Optimization and application of deep learning algorithm in library information retrieval system

The library information retrieval system (IRS) has exposed some problems, such as insufficient semantic understanding and low retrieval efficiency, in the face of massive digital resources and diverse user needs. In order to solve these problems, this article proposes an improved scheme of library IRS based on deep learning (DL) algorithm optimization. Transformer model and its multi-head self-attention mechanism are introduced to improve the semantic understanding ability and retrieval accuracy of the system. This article constructs a high-quality training set by preprocessing library collection data, designs optimization algorithms based on the Transformer model, and assesses the optimization effect by comparing the university library dataset with the traditional Term Frequency Across Document Frequency (TF-IDF) algorithm. The results show that the retrieval accuracy of the optimized algorithm is improved by 18.55% compared with the traditional method, the average response time is shortened from 3.8 seconds to 1.5 seconds, and the user satisfaction score reaches 4.6 (out of 5). The optimization algorithm also performs well in comprehensive indexes such as recall and F1 value, which proves its stability and robustness in complex query scenarios.