지방정부 복지 보도자료에 대한 확률 및 문맥 기반 토픽 분석

Abstract

This study analyzes the communication methods and topic structures of local government welfare policies through topic modeling of welfare-related press releases issued by the Gyeongsangnam-do Provincial Government. A total of approximately 9,800 documents published from 2016 to 2024 were examined, categorized into pre-COVID-19, outbreak, and post-pandemic periods. Two topic modeling techniques were applied: probabilistic Latent Dirichlet Allocation (LDA) and context-based BERTopic. LDA effectively identified structured policy changes such as youth employment, infectious disease response, and welfare blind spot identification, while BERTopic captured contextual flows and thematic interconnections within welfare messages. Both models revealed clear diversification and refinement of topics in the post-pandemic period, indicating a multi-layered transformation of welfare policy communication. This study provides a systematic comparison of probabilistic and contextual topic models in analyzing Korean local government policy communications, offering methodological insights into how different approaches capture changes in policy discourse during crises, and highlighting both theoretical and practical implications for public communication research and policy messaging.