Google AI: Exploring Multilingual Neural Machine Translation
Due to advances in neural machine translation (NMT), there has been a lot of developments in Machine Translation (MT) in the recent past. But just like with many other neural network models, the success of NMT hinges on massive and high-quality training data.
When it comes to MT, it’s easy to find data for some languages but not for others. According to Google AI, Multilingual NMT can help through the learning signal from one language to benefit the quality of translation to other languages. But is it possible to train a single model using available data, irrespective of the huge differences across languages in data size, complexity, scripts, and domains?
Towards Universal Machine Translation
Google AI researchers recently introduced efforts towards building a universal neural machine translation (NMT) system with the ability to translate between any language pair. They built a single massively multilingual NMT model that was trained on over 25 billion examples and capable of handling approximately 103 languages.
The suggested system shows effective transfer learning ability which in turn significantly improves the translation quality of low-resource languages while keeping high-resource language translation quality on-par with competitive bilingual baselines. The researchers have provided an in-depth analysis of various aspects of model building that are crucial to achieving quality and practicality in universal NMT.
Potential Uses and Effects
Taking into consideration the demand for high-quality training data needed to achieve model accuracy, it’s crucial for researchers to think out of the box when it comes to situations where data is scarce or not presently attainable.
This work helps present issues in multilingual NMT research that need to be worked on and taken into consideration. Apart from reducing operational costs, multilingual models improve performance on low and zero-resource language pairs due to joint training and consequent positive transfer from higher resource languages.
Such multilingual models provide a way to easily extend to new languages, domains, and down-stream tasks, even when parallel data is unavailable. And while the work provides a prototype for a high-quality universal translation system, there’s still a lot missing and a long way to go.
This work makes multilingual NMT a testbed for machine learning practitioners and theoreticians interested in exploring the annals of multi-task learning, meta-learning, training dynamics of deep nets and much more according to Google AI researchers.
Read more: https://ai.googleblog.com/2019/10/exploring-massively-multilingual.html