Facilitating Online Petition Topic Categorization in South Korea: A BERT-Based Approach

Online petition systems facilitate citizen engagement with government agencies by providing a platform for direct expression of policy concerns. The increasing volume and diverse topical range of submissions, however, present significant challenges in routing petitions to the relevant departments in public administration. This study presents an automated approach for topic categorization of online petitions in South Korea utilizing a BERT-based framework. The methodology integrates data from both legacy (Blue House) as well as current (National Assembly) online petition platforms and fine-tunes KR-BERT, a Korean-specific BERT variant to capture domain-specific linguistic features. The study employs selective layer freezing to balance computational efficiency with domain adaptation. Experiments conducted demonstrate that models trained on combined historical and recent data outperform those relying exclusively on data from the currently used platform. The bestperforming configuration achieved 76% accuracy and an F1score of 0.75 over the 12 petition categories. Although there are inherent challenges in aligning data from older systems and platforms with newer ones, the findings suggest that incorporating historical data and implementing tailored layer-freezing strategies can considerably enhance categorization outcomes in resource-constrained public-sector environments.