To solve structural misalignment and hallucinations in document-to-document translation, the authors introduce the Sentence Translation Alignment Rate (STAR) metric. The proposed StarPO framework uses dynamic alignment masks during preference optimization to prioritize fixing misaligned segments, enhancing global coherence in news and literary text.
HOW THIS AFFECTS YOU
●
builderThis provides a method to reduce hallucinations in document-scale translation tasks.
●
researcherYou can use STAR as a more granular metric for evaluating structural integrity in long-form translation.