Alibaba Cloud PAI improves neural machine translation training efficiency by 5 times

Abstract: In the past two years, Neural Machine Translation (NMT: Neural Machine Translation) technology has emerged, and the quality of translation has been greatly improved. Unfortunately, the training cost of NMT systems is very high, limiting the widespread use of this technique. The Alibaba translation team achieved a 5-fold leap in model training efficiency by using the Alibaba Cloud machine learning platform PAI, which has been applied to the English-Russian e-commerce translation quality optimization project.

In the past two years, Neural Machine Translation (NMT: Neural Machine Translation) technology has emerged, and the quality of translation has been greatly improved. Unfortunately, the training cost of NMT systems is very high, limiting the widespread use of this technique.

The Alibaba translation team achieved a 5-fold leap in model training efficiency by using the Alibaba Cloud machine learning platform PAI, which has been applied to the English-Russian e-commerce translation quality optimization project.


1. What is NMT

The emergence of the term NMT dates back to September 1, 2014. The research group of Professor Bengio of the University of Montreal in Canada published their latest research results on the open paper website arxiv "neural machine translation by jointly learning to align and translate', NMT entered people's field of vision.

They designed a set of neural networks to encode source language sentences into a vector using an encoder, and then use a decoder to decode the vector to produce translations. At the same time, an attention mechanism is introduced to further improve the translation quality.


2. How AliTranslation uses PAI

Inside Alibaba, AliTranslation is responsible for providing multilingual services for 1688 International Station, AliExpress, etc. Some Chinese information filled in by Chinese sellers will be automatically translated into multiple languages ​​by machines. The team also provides services for Dingding, Southeast Asian e-commerce company Lazada, etc.

Last year, they applied NMT technology to communication scenarios for the first time. Although the translation quality has improved a lot, the model training takes too long. 30 million training data generally needs to be trained for more than 20 days on a single GPU card to get a preliminary usable model.

After that, they tried to develop an NMT system that supports distributed training on the Alibaba Cloud machine learning platform PAI, and completed the first version at the end of March. In the British-Russian e-commerce translation quality optimization project, the distributed NMT system greatly improved the training speed, shortening the model training time from 20 days to 4 days.

Figure: Convergence acceleration ratio obtained

on obstacle.


3. What is PAI?

PAI is the first major tool released by Alibaba's "NASA" plan, which is fully compatible with the world's mainstream deep learning open source frameworks. At the same time, the bottom layer provides powerful cloud heterogeneous computing resources, including CPU, GPU, and FPGA. On the GPU side, multi-card scheduling can be flexibly implemented.

Inside Alibaba, PAI is already widely used. Taobao search uses PAI's parameter server, which can disperse models with tens of billions of features to dozens or even hundreds of parameter servers, breaking the scale bottleneck. Finally, the search results are sorted based on the characteristics of products and users.

In the past year, Alibaba Cloud has assisted customers in landing a number of major AI applications. But in order for artificial intelligence to truly become an inclusive technology, it needs a production tool that is available to everyone. PAI was born for this.