HiMoE-CCTC: A Hierarchical Mixture-of-Experts Framework for Citizen Complaint Text Classification

Citizen complaint text classification is a critical task in smart city governance, yet manual annotation suffers from inefficiency and inconsistency, particularly given the hierarchical taxonomies used to organize complaint categories. Existing approaches neglect this hierarchical structure and lack specialization for diverse complaint domains. To address these issues, we propose a Hierarchical Mixture-of-Experts (HiMoE-CCTC) framework that explicitly models taxonomy through cascading layers with progressive knowledge sharing. Each lower level integrates inherited knowledge from preceding higher layers with complementary insights from the shared expert pool through level-specific gating, enabling progressive refinement from general to specific categories. Experiments on real-world 12345 hotline citizen complaint data demonstrate that HiMoE-CCTC significantly outperforms state-of-the-art baseline models, achieving improvements of 1.30% in Micro-F1 and 5.40% in Macro-F1, with superior accuracy and hierarchical consistency for intelligent urban management systems.