Attention Model

- Hard Attention/Soft Attention[attention model survey](https://arxiv.org/abs/1601.06823)

- Self-Attention[1](https://arxiv.org/abs/1705.04304) [2](https://arxiv.org/abs/1703.03130) 类似 NMT的[coverage](https://arxiv.org/pdf/1601.04811.pdf)， 也和ASR中的 (ARGS)[https://arxiv.org/abs/1506.07503]  相似 ，attention都是自回归

输出减少trigram（三元） 将attention用于decode

    - Attention is all you need

- Global Attention/Local Attention/

- [monotonic Online Attention](https://arxiv.org/abs/1704.00784)