토픽모델링과 대규모 언어모델을 활용한 핀테크 결제 앱 부정 리뷰의 사용자 불만 요소 심각도 분석
This study asks which complaints are most severe in negative reviews of three leading Korean fintech payment applications (Toss, Kakao Pay, and Naver Pay). Latent Dirichlet Allocation on 14,000 Google Play reviews yielded seven topics, to which a residual category was added, and a large language model assigned one of the eight resulting categories to every review. Six of these categories, referred to as complaint types, were divided into core and supplementary services following Wirtz and Lovelock's service structure, and severity was assessed by star rating and by stated intention to discontinue use. Supplementary types ranked highest on both measures, although the type with the highest share of one-star reviews differed by application. All 1,714 one-star reviews of those types were then analyzed qualitatively, and obstruction of the core service could be determined for 1,641. Discontinuation intention was stated more often where a supplementary element obstructed use of the core service. Within such obstruction it was concentrated in the design-induced form, in which the element worked as intended but its design or policy displaced the core service (22.2% vs. 13.2% for complaints involving no obstruction; common odds ratio 2.28, 95% CI 1.51–3.44, stratified by application). Frequency and severity therefore do not coincide: the most severe complaints arise where a supplementary element's design displaces the core service.