Learning Language-Specific Layers for Multilingual Machine Translation

TL;DR

This paper introduces Language-Specific Transformer Layers (LSLs) for multilingual machine translation, enhancing capacity without increasing computation, leading to significant translation quality improvements across multiple languages.

Contribution

The paper proposes a novel method of adding language-specific layers within Transformer models, optimized via neural architecture search, to improve multilingual translation performance.

Findings

Achieved 1.3 chrF and 1.5 spBLEU improvements with separate decoder architecture.

Achieved 1.9 chrF and 2.2 spBLEU improvements with shared decoder architecture.

Demonstrated effective capacity increase without additional computational cost.

Abstract

Multilingual Machine Translation promises to improve translation quality between non-English languages. This is advantageous for several reasons, namely lower latency (no need to translate twice), and reduced error cascades (e.g., avoiding losing gender and formality information when translating through English). On the downside, adding more languages reduces model capacity per language, which is usually countered by increasing the overall model size, making training harder and inference slower. In this work, we introduce Language-Specific Transformer Layers (LSLs), which allow us to increase model capacity, while keeping the amount of computation and the number of parameters used in the forward pass constant. The key idea is to have some layers of the encoder be source or target language-specific, while keeping the remaining layers shared. We study the best way to place these layers using a neural architecture search inspired approach, and achieve an improvement of 1.3 chrF (1.5 spBLEU) points over not using LSLs on a separate decoder architecture, and 1.9 chrF (2.2 spBLEU) on a shared decoder one.