Abstract
In recent years, with the rapid development of deep learning, big data mining, and natural language processing (NLP) technologies, the application of NLP in the field of recommendation systems has attracted significant attention. However, current text recommendation systems still face challenges in handling word distribution assumptions, preprocessing design, network inference models, and text perception technologies. Traditional RNN neural network layers often encounter issues such as gradient explosion or vanishing gradients, which hinder their ability to effectively handle long-term dependencies and reverse text inference among long texts. Therefore, this paper proposes a new type of depth-aware neural network recommendation model (Hourglass Deep-aware neural network Recommendation Model, HDARM), whose structure presents an hourglass shape. This model consists of three parts: The top of the hourglass uses Word Embedding for input through Finetune-Bert to process text embeddings as word distribution assumptions, followed by utilizing bidirectional LSTM to integrate Transformer models for learning critical information. The middle of the hourglass retains key features of network outputs through CNN layers, combined with pooling layers to extract and enhance critical information from user text. The bottom of the hourglass avoids a decline in generalization performance through deep neural network layers. Finally, the model performs pattern matching between text vectors and word embeddings, recommending texts based on their relevance. In experiments, this model improved metrics like MSE and NDCG@10 by 8.74% and 10.89% respectively compared to the optimal baseline model.