To enhance domain adaptation in transfer learning,we propose a novel approach based on adapter fusion to integrate multi-domain knowledge.Initially,base adapters are trained separately across various domains to learn domain alignment and task representation information.Subsequently,these adapters,excluding the one from the target domain,are fused to transfer multi-domain knowledge for various tasks in the target domain.Experimental results on the multi-domain text sentiment classification AMAZON dataset and the multi-genre natural language inference MNLI dataset show that our approach achieved top F1 scores of 85.79 and 76.68,respectively.Furthermore,in the more challenging domain generalization scenario,F1 scores of 85.96 and 76.74 were achieved.The proposed method effectively enhances domain adaptation and generalization capabilities in transfer learning.
关键词
迁移学习/域适应/域泛化/适配器
Key words
transfer learning/domain adaptation/domain generalization/adapter
维普
万方数据