Unsupervised Embeddings with Graph Auto-Encoders for Multi-domain and Multilingual Hate Speech Detection