Multi-level encoder architectures for event causality identification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Event causality identification (ECI) is crucial for understanding text.
• Effectively encoding contextual information in lengthy documents is challenging for ECI.
• Existing methods often overlook event interactions and the main event's role.

Purpose of the Study:

• To propose a novel Multi-level Encoder Architectures (MEA) for enhanced event causality identification.
• To effectively capture context at sentence, event, event-pair, and discourse levels.
• To leverage the main event's significance and inter-event relationships within documents.

Main Methods:

• Utilizing pre-trained language models for sentence-level event representation.
• Constructing an event graph and employing graph neural networks for event-level analysis.
• Applying self-attention mechanisms for event-pair and discourse-level relationship modeling.
• Implementing a multi-level encoding strategy tailored to different contextual granularities.

Main Results:

• Demonstrated the effectiveness of each level of the MEA framework.
• Showcased significant improvements in event causality identification performance.
• Validated the approach on two widely-used public datasets.

Conclusions:

• The proposed Multi-level Encoder Architectures (MEA) effectively encode multi-level contextual information for ECI.
• The framework successfully captures complex event interactions and the role of the main event.
• MEA offers a robust and versatile approach for improving ECI in lengthy documents.