[论文阅读笔记 --- 13] StructBERT: Incorporating Language Structures into Pre-training for Deep Language Un
Bert 预训练思路
文章图片
Bert 预训练模型 Bert 预训练两大子任务
- Mask Token Prediction: 对于Mask的位置,多分类任务,从此表中预测处正确的词
- Next Sentence Prediction: 输入两个句子,判断S1和S2是否是上下句的关系。
![[论文阅读笔记 --- 13] StructBERT: Incorporating Language Structures into Pre-training for Deep Language Un](https://img.it610.com/image/info8/bb6ee81379ac47428569cc9b13b24a14.jpg)
文章图片
Mask Token Prediction
![[论文阅读笔记 --- 13] StructBERT: Incorporating Language Structures into Pre-training for Deep Language Un](https://img.it610.com/image/info8/df3914957c8d4acb9441a4ca306ee62a.jpg)
文章图片
Next Sentence Prediction StructBert
![[论文阅读笔记 --- 13] StructBERT: Incorporating Language Structures into Pre-training for Deep Language Un](https://img.it610.com/image/info8/65535b9a54e34d18b702cbee2cb2c0ae.jpg)
文章图片
对于单个句子,考虑Word-Leval Prediction
(1) 预测被Mask的词
(2) 选择一些不包含Mask词的连续三元组,打乱三元组的顺序,预测重建该三元组
![[论文阅读笔记 --- 13] StructBERT: Incorporating Language Structures into Pre-training for Deep Language Un](https://img.it610.com/image/info8/b7a5f02ce2e748ab85c6ee7bfba436a2.jpg)
文章图片
最后在句子被预测处理过的位置所对应的正确单词。
对于两个句子,考虑Sentence-Level Prediction
![[论文阅读笔记 --- 13] StructBERT: Incorporating Language Structures into Pre-training for Deep Language Un](https://img.it610.com/image/info8/9e38409abbae4e7e885807896415e011.jpg)
文章图片
【[论文阅读笔记 --- 13] StructBERT: Incorporating Language Structures into Pre-training for Deep Language Un】(1) 考虑为三分类任务,给定句子对S1和S2,存在以下三种情况,S2是S1的下一句,S2是S1的上一句,S1和S2没有上下句关系。
推荐阅读
- 论文阅读笔记|Deep Sliding Shapes for Amodal 3D Object Detection in RGB-D Images(CVPR2016)(1)
- 清华大佬SCI写作心得
- MATLAB之机器学习——BP神经网络
- KGSF(通过基于语义融合的知识图谱来改善会话推荐系统 KDD2020)