Mixup Decoding for Diverse Machine Translation

TL;DR

This paper introduces MixDiversity, a novel decoding method for diverse machine translation that uses mixup in the latent space to generate varied translations, improving diversity and faithfulness without extra training.

Contribution

The paper proposes a new mixup-based decoding approach for diverse translation, enabling controllable trade-offs between diversity and faithfulness without additional training.

Findings

Outperforms previous diverse translation methods on multiple benchmarks.

Effectively balances translation diversity and faithfulness through interpolation weight control.

Achieves significant improvements in translation quality and variety.

Abstract

Diverse machine translation aims at generating various target language translations for a given source language sentence. Leveraging the linear relationship in the sentence latent space introduced by the mixup training, we propose a novel method, MixDiversity, to generate different translations for the input sentence by linearly interpolating it with different sentence pairs sampled from the training corpus when decoding. To further improve the faithfulness and diversity of the translations, we propose two simple but effective approaches to select diverse sentence pairs in the training corpus and adjust the interpolation weight for each pair correspondingly. Moreover, by controlling the interpolation weight, our method can achieve the trade-off between faithfulness and diversity without any additional training, which is required in most of the previous methods. Experiments on WMT'16 en-ro, WMT'14 en-de, and WMT'17 zh-en are conducted to show that our method substantially outperforms all previous diverse machine translation methods.