Scholarlink

<p>Chatbots serve as a valuable tool for enhancing customer interactions. The traditional debt collection process is time-consuming and requires significant human effort. In this research, we have assessed the effectiveness of proposed chatbot systems in optimizing the debt collection process. Additionally, this study examined the customer risk category based on payment history using Deep Reinforcement Learning. This research developed a chatbot using prompt engineering and handled the hallucinated response using Parameter Efficient Fine-Tuning techniques (PEFT). Based on the customer risk category, the proposed model offers flexible payment options via the chatbot model. Crafted a customized prompt that consists of sample debt collection interactions, rules, and regulations of the debt collection procedure. When used for prolonged periods, the model may occasionally produce irrelevant responses. In order to address this issue, we have created a customized dataset that captures the preferred human response associated with the irrelevant response. This customized dataset was trained using PEFT to generate the relevant response. A comparative analysis was made between GPT 3.5 and Llama 2 7B chat model through prompt engineering and statistically verified using a t-test. The experimental results showed that there was no significant variation in the chatbot performance between the two Large Language models. The proposed credit risk analysis provides a higher 4.35% classification accuracy than the existing algorithms, which was verified using hypothesis testing. Statistical analysis showed that LoRA fine-tuning resulted in a 25.48% improvement in the ROUGE score compared to the baseline model. This research not only compares multiple LLM-based chatbots but also evaluates them with existing chatbot models in the literature. The Large Language model-based chatbot provides more significant results than the existing chatbot for the daily dialog dataset, with a 61.39% improvement in BLEU score and 129% improvement in ROUGE.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Building an effective chatbot for optimizing the debt collection process through large language models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月11日

查看全文
  • Keerthana Sivamayilvelan,
  • Elakkiya Rajasekar,
  • Subramaniyaswamy Vairavasundaram,
  • Santhi Balachandran,
  • Ketan Kotecha,
  • Ambarish Kulkarni

摘要

Chatbots serve as a valuable tool for enhancing customer interactions. The traditional debt collection process is time-consuming and requires significant human effort. In this research, we have assessed the effectiveness of proposed chatbot systems in optimizing the debt collection process. Additionally, this study examined the customer risk category based on payment history using Deep Reinforcement Learning. This research developed a chatbot using prompt engineering and handled the hallucinated response using Parameter Efficient Fine-Tuning techniques (PEFT). Based on the customer risk category, the proposed model offers flexible payment options via the chatbot model. Crafted a customized prompt that consists of sample debt collection interactions, rules, and regulations of the debt collection procedure. When used for prolonged periods, the model may occasionally produce irrelevant responses. In order to address this issue, we have created a customized dataset that captures the preferred human response associated with the irrelevant response. This customized dataset was trained using PEFT to generate the relevant response. A comparative analysis was made between GPT 3.5 and Llama 2 7B chat model through prompt engineering and statistically verified using a t-test. The experimental results showed that there was no significant variation in the chatbot performance between the two Large Language models. The proposed credit risk analysis provides a higher 4.35% classification accuracy than the existing algorithms, which was verified using hypothesis testing. Statistical analysis showed that LoRA fine-tuning resulted in a 25.48% improvement in the ROUGE score compared to the baseline model. This research not only compares multiple LLM-based chatbots but also evaluates them with existing chatbot models in the literature. The Large Language model-based chatbot provides more significant results than the existing chatbot for the daily dialog dataset, with a 61.39% improvement in BLEU score and 129% improvement in ROUGE.