Zero-Shot Translation using Diffusion Models

TL;DR

This paper introduces a novel zero-shot neural machine translation method using diffusion models, enabling non-autoregressive translation between unseen language pairs by conditioning on source sentences.

Contribution

It adapts diffusion probabilistic models for text translation, demonstrating zero-shot capabilities and non-autoregressive translation, which are new advancements in NMT.

Findings

Effective zero-shot translation between unseen language pairs

Non-autoregressive translation achieved with diffusion models

Model outperforms traditional methods in certain zero-shot scenarios

Abstract

In this work, we show a novel method for neural machine translation (NMT), using a denoising diffusion probabilistic model (DDPM), adjusted for textual data, following recent advances in the field. We show that it's possible to translate sentences non-autoregressively using a diffusion model conditioned on the source sentence. We also show that our model is able to translate between pairs of languages unseen during training (zero-shot learning).