GroupReduce: Block-Wise Low-Rank Approximation for Neural Language Model Shrinking