Unsupervised Machine Translation Using Monolingual Corpora Only

Introduction

The paper presents an unsupervised method to machine translation using only monoligual corpora without any alignment between sentences or documents. Monoligual corpora are text corpora that is made up of one language only. This contrasts with the usual translation approach that uses parallel corpora, where two corpora are the direct translation of each other and the translations are aligned by words or sentences.

The general approach of the methodology is to first use a unsupervised word-by-word translation model proposed by [Conneau, 2017], then iteratively improve on the translation by utilizing 2 architectures:

1. A denoising auto-encoder to reconstruct noisy versions of sentences for both source and target languages.
2. A discriminator to align the distributions of the source and target languages in a latent space.

Methodology

The objective function that proposed by the paper is a combination of 3 component objective functions:

1. Reconstruction loss of the denoising auto-encoder
2. Cross domain loss of the auto-encoder
3. Adversarial cross entropy loss of the discriminator

Reconstruction Loss

\begin{align} \mathcal{L}_{auto}(\theta_{enc}, \theta_{dec}, \mathcal{Z}, \ell) = E_{x\sim D_\ell, \hat{x}\sim d(e(C(x),\ell),\ell)}[\Delta(\hat{x},x)] \end{align}

References

1. [Lample, 2018] Lample, G., Conneau, A., Ranzato, M., Denoyer, L., "Unsupervised Machine Translation Using Monolingual Corpora Only". arXiv:1711.00043

1. [Conneau, 2017] Conneau, A., Lample, G., Ranzato, M., Denoyer, L., Jégou, H., "Word Translation without Parallel Data". arXiv:1710.04087