| 摘 要: 为了解决多模态数字图书馆跨域检索中的信息碎片化和模态不统一问题,提出一种基于深度迁移学习的跨域检索算法。首先,利用BERT预训练模型和SWTR模型提取多模态资源特征,通过CNN网络融合,以解决资源的跨域差异;其次,采用深度迁移学习将融合特征迁移至数字图书馆资源域,得到各特征对应资源的类标签;最后,根据查询本体与类标签的词形、语义相似度构建索引,实现精准跨域检索。实验结果表明,该算法在特征提取和资源检索准确性方面均表现优异。 |
| 关键词: 深度迁移学习 多模态数字图书馆 资源跨域检索 CNN网络 相似度 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 江苏省高校哲学社会科学研究项目(2022SJYB1499);江苏省高校图工委教改研究项目(2024JTYB01);苏州市图书馆学会研究项目(25-C-008) |
|
| Deep Transfer Learning-Based Cross-Domain Retrieval Algorithm for Multimodal Digital Library Resources |
|
LIU Feng
|
(Suzhou University of Technology Library, Suzhou 215500, China)
liuf@szut.edu.cn
|
| Abstract: To address information fragmentation and modal heterogeneity in cross-domain retrieval of multimodal digital libraries, this paper proposes a cross-domain retrieval algorithm based on deep transfer learning. Firstly, BERT and SWTR pre-trained models extract multimodal resource features, integrated via CNN to resolve cross-domain discrepancies. Secondly, deep transfer learning migrates fused features to the digital library domain, generating class labels for resources. Finally, indices are constructed based on lexical and semantic similarity between query ontologies and class labels, achieving precise cross-domain retrieval. Experimental results demonstrate the algorithm’s excellence in feature extraction and retrieval accuracy. |
| Keywords: deep transfer learning multimodal digital library resource cross-domain retrieval CNN similarity |
