Adaptation for Neural Machine Translation
目标：7月把NMT adaptation文章通读下
场景：假设储备了大量新闻领域双语语料，少量科技领域双语语料（或者没有），科技领域单语语料（大量、少量或者没有）的大菜鸟翻译公司，接到一个科技领域的翻译项目。如何使用现有资源去尽可能的把科技领域的翻译做好。
问题：这里的科技领域就是in-domain，新闻领域是out-domain。如何使用数量有限的in-domain parallel corpus和相对丰富的out-domain parallel corpus来更好地提升in-domain translation performance。
论文阅读：
1. A Survey of Domain Adaptation for Neural Machine Translation
https://arxiv.org/pdf/1806.00258.pdf
文章给做了个简单的归类：
1. Data Centric
- Using Monolingual Corpora
- Synthetic Paralle Corpora Generation
- Using Out-of-Domain Parallel Corpora
- Multi-Domain
- Data Selection
2. Model Centric
- Training Objective Centric
- Instance/Cost Weighting
- Fine Tuning
- Mixed Fine Tuning
- Regularization
- Architecture Centric
- Deep Fusion
- Domain Discriminator
- Domain Control
- Decoding Centric
- Shallow Fusion
- Ensembling
- Neural Lattice Search
关于Data Centric，第三个Using Out-of-Domain Parallel Corpura好理解，在使用out-domain parallel corpus时候，如果是所有的out-domain data带有区分性地使用起来，就是Multi-Domain，如果只是使用某种标准对out-domain parallel corpus中的data进行挑选使用，则是Data Selection。
关于Model Centric，分别在训练阶段（Training）、解码阶段（Decoding）和模型结构上（Architecture Centric）进行相应的调整。感觉比较有意思的是Domain Discriminator，Encoder后面分别挂一个负责target sentence generation的decoder和一个负责predict source domain的Discriminator。借助Discriminator来帮助Encoder捕获domain information。
2. Sentence Selection and Weighting for Neural Machine Translation Domain Adaptation
https://ieeexplore.ieee.org/abstract/document/8360031/
3. Document-Level Adaptation for Neural Machine Translation
http://www.aclweb.org/anthology/W18-2708
4. Instance Weighting for Neural Machine Translation Domain Adaptation
http://www.aclweb.org/anthology/D/D17/D17-1155.pdf
5. An Empirical Comparison of Simple Domain Adaptation Methods for Neural Machine Translation
https://arxiv.org/pdf/1701.03214.pdf
6. Multi-Domain Neural Machine Translation through Unsupervised Adaptation
http://www.aclweb.org/anthology/W/W17/W17-4713.pdf
7. Dynamic Data Selection for Neural Machine Translation
https://arxiv.org/pdf/1708.00712.pdf
8. Cost weighting for neural machine translation domain adaptation
http://www.aclweb.org/anthology/W17-3205