Viterbi Decoding of Directed Acyclic Transformer for Non-Autoregressive Machine Translation

TL;DR

This paper introduces a Viterbi decoding method for DA-Transformer in non-autoregressive machine translation, improving translation accuracy without sacrificing decoding speed.

Contribution

It presents a Viterbi decoding framework that guarantees optimal translation and decoding path for DA-Transformer models.

Findings

Consistent performance improvement over baseline DA-Transformer

Maintains similar decoding speedup as original models

Enhances translation accuracy in non-autoregressive MT

Abstract

Non-autoregressive models achieve significant decoding speedup in neural machine translation but lack the ability to capture sequential dependency. Directed Acyclic Transformer (DA-Transformer) was recently proposed to model sequential dependency with a directed acyclic graph. Consequently, it has to apply a sequential decision process at inference time, which harms the global translation accuracy. In this paper, we present a Viterbi decoding framework for DA-Transformer, which guarantees to find the joint optimal solution for the translation and decoding path under any length constraint. Experimental results demonstrate that our approach consistently improves the performance of DA-Transformer while maintaining a similar decoding speedup.