Modeling document causal structure with a hypergraph for event causality identification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Document-level event causality identification (ECI) seeks causal links between events.
• Current graph neural network methods struggle to capture all relevant event connections for ECI.
• Event causal relations are often interdependent, suggesting a need for more sophisticated modeling.

Purpose of the Study:

• To propose a novel neural causal hypergraph model (NCHM) for document-level event causality identification.
• To address the limitations of existing methods in modeling interdependent event causal relations.
• To improve the accuracy of identifying causal relationships between events within a document.

Main Methods:

• Utilized a hypergraph to represent interdependent event causal relations as a document causal structure.
• Developed a pairwise event semantics learning module (PES) using prompt learning for event representation and causal connection identification.
• Implemented a document causal structure learning module (DCS) with a hypergraph convolutional neural network for document-wise event representation.
• Concatenated pairwise and document-wise event representations for the final ECI task.

Main Results:

• The proposed NCHM significantly outperforms state-of-the-art algorithms on both EventStoryLine and English-MECI corpora.
• The hypergraph approach effectively models the complex, interdependent nature of event causality.
• The PES and DCS modules contribute to enhanced event representation learning for ECI.

Conclusions:

• The neural causal hypergraph model (NCHM) offers a superior approach to document-level event causality identification.
• Modeling interdependent causal relations via hypergraphs is crucial for accurate ECI.
• NCHM demonstrates the potential of advanced graph-based neural networks for complex NLP tasks.