关键词: 人工智能, 精准荐读, 高校图书馆, 馆员复核, 阅读推广, 试点设计

Abstract:

[Purpose/Significance] This study addresses the following issues in university-library reading promotion: scattered thematic resources, coarse demand identification, insufficient operational constraints on AI-generated content, and an evaluation system dominated by dissemination counts. It proposes an AI-assisted precision reading recommendation mechanism that can be implemented with existing systems. The study addresses the need to connect reading promotion with learning support in the context of developing public reading services and educational digitalization. Precision is defined as the degree to which content aligns with verified themes and reader-selected scenarios. The contribution lies in specifying the relationship between AI processing, local resource evidence, librarian responsibility, and service evaluation within a manageable workflow. [Method/Process] A case study and process modelling were combined. Shaanxi Normal University Library was selected for its established reading-promotion activities, thematic collections, information-literacy education, learning spaces, and digital service entry points. The 2025 reading report and aggregate business statistics are primary sources for examining service foundations and operational constraints. The research and practice context of published research and institutional accounts of AI applications were reviewed. The exploratory single-case analysis links service foundations, operational constraints, mechanism stages, and observable indicators, distinguishing documented service conditions from proposed workflow requirements. Available resources and entry points were first identified; gaps in resource fields, review rules, and evaluation records were then translated into workflow requirements. Process modelling specifies responsible roles, inputs, outputs, control points, and the handling of rejected or unavailable results. Development-stage prototype checks retain demand inputs, candidate snapshots, raw model responses, validation outcomes, and librarian amendments to examine how generation and review are connected. Aggregate statistics describe existing service conditions; reader-facing outcomes require a subsequent service trial. [Results/Conclusions] The proposed framework comprises a five-stage business chain: thematic resource organization, scenario-based demand identification, AI-assisted generation, librarian review and reading guidance, and evaluation with feedback. Collaborative quality control, data governance, and exception handling support the chain. The AI component uses an instruction-following large language model capable of processing Chinese resource descriptions and producing structured output. The library-controlled backend selects verified candidate records, assembles task instructions and supporting descriptions, calls an approved model service, and validates the returned fields. AI classifies resources, proposes reading order, and drafts recommendation reasons grounded on the supplied descriptions. Bibliographic details and access links are retrieved through resource identifiers. Structural validity alone cannot establish factual accuracy, so librarians still review the evidence and suitability of each draft. With a small candidate pool, verified records can be supplied directly. The need for additional semantic retrieval or retrieval-augmented generation should be evaluated when resources and feedback accumulate. A micro-pilot on introductory AI-literacy reading is designed. Librarians first screen ten to fifteen resources and organize materials around basic concepts, tool use, educational applications, and ethical governance. One principal entry point and an observation window of two to four weeks limit the initial scope. A development prototype has produced logged outputs through the DeepSeek API and a local Qwen3-1.7B model, with librarian's review documented for a demonstration set. In local development, a missing mandatory review field was rejected by backend validation; a subsequent output passed structural checks after a JSON Schema constraint was introduced. Manual review of the demonstration recommendations replaced general suitability statements with specific reading tasks and clarified the intended use of different resources. These observations concern workflow checks, while reader-facing performance data are not yet available. Before release, normal requests, insufficient evidence, invalid resource identifiers, and interface failures should be checked. Evaluation records AI-output approval and revision, librarian preparation and review time, recommendation reach, resource-detail visits, borrowing or access intentions, and both positive and negative feedback. Rates require consistent denominators and observation windows, and session counts must not be interpreted as numbers of individual readers. The study provides a low-barrier, auditable starting point for university libraries without mature recommendation platforms. Further work should extend prototype checks, complete small-scale service trials, compare manual and AI-assisted workflows using the same theme, candidate pool, task requirements, and service entry point, and examine content quality, total workload, and applicability across themes. The present design has not established improvements in borrowing, resource use, or reading quality.

Key words: artificial intelligence, precision reading recommendation, university library, librarian review, reading promotion, pilot study