A Study on a Framework for Initial Counseling for Vulnerable Populations in Welfare Blind Spots Based on LLM

Abstract

This study proposes a method to enhance the quality of counseling and automate initial counseling in welfare counseling systems by utilizing Large Language Models (LLMs) to correct errors occurring in the Speech-to-Text (STT) process. The Silver framework consists of an STT correction and evaluation model, a conversation model, and a summary model, aiming to simultaneously improve the flexibility and accuracy of counseling. In this study, the STT correction and evaluation model improved the quality of text correction, while the conversation model enabled natural conversation. Additionally, the summary model effectively organized and verified counseling content. Experimental results demonstrated that the STT correction model achieved 82 % accuracy in evaluation and 74 % accuracy in correction, while the conversation progress and summary models recorded 68 % and 92 % accuracy, respectively. These findings highlight the effective applicability of LLM-based technologies in welfare counseling. © 2025 IEEE.