The Sensitivity of Word Embeddings-based Author Detection Models to Semantic-preserving Adversarial Perturbations