Speeding Up Neural Machine Translation Decoding by Cube Pruning