🧑‍🔬 Mia Chen

Mia Chen works on neural machine translation, with papers here dating from 2018 and 2019. The work asks what really drives translation quality: which improvements belong to a model architecture and which to the training methods packaged with it, and how far a single shared network can be stretched across many languages at once. The approach is empirical and large in scale, comparing Transformer-based and related systems under carefully matched conditions. It spans machine translation and multilingual and cross-lingual NLP.

Both of Chen's papers here are widely cited. The Best of Both Worlds (2018) rebuilt recent architectures such as ConvS2S and the Transformer within one framework and gave them identical preprocessing. This made it possible to separate the effect of each design from the modeling and training techniques that usually come with it. The systems were evaluated on the WMT'14 English→French and English→German tasks. Massively Multilingual Neural Machine Translation in the Wild (2019) trained one model covering 103 languages on a web-gathered corpus of more than 25 billion sentence pairs into and out of English. It set this model against separate bilingual Transformer baselines for each direction and examined vocabulary, how training data is sampled, and model size.

The newer of the two papers moves from controlled benchmark comparisons to multilingual translation under realistic conditions, where the amount of data, the subject matter and the level of noise differ widely from one language to the next. Chen's bylines list Google in Mountain View from 2018 to 2022. Orhan Firat, Ankur Bapna, Melvin Johnson, George Foster and Wolfgang Macherey appear on both papers here. Other co-authors include Ashish Vaswani, Mike Schuster, Colin Cherry and Dmitry Lepikhin.

Mia Chen has 2 papers in our library, in 2018–2019, most of them on Machine translation, Multilingual & cross-lingual NLP; often with Ankur Bapna, George Foster, Melvin Johnson.