Human-in-the-Loop Cultural Schema Annotation for Metaphor Translation: A Pilot Framework

Abstract

Metaphor translation presents a persistent challenge in cross-lingual communication because source and target languages often organize metaphorical meaning through different cultural schemas. This pilot study proposes a human-in-the-loop annotation framework for tracing cultural schema transfer in Chinese-English metaphor translation. The framework annotates 500 Chinese-English metaphor pairs across four dimensions: source-language cultural schema, target-language cultural schema, transfer strategy, and translator proficiency level. BERT-wwm-assisted pre-annotation is used to generate candidate labels, which trained human annotators revise and adjudicate. An exploratory CLIP-based similarity measure is used only as a heuristic indicator of schema alignment, and a Graph Neural Network models associations among schema labels, proficiency levels, and observed transfer-strategy labels. In the pilot corpus, the GNN reached 82% accuracy in predicting adjudicated transfer-strategy labels. Corpus patterns indicated that intermediate translators favored literal translation (68%) when schema similarity was low, whereas advanced translators more often used adaptation or substitution strategies (74%). These findings are presented as preliminary evidence for the feasibility of schema-aware annotation and label prediction, not as evidence of downstream translation-quality gains or broad model generalization. The study contributes a reproducible pilot framework for making cultural-schema decisions more explicit in translation pedagogy and computational annotation while identifying the validation steps required for larger-scale work.