Scholarlink

<p>Predicting user satisfaction for chatbots in customer service operations is important for their successful uptake. Based on chatbot conversation logs and corresponding satisfaction scores from a much used intent-based customer service chatbot, we developed models for predicting user satisfaction on the basis of conversation log data. We found significant covariation between satisfaction and conversation characteristics reflecting in the log data, suggesting efficient chatbot interactions. We found a prediction model including data on generic conversation characteristics, such as the number of user messages and predicted intents, to explain 10% of the variation in user satisfaction. A model also including domain-specific information in addition to generic conversation characteristics, specifically on the types of predicted chatbot intents, explained 27% of the variation. Substantial variation in prediction model performance was identified between different areas of support, suggesting the need to tailor prediction models to different areas of support provided. We conclude by discussing the implications of our findings and suggesting further research.</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

Conversation logs as a source of insight: predicting user satisfaction for customer service chatbots

  • 期刊论文

  • 发表日期：

    2025年4月4日

查看全文
  • Fehime Ceren Ay,
  • Eleonora Freddi,
  • Asbjørn Følstad,
  • Stig Hodnebrog,
  • Knut Kvale,
  • Olav Alexander Sell,
  • Simen Ulsaker

摘要

Predicting user satisfaction for chatbots in customer service operations is important for their successful uptake. Based on chatbot conversation logs and corresponding satisfaction scores from a much used intent-based customer service chatbot, we developed models for predicting user satisfaction on the basis of conversation log data. We found significant covariation between satisfaction and conversation characteristics reflecting in the log data, suggesting efficient chatbot interactions. We found a prediction model including data on generic conversation characteristics, such as the number of user messages and predicted intents, to explain 10% of the variation in user satisfaction. A model also including domain-specific information in addition to generic conversation characteristics, specifically on the types of predicted chatbot intents, explained 27% of the variation. Substantial variation in prediction model performance was identified between different areas of support, suggesting the need to tailor prediction models to different areas of support provided. We conclude by discussing the implications of our findings and suggesting further research.