seq2seq
![](https://img.haomeiwen.com/i1507799/8f3c91a72e86aeb3.png)
![](https://img.haomeiwen.com/i1507799/ff2644a2d3b51faa.png)
- 离散的词ID转换为词向量
与Encoder 中的这个步骤是一样的, 只不过embedding矩阵与Encoder的可能不一样,比如翻译源语言与目标语言需要使用不同的embbedding矩阵,但是如文本摘或是文本风格改写这种就可以使用同一个embedding矩阵。
- 由encoder的输出结合decoder的prev_hidden_state生成energy
![](https://img.haomeiwen.com/i1507799/2aeb038ed3f1e319.png)
- 由energy 到概率
![](https://img.haomeiwen.com/i1507799/e224c34aae11a53a.png)
- context 向量合成
![](https://img.haomeiwen.com/i1507799/726389be9f475155.png)
- prev_hidden_state, 词向量, context向量通过GRU单元生成下一时刻hidden_state
![](https://img.haomeiwen.com/i1507799/cb9c6fca2e0cc020.png)
Pointer Network
![](https://img.haomeiwen.com/i1507799/891f9a014fe3e117.png)
![](https://img.haomeiwen.com/i1507799/275ea32ea9030046.png)
网友评论