Answers are grounded in this paper and cite their sources.
Translation quality assessment requires localized error feedback, semantic evaluation, and sensitivity to document context and application domain. This study evaluates a shared-encoder framework integrating word-level error detection, sentence-level semantic alignment, document-level modeling, source-associated representation learning, and MAML-based domain adaptation. Three research questions examine multi-granularity assessment, cross-task representation learning, and adaptation across support-set sizes. On the official WMT23 MQM-scored subsets, the full model reaches Pearson correlations of 0.769 for Chinese–English sentences, 0.781 for English–German paragraphs, and 0.762 for Hebrew–English sentences. The corresponding numerical gains over xCOMET are 0.023–0.024 correlation points. Human–machine classification accuracy is 92.7% in distribution and 86.4% under leave-one-MT-system-out evaluation. On 2000 machine-origin post-edited translations, machine-class recall is 71.3%. Frozen-representation probes recover four linguistic properties with instance-level R 2 values of 0.52–0.72. With 50 labeled support items and 500 labeled validation items per domain, the model reaches correlations of 0.778, 0.765, and 0.748 for medical, legal, and literary translation, respectively. Token-level error-detection F1 ranges from 0.748 to 0.766. Component ablations link the quality-assessment and source-classification tasks through shared representations, while the adaptation results quantify performance under a fixed target-domain model-selection protocol.
No takes yet. Share an insight, caveat, or question.
Lv et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: