Multi-Agent Dual Learning

[kweonwooj]

Abstract

• propose multi-agent dual learning framework to boost performance of neural machine translation
• dual learning leverages the duality between primal task (X->Y) and dual task (Y->X)
• SOTA score on WMT 2014 EnDe BLEU : 30.67 (+2.2 compared to Transformer_big)

Details

Introduction

• Dual Learning
  • formulated as a two-agent system where primal model learns f : X -> Y mapping and dual model learns g : Y -> X mapping.
  • given x in X, delta (x, g(f(x))) is the reconstruction loss function used for training signal.
  • Theoretically, monolingual corpus is sufficient to learn NMT model in dual learning framework.
  • refer to original dual learning paper (Xia et al. 2016) accepted at NIPS2016 for more details
• Multi-agent Dual Learning
  • instead of single f and g, multi agent system uses N - 1 additional agents in each side, pre-trained with parallel corpus via different random seed. Ensemble effect boosts the quality of feedback signal.

Algorithm

Results

• Experimental Settings
  • Model : Transformer Big
  • compare with Knowledge Distillation (KD), Back Translation (BT) and two-agent Dual Learning (Dual) each with single and multi-agent
• IWSLT En <-> De
  • KD improves BLEU little, BT has no effect, Dual-5 improves BLEU best
    
• IWSLT Es, Ru, He -> En
  • result is consistent throughout various language pairs in IWSLT
    
• WMT 2014 En <-> De Bilingual
  • KD improves BLEU little, BT has no effect, Dual-5 improves BLEU best (SOTA)
    
• WMT 2014 En <-> De Monolingual
  • also, performs best in unsupervised NMT (SOTA)
    

Image Translation

• compares Multi-Agent Dual Learning with CycleGAN in image translation, with MADL showing more robust and cleaner image translation

Personal Thoughts

• Multi-Agent pre-trained models provide good initialization point and improve the quality of feedback signal
• Existing dual learning seemed to have only theoretical merit, not practical enough. But this paper uncovers the practical merit as well.
• Seems to work across various languages

Link : https://openreview.net/pdf?id=HyGhN2A5tm
Authors : Anonymous