Scholarlink

Open-domain chatbots leveraging Transformer models like GPT-3 and Transformer-XL have demonstrated impressive text generation capabilities. However, their industrial applications extend beyond text generation, requiring continuous learning from user feedback while preserving data privacy. This study explores the integration of Federated Learning (FL) into chatbot training, enabling self-learning from user interactions without compromising personally identifiable information (PII). We implement intent classification models using both traditional centralized learning and FL-based approaches, comparing their effectiveness. Additionally, we propose a Self-Feeding mechanism to address post-deployment data drift and mitigate training bias. A comprehensive chatbot architecture integrating FL and self-learning is presented, along with an evaluation of various neural network structures for intent classification. The results demonstrate the viability of privacy-preserving, continuously learning chatbots. Future work aims to extend FL’s application from Natural Language Understanding (NLU) to Natural Language Generation (NLG), enhancing chatbot adaptability and performance.

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

Privacy-Preserving Self-Learning Chatbot with Federated Intelligence

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pankhuri Kulshrestha,
  • Asra Aslam,
  • Mohammad Samar Ansari

摘要

Open-domain chatbots leveraging Transformer models like GPT-3 and Transformer-XL have demonstrated impressive text generation capabilities. However, their industrial applications extend beyond text generation, requiring continuous learning from user feedback while preserving data privacy. This study explores the integration of Federated Learning (FL) into chatbot training, enabling self-learning from user interactions without compromising personally identifiable information (PII). We implement intent classification models using both traditional centralized learning and FL-based approaches, comparing their effectiveness. Additionally, we propose a Self-Feeding mechanism to address post-deployment data drift and mitigate training bias. A comprehensive chatbot architecture integrating FL and self-learning is presented, along with an evaluation of various neural network structures for intent classification. The results demonstrate the viability of privacy-preserving, continuously learning chatbots. Future work aims to extend FL’s application from Natural Language Understanding (NLU) to Natural Language Generation (NLG), enhancing chatbot adaptability and performance.