Effective Strategies in Zero-Shot Neural Machine Translation

TL;DR

The paper tackles zero-resource translation in multilingual NMT by proposing two lightweight strategies to reduce language bias and training cost. It evaluates these methods—Target Dictionary Filtering and Language as a Word Feature—within a multilingual NMT framework and against pivot and back-translation baselines. Results show meaningful BLEU gains in real zero-shot settings (notably German→Dutch) and substantial reductions in training time and vocabulary size when using language features. These findings demonstrate practical, scalable improvements for zero-shot translation under limited bilingual data, with potential applicability to broader multilingual deployments.

Abstract

In this paper, we proposed two strategies which can be applied to a multilingual neural machine translation system in order to better tackle zero-shot scenarios despite not having any parallel corpus. The experiments show that they are effective in terms of both performance and computing resources, especially in multilingual translation of unbalanced data in real zero-resourced condition when they alleviate the language bias problem.

Figures (2)

• Figure 1: Effect of target dictionary filtering on the decoding process using beam search.
• Figure 2: The NMT architecture which allows the integration of linguistic information as word features.