Improving Low-Resource Sequence Labeling with Knowledge Fusion and Contextual Label Explanations