北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2025, Vol. 48 ›› Issue (1): 46-51.

• 论文 • 上一篇    下一篇

基于知识增强的多方人机主动对话策略

黄宏程1,2,  陈欣瑞1,  胡 敏1   

  1. 1. 重庆邮电大学 通信与信息工程学院 2. 重庆市通信软件工程技术研究中心
  • 收稿日期:2023-11-28 修回日期:2023-12-25 出版日期:2025-02-26 发布日期:2025-02-25
  • 通讯作者: 黄宏程 E-mail:huanghc@cqupt.edu.cn

Multi-Party Human-Computer Active Dialogue Strategy Based on Knowledge Enhancement

HUANG Hongcheng1,2,  CHEN Xinrui1,  HU Min1   

  • Received:2023-11-28 Revised:2023-12-25 Online:2025-02-26 Published:2025-02-25
  • Contact: Hong-Cheng Huang E-mail:huanghc@cqupt.edu.cn

摘要: 针对现有多方人机对话系统容易忽略在对话过程中可能被遗忘的发言者,导致用户交互主动性不强、用户对话体验感差的问题,提出了一种基于知识增强的多方人机主动对话策略。该策略借助知识图谱作为外部知识,学习对话中可能被遗忘的特定个体偏好,并结合当前群体对话需求设计个性化回复策略,促进个体主动参与群体对话。首先,使用图注意力机制学习特定个体兴趣实体表征,并引入时间权重聚合得到个体深层次偏好表示;然后,沿着当前对话主题知识子图路径触发多条邻居集合,主动捕捉个体的高阶个性化兴趣主题;最后,通过在语义层面分析当前群体对话语境和需求,评估群体对候选主题的满意度,得到既满足个体偏好又能照顾群体感受的最优对话主题。实验结果表明,融合外部知识和关注特定说话人的多方人机对话系统能有效地提升其响应内容丰富度和参与者交互满意度,推动多方对话持续进行。

关键词: 人机交互, 多方对话, 对话策略, 知识图谱

Abstract: In view of the problem that the existing multi-party human-computer dialogue system is prone to ignore speakers who may be forgotten during the dialogue process, leading to weak user interaction initiative and poor user dialogue experience,a multi-party human-computer active dialogue strategy based on knowledge enhancement is proposed. The strategy uses a knowledge graph as external knowledge to learn the preferences of specific individuals who may be overlooked in the dialogue, combine the knowledge with the current group dialogue requirements to design a personalized response strategy, thereby encouraging individual active participation in group conversations. Initially, a graph attention mechanism is employed to learn the representation of interest entities specific to individuals, and a deep preference representation for individuals is obtained by introducing time-weighted aggregation. Subsequently, multiple neighbor sets are triggered along the path of the current dialogue topic knowledge subgraph to actively capture the high-level personalized interest topics of individuals. Finally, through semantic-level analysis of the current group dialogue context and requirements, the satisfaction of the group with candidate topics is evaluated, and the optimal dialogue topic is generated, which both satisfy individual preferences and consider group sentiments. Experimental results indicate that a multi-party human-computer dialogue system, which integrates external knowledge and focuses on specific speakers, effectively enhances the richness of response content and interaction satisfaction from participants, fostering continuous multi-party dialogue.

Key words: human-computer interaction ,  multi-party interaction ,  dialogue strategy ,  knowledge graph

中图分类号: