Interpreting Word-Level Hidden State Behaviour of Character-Level LSTM Language Models