Library Resource Retrieval via Joint Feature Embedding and Cross-Modal Hashing Optimization Algorithms

In the context of library informatization, this study proposes a library resource retrieval framework based on joint feature embedding and cross-modal hash optimization. This framework can solve the retrieval scenarios of multimodal, high-dimensional heterogeneous data in digital libraries. This framework integrates text semantics, structure graphs and user behaviors into a unified embedding space to achieve efficient and accurate cross-modal retrieval. Specifically, this study first constructs a baseline retrieval model based on TF-IDF vectorization and RankNet ranking optimization; then introduces a joint feature embedding mechanism. This mechanism can fuse multimodal features into a low-dimensional semantic space to enhance representation discriminability. Finally, this study proposes a cross-modal hash optimization algorithm to compress high-dimensional features into compact binary codes while maintaining semantic consistency. Experimental results on a multimodal library dataset containing 3,000 text-image-user behavior data directly show that this method is significantly better than the comparison model in terms of mAP (83.7%), NDCG@20 (85.2%) and Top-10 accuracy (88.5%). Ablation experiments further verify the contribution of each modality, with text, image and behavior data bringing mAP improvements of 9.2%, 7.6% and 5.3% respectively.