Search papers, labs, and topics across Lattice.
This paper introduces the Sentence Translation Alignment Rate (STAR), a novel metric designed to quantify sentence-level structural fidelity in document-to-document machine translation. By implementing STAR-masked Preference Optimization (StarPO), the authors optimize translation quality by focusing on misaligned segments, leading to improved global coherence and structural integrity in translations. Experimental results show that StarPO enables smaller models to outperform larger proprietary systems like GPT-4o, highlighting its effectiveness in enhancing translation performance while maintaining token efficiency.
Compact models can now outperform massive systems like GPT-4o in document translation by leveraging a new metric that prioritizes structural fidelity.
Large Language Models (LLMs) have enabled a shift from sentence-level to document-to-document (Doc2Doc) machine translation, promising improved global coherence. However, document-to-document generation in a single pass frequently suffers from structural misalignment, manifesting as sentence omissions or hallucinations that violate the core requirement of source-target correspondence. To address this, we introduce Sentence Translation Alignment Rate (STAR), an auxiliary metric that explicitly quantifies sentence-level structural fidelity. Building on this, we propose STAR-masked Preference Optimization (StarPO), a framework that ranks document-level hypotheses by structural quality and utilizes a dynamic alignment mask to focus optimization on misaligned segments. Experimental results across news and literary domains demonstrate that StarPO significantly enhances translation quality and structural integrity. Notably, StarPO allows compact models to surpass the performance of massive proprietary systems like GPT-4o while maintaining superior token efficiency.
