Adversarial Transfer Learning for Chinese Named Entity Recognition with Self-Attention Mechanism 论文解

Adversarial Transfer Learning for Chinese Named Entity Recognition with Self-Attention Mechanism 论文解读

 

前言:论文主要引入了两个创新点:

  1. 如何应用CWS(chinese word segment)的信息
  2. 在bilstm和crf层中间加了self-attention(第一次引入到NER)

过程

Adversarial Transfer Learning for Chinese Named Entity Recognition with Self-Attention Mechanism 论文解_第1张图片

 

从整个图中可以看出包括以下几个模块

  1. embedding layer
  2. Shared-private feature extractor
  3. Self-attention
  4. Task-specific crf
  5. Task discriminator
  6. Training
  1. embedding layer

主要包括左边NER模型的embedding和CWS的embedding都是独立的,即embedding向量不共享(CWS的embedding纬度多大??也是n吗)

(2)Shared-private feature extractor

都是Bi-LSTM来提取特征,shared feature extractor是怎么提取特征的,输入是各个embedding的相加?

  1. Self-attention

其实就是multi-head self-attention,可以看下attention is all you need这篇paper

  1. Task-specific crf

就是crf,只是NER和CWS对呀的label不一样,所以反向传播更新的有差异

 

 

 

你可能感兴趣的:(每天一篇自然语言顶会paper)