Dual Learning for Machine Translation

[flrngel]

https://arxiv.org/abs/1611.00179 paper from USTC, PKU, Microsoft Research (NIPS 2016)

Summary

Model

1. Prepare 2 agents LM_{a,b} which learned languages(en,fr; WMT14) that outputs log probability
2. 2 translation models P(•|s;Θ_{AB,BA}) are needed
3. Feed P's output to LM and use policy gradient to train
4. Flip data and restart training until model converges

Abstract

• dual-NMT uses reinforcement learning process
• and it works very well

1. Introduction

• Parallel data are costly in Machine Translation(MT)
• Two methods using monolingual data, proposed before this paper
  1. use monolingual data, and then integrate with parallel bilingual data trained model
  2. generate pseudo pair (untrustable method)
     1. train aligned parallel corpora model
     2. generate pseudo bilingual sentence pair
     3. use subsequent learning process
• Dual learning mechanism
  • two agent communication game
  • feedback based

2. Background: Neural Machine Translation

• talks about typical NMT model with using attention

3. Dual Learning from Neural Machine Translation

5. Discussions

• Dual learning is generally applicable to (and already forms in)
  • Speech recognition vs Text to speech
  • Image caption vs image generation
  • Question answering vs Question generation
  • Search vs Keyword extraction
  • etc.
• Not restricted to two tasks (can be generalized as close-loop learning)
• Not only pair of languages, but also can use tuple of multiple(3+) monolinugal data