Character-Level Language Modeling with Deeper Self-Attention