MERIT: Multilingual Expert-Reward Informed Tuning for Chinese-Centric Low-Resource Machine Translation

TL;DR

MERIT is a novel multilingual tuning framework that significantly improves Chinese-centric low-resource machine translation by combining language-specific techniques and reward-guided optimization.

Contribution

It introduces a unified Chinese-centric translation framework with a new reward-guided optimization method for low-resource Southeast Asian languages.

Findings

Targeted data curation improves translation quality.

Reward-guided optimization outperforms model scaling.

Framework enhances translation for Lao, Burmese, Tagalog.

Abstract

Neural machine translation (NMT) from Chinese to low-resource Southeast Asian languages remains severely constrained by the extreme scarcity of clean parallel corpora and the pervasive noise in existing mined data. This chronic shortage not only impedes effective model training but also sustains a large performance gap with high-resource directions, leaving millions of speakers of languages such as Lao, Burmese, and Tagalog with persistently low-quality translation systems despite recent advances in large multilingual models. We introduce \textbf{M}ultilingual \textbf{E}xpert-\textbf{R}eward \textbf{I}nformed \textbf{T}uning (\textbf{MERIT}), a unified translation framework that transforms the traditional English-centric ALT benchmark into a Chinese-centric evaluation suite for five Southeast Asian low-resource languages (LRLs). Our framework combines language-specific token prefixing (LTP) with supervised fine-tuning (SFT) and a novel group relative policy optimization (GRPO) guided by the semantic alignment reward (SAR). These results confirm that, in LRL{\textrightarrow}Chinese translation, targeted data curation and reward-guided optimization dramatically outperform mere model scaling.