| 摘 要: 依据对话机器人回复方式的不同，可以将对话机器人分为基于生成式的对话机器人和基于检索式的对话机器人。本文主要针对上下文信息的使用，将这两类机器人继续分类为单轮检索、多轮检索、单轮生成和多轮生成对话机器人，对各对话机器人的构建进行了总结和阐述，列举了应用场景和算法实现挑战，并通过两个工业对话机器人构建中基于生成式和基于检索式的应用，分析了目前两类对话机器人落地应用的主要难点和解决方法。最后对未来的研究趋势和方向进行了展望。 |
| 关键词: 基于生成的对话机器人；基于检索的对话机器人；单轮对话；多轮对话 |
|
中图分类号: TP311.5
文献标识码: A
|
|
| Overview of Algorithm Design and Implementation of Generative and Retrieving Chatbots |
|
WANG Yu1,2, WANG Che1,2, YU Dan1,2
|
( 1. Dalian Neusoft University of Information, Dalian 116023, China ;
2. Research Institute, Dalian Neusoft Education Technology Group Co. Limited, Dalian 116023, China)
wangyu@neuedu.com; wangche@neuedu.com; yudan@neuedu.com
|
| Abstract: Responses of chatbots are generated by methods that are either generation-based or retrieval-based. This paper mainly focuses on the use of contextual information, to further classify these two types of chatbots into single-turn retrieval, multi-turn retrieval, single-turn generation, and multi-turn generation chatbots. And then, the paper summarizes and elaborates the construction of each type of these chatbots, and lists application scenarios and algorithm implementation challenges. Through generative and retrieval-based applications in construction of industrial chatbots, main difficulties and solutions are analyzed. Finally, the future research trends and directions are considered. |
| Keywords: generation-based chatbot; retrieval-based chatbot; single-turn dialogue; multi-turn dialogue |
