Learning to Contextually Aggregate Multi-Source Supervision for Sequence Labeling