Abstract: 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 fine-tune 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.

Key words: text summarization generation, few-shot learning, transfer learning, pre-trained model, adapter, meta learning

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。

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