Transfer learning guided by cultural distance enables effective cross cultural content adaptation

Abstract

Artificial intelligence-driven multilingual text generation technology offers new approaches for cross-cultural content dissemination. However, existing models still struggle to effectively handle cross-cultural differences in value orientations, narrative conventions, and pragmatic preferences during content adaptation, which often leads to communication losses in the target culture. To address this problem, this paper proposes a transfer learning-driven cultural discount reduction network (TL-CDRNet). Built upon a pretrained multilingual encoder-decoder Transformer, the model incorporates a shared-private transfer adaptation module, a cultural distance guidance module, a transferability filtering module, and a transfer-feedback cultural discount optimization module, enabling synergistic optimization of cross-cultural knowledge transfer and target-cultural expression adaptation. Experiments are conducted on three tasks: film synopsis adaptation, tourism promotional text rewriting, and cultural creative/intangible cultural heritage product description adaptation. Results show that TL-CDRNet outperforms baseline methods in terms of BLEU, ROUGE-L, BERTScore, and semantic consistency, reducing the cultural discount score to 0.298 and increasing the cultural discount reduction rate to 43.7%. Moreover, the model exhibits satisfactory stability and generalization ability under low-resource and cross-domain conditions. This study demonstrates that transfer learning-based AI models can more effectively achieve cross-cultural content adaptation while preserving original semantics, offering a new technical pathway for intelligent text generation in cultural communication scenarios.