SAINF: Intrinsic Self-Correction for Robust Machine Translation with Large Language Models
Yin ZHANG , Jiahe LI , Saisai HU , Tianyi ZHANG , Yuehan CUI , Shaolin ZHU , Deyi XIONG
Large language models (LLMs) have demonstrated significant capabilities in machine translation (MT), yet they often exhibit robustness deficits, particularly with low-frequency, ambiguous, or culturally-specific terms. While existing methods like Chain-of-Dictionary (CoD) prompting offer partial remedies, they can lack semantic depth or, like interpretation-based approaches, depend on external components such as Quality Estimation (QE) models. To address these limitations, we propose the Self-Adaptive Intrinsic Framework (SAINF), which leverages an LLM’s intrinsic capabilities for self-assessment and refinement. SAINF introduces a conditional workflow: it first employs an efficient CoD translation and prompts the LLMto intrinsically evaluate its quality. If this self-assessment score is below a predetermined threshold, a targeted refinement process is activated. This process involves the LLM intrinsically identifying problematic terms, generating and validating corrective interpretations, and revising the translation accordingly. Experiments on the ChallengeWMT benchmark demonstrate that SAINF significantly improves translation quality and robustness across diverse language pairs, outperforming established baselines.
Large Language Models / Machine Translation / Cross-lingual.
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(2063KB)
