One Source, Two Targets: Challenges and Rewards of Dual Decoding

TL;DR

This paper explores the concept of dual decoding in machine translation, where two target texts are generated simultaneously with interdependence, highlighting its challenges and benefits for applications like multi-target translation.

Contribution

It introduces the dual decoding framework, analyzes its implementation challenges, and demonstrates its advantages over independent translation outputs across four applications.

Findings

Dual decoding improves consistency between outputs

Joint generation enables controlled variation of translations

Challenges include increased complexity and implementation difficulties

Abstract

Machine translation is generally understood as generating one target text from an input source document. In this paper, we consider a stronger requirement: to jointly generate two texts so that each output side effectively depends on the other. As we discuss, such a device serves several practical purposes, from multi-target machine translation to the generation of controlled variations of the target text. We present an analysis of possible implementations of dual decoding, and experiment with four applications. Viewing the problem from multiple angles allows us to better highlight the challenges of dual decoding and to also thoroughly analyze the benefits of generating matched, rather than independent, translations.