Curriculum Learning for Cross-Lingual Data-to-Text Generation With Noisy Data