Aligning Translation-Specific Understanding to General Understanding in Large Language Models

Although large language models (LLMs) have shown surprising languageunderstanding and generation capabilities, they have yet to gain arevolutionary advancement in the field of machine translation. One potentialcause of the limited performance is the misalignment between thetranslation-specific understanding and general understanding inside LLMs. Toalign the translation-specific understanding to the general one, we propose anovel translation process xIoD (Cross-Lingual Interpretation of Difficultwords), explicitly incorporating the general understanding on the contentincurring inconsistent understanding to guide the translation. Specifically,xIoD performs the cross-lingual interpretation for the difficult-to-translatewords and enhances the translation with the generated interpretations.Furthermore, we reframe the external tools of QE to tackle the challenges ofxIoD in the detection of difficult words and the generation of helpfulinterpretations. We conduct experiments on the self-constructed benchmarkChallengeMT, which includes cases in which multiple SOTA translation systemsconsistently underperform. Experimental results show the effectiveness of ourxIoD, which improves up to +3.85 COMET.