Hierarchical Framework for Automated Enterprise Document Classification

Citations

WEB OF SCIENCE

0

초록

Enterprises generate and share large volumes of documents through cloud storage and collaboration platforms. Indiscriminate retention increases storage costs, whereas uncontrolled sharing raises the risk of data leakage and insider misuse, creating a practical need for large-scale security-grade classification. Manual labeling is costly, error-prone, and often inconsistent across organizations. This study proposes a hierarchical multitask learning framework for enterprise document security classification that jointly predicts three security grades (classified, sensitive, and open) along with the originating department. Each sentence is encoded using a finetuned KLUE-RoBERTa-Base model. Additionally, a lightweight document-level transformer with attention pooling aggregates sentence representations for document-level prediction. This architecture supports long-document processing beyond the 512-token input limit of encoder-based models, replacing heuristic voting with a learnable aggregation mechanism. Experiments on 635 real-world enterprise documents from four departments revealed that the proposed framework achieved a mean combined macro-F1 of 89.30% under stratified 10-fold cross-validation with three random seeds and outperformed GPT-4.1 zero-shot classification. Attention weights provide sentence-level interpretability by identifying sentences that most strongly influence the predictions of the model. These results indicate that the framework offers a practical, interpretable basis for policy-aligned document governance in enterprise environments under real-world operational constraints.

키워드

Document Classification; Information Sensitivity; Access Control Policy; Hierarchical Transformer; Multitask Learning
제목
Hierarchical Framework for Automated Enterprise Document Classification
저자
Choi, Yeji; Chang, Hangbae; Lee, Jung A.
DOI
10.22967/HCIS.2026.16.067
발행일
2026-11
유형
Article
저널명
Human-centric Computing and Information Sciences
권
16