Retrieval-Augmented Department Routing of Civil Complaints Using a Work-Item Knowledge Base

초록

Automated routing of civil complaints to administrative departments is a key technology for improving public service efficiency. However, civil complaint data in practice can be short, informal, and tied to specific regional and temporal contexts, while administrative organizations and the work items each department handles are continuously revised. Conventional text classifiers that learn a fixed mapping from complaint text to a fixed set of department labels are therefore difficult to maintain. This paper proposes a retrieval-augmented department-routing framework that does not predict departments directly. Instead, it retrieves relevant work items from a work-item knowledge base, re-ranks the candidates with a cross-encoder and a local large language model, and aggregates them at the department level to return a top-k recommendation. Because departments are linked to retrieved work items rather than learned as labels, the system can reflect organizational reorganization by updating the knowledge base, can support routing under department-level imbalance by relying on work-item descriptions rather than department-label frequency alone, and provides retrieval-grounded explanations for each recommendation. We further introduce a cascade reranking scheme that combines the cross-encoder with the local language model. On a real civil-complaint data set from the city of Changwon, Republic of Korea, enriching the work-item knowledge base with task keywords improves top-five accuracy from 37.4 percent to 80.8 percent over a taxonomy-only baseline. Under a synthetic transfer-of-duties scenario, simply updating the knowledge base raises top-five accuracy from 2.2 percent to 82.1 percent on affected complaints without any model retraining. © 2026 IEEE.

키워드