Argument centric causal intervention for cross document event coreference resolution.

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence (AI)

Background:

• Cross-document event coreference resolution (CD-ECR) is crucial for understanding information across texts.
• Existing CD-ECR methods often rely on trigger features, leading to spurious correlations and performance degradation.

Purpose of the Study:

• To propose a novel CD-ECR method, Argument-Centric Causal Intervention (ACCI), that addresses limitations of trigger-based approaches.
• To mitigate spurious correlations between lexical features and coreference by focusing on argument semantics.

Main Methods:

• Constructing a structural causal graph to identify confounding dependencies between triggers and coreference labels.
• Employing backdoor-adjusted interventions to isolate the causal effect of argument semantics.
• Integrating counterfactual reasoning and argument-aware enhancement modules within an end-to-end framework.

Main Results:

• ACCI achieves state-of-the-art performance on benchmark datasets.
• Demonstrated significant improvements with CoNLL F1 scores of 88.4% on ECB+ and 85.2% on GVC.
• The method effectively reduces spurious correlations, enhancing the reliability of coreference decisions.

Conclusions:

• ACCI offers a robust and causally-informed approach to CD-ECR.
• The proposed method enhances NLP model performance by focusing on semantic grounding rather than superficial lexical cues.
• The availability of the implementation facilitates further research and application in the field.