Non-Autoregressive Machine Translation with Disentangled Context Transformer

TL;DR

This paper introduces the DisCo transformer, a non-autoregressive machine translation model that generates all tokens simultaneously using disentangled contexts, leading to faster inference with competitive translation quality.

Contribution

The paper proposes a novel attention-masking model and inference algorithm for non-autoregressive translation, enabling parallel token generation and improved decoding speed.

Findings

Achieves comparable or better translation quality than autoregressive models.

Significantly reduces decoding time across multiple translation tasks.

Demonstrates effectiveness on 7 translation directions with various data sizes.

Abstract

State-of-the-art neural machine translation models generate a translation from left to right and every step is conditioned on the previously generated tokens. The sequential nature of this generation process causes fundamental latency in inference since we cannot generate multiple tokens in each sentence in parallel. We propose an attention-masking based model, called Disentangled Context (DisCo) transformer, that simultaneously generates all tokens given different contexts. The DisCo transformer is trained to predict every output token given an arbitrary subset of the other reference tokens. We also develop the parallel easy-first inference algorithm, which iteratively refines every token in parallel and reduces the number of required iterations. Our extensive experiments on 7 translation directions with varying data sizes demonstrate that our model achieves competitive, if not better, performance compared to the state of the art in non-autoregressive machine translation while significantly reducing decoding time on average. Our code is available at https://github.com/facebookresearch/DisCo.