Multi-level encoder architectures for event causality identification

doi:10.1016/j.neunet.2025.108356

Multi-level encoder architectures for event causality identification

2025-11-24

0

PRE

AI

DOI：10.1016/j.neunet.2025.108356

Original

Original request for help

Save

Abstract

En 中文

• The Multi-Level Encoder Architecture (MEA) is proposed to explicitly capture event information at multiple levels: sentence-level, event-level, event-pair-level, and discourse-level. By integrating information from these diverse levels, the MEA enhances event representation and significantly improves the accuracy of causality predictions. • A novel method for constructing graphs based on major events is introduced. This method can be seamlessly incorporated into Graph Neural Networks (GNNs) to model event relationships and enhance the encoding process. • Comprehensive evaluations conducted on two widely used public benchmarks for Event Causality Identification demonstrate that our model outperforms existing methods. This highlights its superior performance and affirms its potential for practical applications.

Multi-level encoder architectures for event causality identification

Abstract

Journal

Organization