On the Effectiveness of Low-Rank Matrix Factorization for LSTM Model Compression