Causal Knowledge Integrated with Attention for Interpretable Event Detection

In recent years, significant breakthroughs have been made in end-to-end models for event detection tasks. Nevertheless, these models are limited in that they solely rely on the source sentence as input and provide detection results, making them opaque black-box models with limited control over learning outcomes. To tackle this limitation, we introduce a novel event detection model that incorporates causal knowledge with attention by constructing a causal graph of event arguments to acquire prior causal knowledge before detection. This approach greatly enhances the interpretability of event detection tasks. Furthermore, we perform an extensive analysis to evaluate the interpretability of our model. The experimental results not only showcase superior performance compared to existing event detection models but also demonstrate the high interpretability of our predicted results.