Towards Semantic Consistency: Dirichlet Energy Driven Robust Multi-Modal Entity Alignment