本文介绍论文《Adversarial Transfer Learning for Chinese Named Entity Recognition with Self-Attention Mechanism》EMNLP2018
(希望可以每周写一篇博客)
任务介绍
中文命名实体识别首先有识别实体边界,然后将他们分类到预定义的类别中,比如PER(人物), LOC(地点)等。
动机
作者观察到
- Weibo NER训练集很小,只有1.3K。
- 中文命名实体识别和中文词分割有相似之处和不同之处
所以作者希望用transfer learning to learn task-shared information, using adversarial training to remove task-specific information.
另外,由于词的类别(PER, LOC, ORG etc.)与词的上下文是相关的,用self-attention来学习这种关联
方法
左边是NER task,右边是CWS task,中间是adversarial training。Adversarial training 用来预测当前输入是哪一个task, NER or CWS,希望shared bilstm学到的信息跟具体任务无关,跟ner and cws task 都有关。Transfer learning to learn task-shared information, adversarial training to remove task-specific information.
代码:
model
- 共享BILSTM
2. CWS 私有BILSTM
3. NER 私有BILSTM
Loss
- NER loss
2. CWS loss
3. Adversarial loss
4. Combine the loss
结果
在SighanNER 和 Weibo NER 都取得了state of the art result. F1 score 90.64, 53.08 respectively.
引用
http://aclweb.org/anthology/D18-1017
https://github.com/CPF-NLPR/AT4ChineseNER
https://vimeo.com/306354811