007. 3.6 What Works Well (and What Can Go Wrong) When We Train Models on Word Embeddings