Goal:
+train an efficient universal model that can translate b/n any language

Progress:
+ work sets a new milestone towards building a single model

Massively Multilingual Neural Machine Translation
in the Wild: Findings and Challenges https://arxiv.org/pdf/1907.05019.pdf

@fbk_mt

Languages in total: 103, can result > 10k translation directions.

Training examples: 25Billion

Core points:
- transfer-learning ability across languages
- benefits low-resource languages
- keeps performance of high-resource languages
- detailed analysis on model training

Open problems areas as mentioned in the paper:
- Data & supervision
- Learning
- Model capacity
- Arch & Vocab

Continues ...