Structured Attention Networks

方法与贡献

使用可微图模型推断，对子序列和句法树进行注意力计算。

建模结构化选择并计算期望表示；对完整预测进行边缘化是不同的操作。

结合主题阅读

把注意力建模为隐变量：学习注意力的概率分布 →

引用

Yoon Kim, Carl Denton, Luong Hoang, Alexander M. Rush. Structured Attention Networks. ICLR 2017.

@inproceedings{kim2017structured,
  title = {{Structured Attention Networks}},
  author = {Kim, Yoon and Denton, Carl and Hoang, Luong and Rush, Alexander M.},
  year = {2017},
  booktitle = {ICLR 2017},
  url = {https://arxiv.org/abs/1702.00887},
  eprint = {1702.00887},
  archivePrefix = {arXiv}
}