Multilingual Cultural Metaphor Transfer Driven by Federated Learning

Traditional federated learning faces significant challenges in multilingual cultural metaphor transfer because nonindependent and identically distributed language data make it difficult to balance local cultural heterogeneity with global semantic consistency. Privacy protection mechanisms may further amplify cultural cognitive bias and reduce transfer quality. To address these issues, this paper proposes a culturally aware personalized federated learning architecture. Based on Hofstede’s cultural dimension theory, cultural feature vectors are constructed, cultural weight functions are defined, and a weighted aggregation strategy is introduced to account for variations in power distance and individualism across clients. Security is enhanced through differential privacy and secure aggregation mechanisms. Experimental results show that, compared with FedAvg, the proposed method improves cross-language metaphor recognition precision and F1-score by 7.4% and 7.1%, respectively. Cultural consistency reaches 0.82, transfer comprehensibility reaches 4.02, and user matching degree reaches 0.79, while the minimum back-inference attack success rate is reduced to 0.5%. The results confirm that the proposed architecture enhances multilingual semantic transfer, cross-cultural consistency, and privacy-preserving collaborative learning.