In response to the current surge in digital library document resources, which has led to problems such as inefficient traditional classification methods, inaccurate subject feature extraction, and insufficient personalization of recommendation systems, this paper aims to build an automated classification and recommendation model based on deep learning to improve the efficiency of document management. A dual-channel feature extraction framework is constructed by integrating the BERT pre-trained model and the TextCNN convolutional neural network. First, the PubMed and CiteSeerX datasets are segmented, stop words are filtered, and Word2Vec word vectorization is performed. Then, the long-distance semantic dependencies are captured through the 12-layer Transformer structure of BERT. Three scales of TextCNN convolution kernels (3 × 300, 5 × 300, and 7 × 300) are used in parallel to extract local text features. After fusion in the feature splicing layer, they are connected to the fully connected layer to realize multi-label classification. A collaborative filtering recommendation model is constructed based on user behavior data, and the cross entropy loss function and Adam optimizer are used for parameter update. Experimental results show that the hybrid model has an average recall rate of 90.98% and an average F1 value of 89.46% on a test set of 200,000 documents. The click-through rate of the recommendation system in the test reaches 32.6%, confirming the significant advantages of deep learning technology in improving document organization efficiency and personalized service quality, and providing an effective technical path for the intelligent transformation of digital libraries.