适用于多领域少样本的元适配器整合学习方法  

Meta-Adapter Integration Learning Approach for Multi-Domain Few-Shot Data

作　　者：于信 马廷淮[1] 彭可兴 贾莉 蒋永溢 YU Xin;MATinghuai;PENG Kexing;JIA Li;JIANG Yongyi(School of Software,Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Computer Science,Nanjing University of Information Science&Technology,Nanjing 210044,China;Reading Academy,Nanjing University of Information Science&Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学软件学院,南京210044 [2]南京信息工程大学计算机学院,南京210044 [3]南京信息工程大学雷丁学院,南京210044

出　　处：《计算机工程与应用》2025年第5期122-133,共12页Computer Engineering and Applications

基　　金：国家自然科学基金(62372243)。

摘　　要：针对多域少样本文本摘要任务中迁移学习面临的诸多挑战,尤其是源域数据的多样性以及目标域数据的数据稀缺性问题,提出了一种创新的学习方法,名为元适配器整合学习方法(meta-adapter integration learning,MAIL)。MAIL使用基于Transformer的预训练模型作为基础模型,融合适配器模块限制模型参数及层数,并采用元学习方法微调适配器。此外,为了增强在不同领域间的迁移和泛化能力,设计了一种元适配器整合算法,旨在最大化利用多域信息,增强模型跨领域泛化能力。实验结果显示,MAIL在标准文本生成评价指标上超越现有主流模型,并能有效应对跨领域迁移中常见的灾难性遗忘、任务干扰和训练不稳定等问题。To overcome the limitations of transfer learning caused by the diversity of source domain data and the lack of target domain data in multi-domain few-shot text summarization tasks,a learning method based on meta-adapter integration learning(MAIL)approach is proposed.MAIL employs a Transformer-based pre-trained model as the foundational model,incorporating an adapter module to constrain model parameters and layer depth,and uses meta-learning to finetune the adapters.Furthermore,a meta-adapter integration learning algorithm is designed to maximize the use of information across multiple domains and enhance the model’s cross-domain generalization capabilities.Experimental results show that MAIL exceeds existing mainstream models on standard text generation evaluation metrics and can effectively address common issues in cross-domain transfer,such as catastrophic forgetting,task interference,and training instability.

关 键 词：文本摘要生成 少样本学习 迁移学习 预训练模型 适配器 元学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]