templates from the ontology as copies to represent event types during inference. Instead of requiring retraining for new event types, the system copies and utilizes pre-defined definition templates to classify events, enabling zero-shot learning while maintaining classification accuracy through template-based matching.
Solution Approach 2:
The event extraction model is designed to handle both predefined event types and new event types universally. By using definition templates that can be dynamically selected and applied, the system achieves multi-functionality, serving both supervised classification and zero-shot adaptation without requiring separate models.
2Measurement precision
If complex event extraction frameworks are implemented, then extraction accuracy is improved, but inference time complexity increases
Solution Approach 1:
The system performs preliminary actions by pre-defining event type definitions and templates offline. During inference, it only needs to retrieve and match these pre-prepared templates against the input text, significantly reducing inference time complexity while maintaining high extraction accuracy through the use of pre-processed definition structures.
Solution Approach 2:
The patent extracts only the essential definition components from the ontology and stores them as compact templates. This extraction of key information allows the system to maintain high accuracy by focusing on critical event characteristics while reducing computational overhead by eliminating redundant information during inference.
3Reliability
If pre-defined event ontologies are used, then classification reliability is improved, but versatility for real applications deteriorates
Solution Approach 1:
The system dynamically selects and applies definition templates from the ontology based on the input text context. This dynamic approach allows the model to adapt to new event types in real applications by retrieving relevant templates on-the-fly, maintaining classification reliability through ontology-based definitions while achieving versatility for unseen event types.
Data Source
AI summary
A method for zero-shot event extraction, performed by a computer device. The method includes training a context encoder, a first definition encoder, and a second definition encoder with auto extracted context-definition alignment data; retrieving a plurality of verbal synsets from a lexical database; refining a representation model based on the context-definition alignment data and the plurality of verbal synsets; encoding a plurality of candidate event type definitions; encoding the refined representation model with the trained context encoder; and determining whether the encoded representation model belongs to one of the plurality of candidate event type definitions based on a cosine similarity between the encoded representation model, the trained context encoder, the first trained definition encoder, and the second trained definition encoder.