文本分类的主流表示大致分为四类
- Bag-of-words representation models
- Sequence representation models
- Structured representation models
- Attention-based methods
已存在的structured representation models中,结构信息要么作为input, 要么根据已有的树库标注的结构进行有监督预测.也有一些人盐焗自动优化结构,但效果一般.
本文提出了一种没有利用明确的结构标注信息而是利用强化学习识别任务相关结构然后构建结构句子表示.
Model的组成部分:
- policy network – defines a policy for structure discovery
- classification network – makes prediction on top of structured sentence representation and facilitates reward computation for the policy network
- the representation models :
(1)Information Distilled LSTM(HS-LSTM) – selects important, task-relevant words to build sentence representation with a two-level LSTM
(2) Hierarchical Structured LSTM(HS-LSTM) –discovers phrase structures and builds sentence representation with a two-level LSTM.
三个模型的工作机制:
三个模型各个分析:
晚饭时更新