京东 | 对话推荐系统调研

最近调研了一下 对话推荐系统（Conversational Recommender System, CRS），包括CRS基础，最新进展，工业落地相关问题，潜在的研究点等等。

下面是CRS的定义和例子

CRS: A recommendation system that can elicit the dynamic preferences of users and take actions based on their current needs through real-time multiturn interactions.

核心：通过多轮对话逐步诱导出用户的偏好并推荐商品。

CRS看起来很美好，但是落地难度还是比较大。

• NLU，NLG部分。人机对话就很难了，能针对商品推荐生成可控文本就更难了。
• 交互策略。什么时候问用户问题？什么时候推荐商品？实时反馈？
• 多个大模型NLP+Rec+RL的联合优化。尤其是RL，工业界真正可用跑起来的还不多。
• 实时性。聊天回复要快呀。
  

简单的试了下淘宝-迪卡侬-CRS，感觉还比较简陋。

调研PPT

参考

• Deep Conversational Recommender Systems: A New Frontier for Goal-Oriented Dialogue Systems. arXiv(2020)
• Tutorial on Conversational Recommendation Systems. RecSys(2020)
• CRSLab: An Open-Source Toolkit for Building Conversational Recommender System. ACL(2021)
• Conversational Recommender System. SIGIR(2018)
• Towards Conversational Search and Recommendation: System Ask, User Respond. CIKM(2018)
• Estimation-Action-Reflection: Towards Deep Interaction Between Conversational and Recommender Systems. WSDM(2020)
• Interactive Path Reasoning on Graph for Conversational Recommendation. KDD(2020)
• Towards Deep Conversational Recommendations. NeurIPS(2018)
• Towards Knowledge-Based Recommender Dialog System. EMNLP-IJCNLP(2019)
• Improving Conversational Recommender Systems via Knowledge Graph based Semantic Fusion. KDD(2020)
• Learning to Ask Appropriate Questions in Conversational Recommendation. SIGIR(2021)
• COOKIE: A Dataset for Conversational Recommendation over Knowledge Graphs in E-commerce
• Leveraging Historical Interaction Data for Improving Conversational Recommender System. CIKM(2020)
• Towards Explainable Conversational Recommendation. IJCAI(2020)
• Improving Conversational Recommender System by Pretraining Billion-scale Knowledge Graph. ICDE(2020)
• You Sound Like Someone Who Watches Drama Movies- Towards Predicting Movie Preferences from Conversational Interactions. ACL(2021)
• Adapting User Preference to Online Feedback in Multi-round Conversational Recommendation. WSDM(2021)
• Unified Conversational Recommendation Policy Learning via Graph-based Reinforcement Learning. SIGIR(2021)