Extracting Interpretable Local and Global Representations from Attention on Time Series