Empowering Catalogers: Evaluating AI Tools for Multilingual Book Classification in the LCC Framework

Abstract

This study evaluates the effectiveness of four internal and external AI tools in supporting subject analysis within a multilingual bibliographic environment and the Library of Congress Classification (LCC) framework. Between January and May 2026, 550 Western-language book titles and abstracts were randomly selected for testing. LC-assigned classification numbers were used as the authoritative answer and were compared with 6,600 AI-generated recommendations using three match levels based on the hierarchical structure of the LCC. Ranking and match-level analyses revealed that the locally developed AI assistant “Lin, Wu-Lin (powered by Google Gemini)” achieved the strongest overall performance, with a Top-1 Accuracy of 48.73% and a Top-3 Accuracy of 59.45%. Although the external tool “CatalogerGPT” performed less effectively in identifying exact classification numbers, it demonstrated relatively stable subject-direction identification, achieving a Close Match Rate of 54.73% and a Broad Match Rate of 69.52%. The findings indicate that locally developed AI tools designed to follow catalogers’ subject-analysis workflows can substantially reduce the time required for preliminary classification assessment. However, because AI-generated outputs remain vulnerable to hallucinations, professional catalogers continue to play an indispensable role in final quality control and complex intellectual evaluation.