智能语音交互中的多意图解析技术研究  

Research on multi-intent parsing technology in intelligent voice interaction

作　　者：唐杰 贾巨涛 李立辉 张鹏 杨驰 TANG Jie;JIA Jutao;LI Lihui;ZHANG Peng;YANG Chi(Gree Electric Appliances,Inc.of Zhuhai,Zhuhai 519070;State Key Laboratory of Air-conditioning Equipment and System Energy Conservation,Zhuhai 519070)

机构地区：[1]珠海格力电器股份有限公司,广东珠海519070 [2]空调设备及系统运行节能国家重点实验室,广东珠海519070

出　　处：《家电科技》2024年第S01期418-421,共4页Journal of Appliance Science & Technology

摘　　要：随着智能家居的广泛普及,用户逐渐习惯于通过语音同系统进行交互,实现智能家居设备控制。但是传统的单意图识别对话系统无法解析这些复杂的指令,导致交互过程生硬、用户体验较差。为了解决这一问题,提出了一种多模块级联的多意图解析对话系统。该系统首先利用意图切分模型将用户的多意图请求切分为单意图请求;接着通过粗粒度意图分类和细粒度意图分类两级分类模型提取单意图类别;最后在对话管理模块中基于预设规则补全领域和词槽信息。使用六种家电品类的数据对该系统性能进行评估,系统整体的基础指令准确率达到100%,泛化指令准确率超过96%,满足实际应用的需求。With the widespread adoption of smart homes,users have gradually become accustomed to controlling multiple devices or functions through a single command.Traditional single-intent recognition dialogue systems cannot parse these complex commands,resulting in a stiffinteraction process and poor user experience.To address this issue,Proposes a multi-module cascade multi-intent parsing dialogue system.This system first uses an intent segmentation model to split the user’s multi-intent request into single-intent requests;then,it extracts the single-intent categories through a two-level classification model,including coarse-grained intent classification and fine-grained intent classification;finally,the dialogue management module completes the domain and slot information based on preset rules.The system’s performance was evaluated using data from six types of home appliances.The basic command accuracy rate for each module and the overall system reached 100%,and the generalized command accuracy rate exceeded 96%,meeting the needs of practical applications.

关 键 词：智能家居 语音交互 对话系统 文本分类 对话管理 

分 类 号：TM925[电气工程—电力电子与电力传动] TP39[自动化与计算机技术—计算机应用技术]