A Large Language Model-Based Framework for Multistage Child Counseling and Abuse Detection = 아동 상담 및 학대 탐지를 위한 대규모 언어 모델 기반 다단계 프레임워크

다국어 초록 (Multilingual Abstract)

Child abuse severely disrupts the healthy growth and development of children, resulting in long-term physical as well as emotional consequences. However, in real-world, despite the continuous increase in reported abuse cases, the chronic shortage of certified child- abuse professionals has significantly increased the workload of individual counselors, making timely intervention increasingly difficult. To reduce counselors’ workload and enhance the efficacy of child abuse detection, we propose Conversational AI for Child Abuse Detection (CACAD). CACAD utilizes a large language model (LLM) to conduct counseling and detect four types of child abuse: neglect, emotional, physical, and sexual. During the question generation process, the LLM serves as the primary agent, supported by two auxiliary modules. In the process of the counseling, abusive question detection module filters out harmful questions to protect children. And the next question category prediction module selects the most appropriate question to guide the counseling flow more flexibly. In addition, during abuse prediction, CACAD leverages uncertainty quantification to dynamically flag cases as pending review for counselor confirmation. Experimental results using a Korean child-and-adolescent counseling dataset show that CACAD achieves an exact match of 0.907 and a macro-F1 score of 0.939 in predicting child abuse categories. Furthermore, in human evaluation, CACAD demonstrated highly reliable performance in counseling sessions. These findings highlight the potential of LLM-based conversational agents to effectively support detection of child abuse through counseling in real-world.

목차 (Table of Contents)

• Ⅰ. INTRODUCTION 1
• II. Related Works 6
• 2.1 Conversational AI In Counseling 6
• 2.2 Child Abuse Detection 7
• III. Dataset Construction 8
• 3.1 Next Question Category Prediction Dataset 9
• 3.2 Abusive Question Detection Dataset 10
• 3.3 Child Abuse Detection Dataset 11
• Ⅳ. Methodology 13
• 4.1 Enhancing Counseling Through Question Flow and Safety Control 14
• 4.1.1 Next Question Category Prediction 14
• 4.1.2 Category Prediction and Follow-up Counseling 14
• 4.1.3 Abusive Question Detection 16
• 4.2 Instruction Tuning and Uncertainty-Aware Detection 17
• Ⅴ. Experiments 18
• 5.1 Evaluation Metrics 18
• 5.1.1 Child Abuse Counseling Evaluation 18
• 5.1.2 Child Abuse Detection Evaluation 19
• 5.2 Experimental Details 20
• 5.3 Baselines 20
• 5.4 Experimental Results 21
• 5.4.1 Evaluation of Child Abuse Detection 21
• 5.4.2 Evaluation of Child Abuse Counseling 22
• 5.4.3 Evaluation of Child Abuse Detection by Category 26
• 5.4.4 Evaluation of Classification Performance Under Uncertainty Thresholds 27
• Ⅵ. Conclusion 32
• Ⅶ Bibliography 33
• Ⅷ 국문 초록 44