问题
训练层级多任务。
关键想法
根据任务关系组合不同的任务,简单任务在下,复杂在上,下层任务对上层进行辅助。这里选择NER,EMD,CR,RE四个任务,分三层。NER-〉EMD-〉CR和RE
模型
-
结构
- Embedding: character emb + Glove + ELMo
- Encoder: Multi-layer BiLSTM for each task
- Decoder: CRF for NER and EMD, Scorer for CR and RE
-
数据
利用标注好的不同训练数据
-
训练 随机选择任务,选择batch
技巧
利用BILOU tagging,不是以前的BIO。
效果
NER,EMD,RE三个任务达到SOTA
评论
- 建设任务层级有效果
- 多任务能加速训练
- 不同层embedding联合能提高效果
其他
论文地址: https://arxiv.org/pdf/1811.06031.pdf
Problem
Perform multi-tasks in a hierarchical manner.
Key Ideas
Construct task hierarchy according to their relation. Here choose NER, EMD (Entity Mention Detection), Coreference Resolution and Relation Extraction. NER-〉EMD-〉CR and RE
Model
-
Structure
- Embedding: character emb + Glove + ELMo
- Encoder: Multi-layer BiLSTM for each task
- Decoder: CRF for NER and EMD, Scorer for CR and RE
-
Data
Use different existing labeled training data
-
Training
Randomly select task and data batch.
Tricks
Use BILOU (Beginning, Inside, Last, Outside, Unit) tagging scheme.
Performance
Get new SOTA performance for NER, EMD, RE.
Comments
- Construct task hierarchichy is useful.
- Multi-task can accelerate training.
- Combine different layer’s embedding can help.
Others
Paper link: https://arxiv.org/pdf/1811.06031.pdf