Improving Performance in AI-based Automatic Classification through Feature Augmentation: A Case Study of KDC

Abstract

The objective of this study is to empirically examine the performance variations of an AI-based Korean Decimal Classification(KDC) automatic classification model through the augmentation of classification features, aiming to identify strategies that improve the consistency and accuracy in automated subject cataloguing of classification numbers. Experiments were conducted using 5,882 bibliographic records, where metadata from the library domain were supplemented with publishing metadata by integrating independent attributes from both sources. Core features(title, author) and KDC extracted from the National Library of Korea’s database were enriched with external features(keywords, book summary, tables of contents) collected from the Korea Publication Industry Promotion Agency’s BNK database. Feature composition was organized into three sets: Feature Set A(title, author), Feature Set B(title, author, keywords), and Feature Set C(title, author, keywords, book summary, tables of contents). Multi-class classification models based on KLUE-BERT were developed for each set, and their performance variations were systematically analyzed. The findings demonstrate that feature enrichment resulted in progressive improvements across all KDC main classes. The Arts(6XX) class exhibited the most substantial improvement, with a 124.24% increase in the F1-score from Feature Set C to Feature Set A. Significant gains were also observed in several other classes, including Science and...