Semantic aware enhanced event causality identification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Event Causality Identification (ECI) research often relies on external knowledge, neglecting intrinsic sentence semantics.
• Existing methods struggle with implicit causal relations and face data limitations, especially in Chinese.

Purpose of the Study:

• To propose a novel framework, Hierarchical Feature Extraction and Prompt-aware Attention (HFEPA), for enhanced ECI.
• To address the scarcity of annotated datasets in Chinese for ECI research.
• To improve the identification of implicit causal relationships between events.

Main Methods:

• Hierarchical Feature Extraction (HFE) module: Extracts event and segment-level features for richer semantic representation.
• Prompt-aware Attention (PAA) module: Leverages pre-trained models to capture implicit causal knowledge and contextual information.
• Development of the Chinese News Causality (CNC) dataset: A large-scale dataset to mitigate data scarcity.

Main Results:

• The HFEPA framework significantly outperforms existing methods on both EventStoryLine and the new CNC dataset.
• The proposed CNC dataset, with 25,629 event mentions and 5,569 causal pairs, is the largest Chinese ECI dataset to date.
• HFEPA effectively captures implicit causal relations by integrating hierarchical features and prompt-aware attention.

Conclusions:

• HFEPA offers a more effective approach to Event Causality Identification, particularly for implicit relations.
• The development of the CNC dataset is a significant contribution to advancing Chinese ECI research.
• The study highlights the importance of intrinsic semantic features and prompt-aware mechanisms in ECI.