端到端的面向任务型对话系统多任务优化模型  

End-to-end multi-task optimization model for task-based dialogue systems

作　　者：赵逢达[1,2,3] 邱梦璐 李贤善 孙永派[1,3] 杨智开 ZHAO Fengda;QIU Menglu;LI Xianshan;SUN Yongpai;YANG Zhikai(School of Information Science and Engineering,Yanshan University,Qinhuangdao 066004,China;School of Information Science and Engineering,Xinjiang University of Science and Technology,Korla 841000,China;The Key Laboratory for Software Engineering of Hebei Province,Qinhuangdao 066004,China)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]新疆科技学院信息科学与工程学院,新疆库尔勒841000 [3]河北省软件工程重点实验室,河北秦皇岛066004

出　　处：《计算机集成制造系统》2023年第11期3592-3599,共8页Computer Integrated Manufacturing Systems

基　　金：新疆维吾尔自治区自然科学基金面上项目(2022D01A59);新疆维吾尔自治区高校科研计划资助项目(自然科学重点项目)(XJEDU2021I029);河北省创新能力提升计划资助项目(22567637H)。

摘　　要：在对话系统中的自然语言理解模块中存在着意图检测与插槽填充两个任务,这两个任务之间存在着极强的相关性,即插槽信息的生成高度依赖于意图信息。然而,现有工作大部分将其视为两个独立任务实现,导致对话系统的准确率无法得到进一步提升。为此,针对对话系统中意图检测任务与插槽填充任务之间的相关性信息,在已有工作的基础上提出了端到端的基于Stack-Propagation(堆栈传播)思想实现的网络模型。该模型在解码器阶段借鉴并改进了Stack-Propagation框架的思想,即将意图检测的结果加入到插槽填充任务的输入中,使用意图检测的结果去进一步指导插槽填充任务的进行。通过在斯坦福多领域对话数据集上进行实验,证明该模型不仅可以充分利用意图检测任务与插槽填充任务之间的相关性信息,还可以通过联合学习达到相互促进的效果,最终有效提高对话系统的准确率。In the natural language understanding module,there are two tasks of intent detection and slot filling,and there is a strong correlation between the two tasks that the generation of slot information is highly dependent on the intent information.However,most of the existing works regard it as two independent tasks to achieve,resulting in that the accuracy of the dialogue system cannot be further improved.To this end,aiming at the correlation information between the intent detection task and the slot filling task in the dialogue system,on the basis of the existing work,an end-to-end network model based on the idea of Stack-Propagation was proposed.The idea of the Stack-Propagation framework in the decoder stage was borrowed and improved,which added the result of intent detection to the input of the slot filling task,and used the result of the intent detection to further guide the slot filling task.Through experiments on SMD dataset,it was proved that the model could not only make full use of the correlation information between the intent detection task and the slot filling task,but also achieve the effect of mutual promotion through joint learning,and finally effectively improve the accuracy of the dialogue system.

关 键 词：对话系统 意图检测 插槽填充 堆栈传播框架 人机交互 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]