AI-Based Book Classification Using Book Titles: Investigating Two-Stage NLP Approach

Assigning library classification codes improves shelving and browse-ability but remains time-consuming and inconsistent across institutions. This paper presents an AI-based method to predict Library of Congress Classification (LCC) classes and Subclasses from book titles only by using Project Gutenberg records. The study compares single- and two-stage prediction, where Stage-1 predicts the first letter of the LCC (AZ) (e.g., P for Language and Literature) and Stage-2 ranks twoletter LCC classes (e.g., PR and PS) within that first-letter group. We benchmark a short-text baseline (TF-IDF with a logistic regression) against a fine-tuned transformer (DistilBERT) under author-grouped train/validation/test split sets. Experiments report Top-1/Top-3 accuracy and Macro-F1, analyze class imbalance under Full and Merge-Rare label schemes, and include a focused case study on adjacent two-letter LCC classes to characterize short-title confusions. Results show that title-only classification is feasible, where the single-stage DistilBERT model achieves the strongest Macro-F1 and the twostage variant improves the baseline (especially under the Full scheme) by reducing off-class errors, but decreases the effectiveness of DistilBERT. Besides, Top-3 suggestions support human cataloging when only titles are available.