[4K4-GS-6b] Language media processing

2:00 PM - 2:15 PM JST(5:00 AM - 5:15 AM UTC)

[4K4-GS-6b-03] Design and Evaluation of an LLM-based Dialogue Control Agent for Career Interviews with Ambiguous Job-Seeker Responses

〇Masato Hidaka¹, Tetsu Sato¹, Naohiro Ikahata¹ (1. PERSOL CAREER CO.,LTD.)

Keywords:

Dialogue processing and dialogue systems,Dialogue control,Large language model (LLM),Career support and interview support,Conversational agents

This paper proposes an LLM-based dialogue control framework for interview-style dialogues to collect qualitative information such as reasons, background, constraints, and priorities. The framework is applicable to qualitative interviews in general, and here we take initial career interviews as a case study. Conventional LLM chatbots often lack explicit mechanisms to control the conversation flow, which leads to missing or repeated questions, insufficient probing, and increased user burden. Our method defines, for each information item, a rubric with priority and depth levels, and a dialogue control agent dynamically selects the next item and depth based on the current acquisition state. We further construct a job-seeker agent that simulates indecisive and ambiguous responses by generating personas from mock interview logs and injecting perturbations such as vague wording and fluctuating conditions according to a perturbation rate r. The value of rrr is estimated from manual labels of perturbation events in the logs, reflecting the observed frequency of indecisive behaviours. Simulation experiments compare the proposed framework with a free-form LLM baseline and show that our method can reduce redundant questions and the number of dialogue turns , indicating that the framework supports short-turn interviews suitable for practical deployment.