On The Alignment Problem In Multi-Head Attention-Based Neural Machine Translation