Scholarlink

<p>With the expanding application of large language models (LLMs), cross-cultural metaphor comprehension has become a crucial problem in knowledge-enhanced semantic analysis. This task requires identifying metaphors, recovering source-to-target conceptual mappings, and generating culturally plausible explanations. However, existing studies often reduce it to cross-lingual transfer or semantic matching, lacking explicit structural modeling of the “cultural conditions–source–target” relationship. Moreover, models easily rely on superficial cues such as lexical overlap, which may limit their robustness in the evaluated cross-cultural settings. To address these issues, we propose CCGA-FT (Cross-Cultural Graph Alignment Fine-Tuning), a knowledge-graph-enhanced fine-tuning method for cross-cultural metaphor comprehension. It integrates three functional modules: cross-cultural graph alignment reasoning to recover culturally constrained source–target mappings, anti-superficial constraint learning to reduce reliance on lexical shortcuts, and knowledge-conditioned low-rank adaptation to inject key cultural paths into parameter updates. Experiments on two representative benchmarks, Meta4XNLI and LCC, show that CCGA-FT achieves consistent improvements in the evaluated settings, especially in overall detection performance and structural mapping recovery. Specifically, it achieves Macro-F1 scores of 79.0 ± 0.6 and 75.9 ± 0.6 on Meta4XNLI and LCC, respectively, alongside 57.3 ± 0.6 Hit@1 on LCC and 0.864 ± 0.005 BERTScore on Meta4XNLI. Robustness and ablation experiments confirm that each component contributes to overall stability and performance. Rather than claiming universal generalizability across all cultural or low-resource scenarios, this study provides initial evidence for a structural mapping approach to cross-cultural metaphor comprehension and offers a verifiable framework for knowledge-enhanced LLMs in the evaluated benchmark settings.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Knowledge graph guided learning for cross cultural english metaphor understanding in large language models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Xiaoyi Liu

摘要

With the expanding application of large language models (LLMs), cross-cultural metaphor comprehension has become a crucial problem in knowledge-enhanced semantic analysis. This task requires identifying metaphors, recovering source-to-target conceptual mappings, and generating culturally plausible explanations. However, existing studies often reduce it to cross-lingual transfer or semantic matching, lacking explicit structural modeling of the “cultural conditions–source–target” relationship. Moreover, models easily rely on superficial cues such as lexical overlap, which may limit their robustness in the evaluated cross-cultural settings. To address these issues, we propose CCGA-FT (Cross-Cultural Graph Alignment Fine-Tuning), a knowledge-graph-enhanced fine-tuning method for cross-cultural metaphor comprehension. It integrates three functional modules: cross-cultural graph alignment reasoning to recover culturally constrained source–target mappings, anti-superficial constraint learning to reduce reliance on lexical shortcuts, and knowledge-conditioned low-rank adaptation to inject key cultural paths into parameter updates. Experiments on two representative benchmarks, Meta4XNLI and LCC, show that CCGA-FT achieves consistent improvements in the evaluated settings, especially in overall detection performance and structural mapping recovery. Specifically, it achieves Macro-F1 scores of 79.0 ± 0.6 and 75.9 ± 0.6 on Meta4XNLI and LCC, respectively, alongside 57.3 ± 0.6 Hit@1 on LCC and 0.864 ± 0.005 BERTScore on Meta4XNLI. Robustness and ablation experiments confirm that each component contributes to overall stability and performance. Rather than claiming universal generalizability across all cultural or low-resource scenarios, this study provides initial evidence for a structural mapping approach to cross-cultural metaphor comprehension and offers a verifiable framework for knowledge-enhanced LLMs in the evaluated benchmark settings.