Non-Adversarial Unsupervised Word Translation

  • 2018-08-13 15:13:05
  • Yedid Hoshen, Lior Wolf
  • 0

Abstract

Unsupervised word translation from non-parallel inter-lingual corpora hasattracted much research interest. Very recently, neural network methods trainedwith adversarial loss functions achieved high accuracy on this task. Despitethe impressive success of the recent techniques, they suffer from the typicaldrawbacks of generative adversarial models: sensitivity to hyper-parameters,long training time and lack of interpretability. In this paper, we make theobservation that two sufficiently similar distributions can be alignedcorrectly with iterative matching methods. We present a novel method that firstaligns the second moment of the word distributions of the two languages andthen iteratively refines the alignment. Extensive experiments on wordtranslation of European and Non-European languages show that our methodachieves better performance than recent state-of-the-art deep adversarialapproaches and is competitive with the supervised baseline. It is alsoefficient, easy to parallelize on CPU and interpretable.

 

Quick Read (beta)

loading the full paper ...