融合标签特征和胶囊注意力的口语理解方法  

Spoken language understanding method fusing label features and capsule attention

作　　者：李丹涛 曾碧[1] 魏鹏飞 蔡佳 LI Dan-tao;ZENG Bi;WEI Peng-fei;CAI Jia(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China;Quality and Safety Testing Center,The Fifth Electronic Research Institute of MIIT,Guangzhou 510000,China;Key Laboratory of MIIT for Intelligent Products Testing and Reliability,The Fifth Electronic Research Institute of MIIT,Guangzhou 510000,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006 [2]工业和信息化部电子第五研究所质量安全检测中心,广东广州510000 [3]工业和信息化部电子第五研究所智能产品质量评价与可靠性保障技术工业和信息化部重点实验室,广东广州510000

出　　处：《计算机工程与设计》2024年第8期2484-2491,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(62076074)。

摘　　要：针对目前意图检测和槽位填充联合学习中未充分考虑交互前标签特征信息的有效提取和融合,缺乏对交互后标签特征的提炼问题,提出一种融合标签特征和胶囊注意力的口语理解方法。主要由意图与槽位标签特征融合交互(label feature fusion interactive, LFFI)和多头胶囊注意力机制(multi-head capsule attention, MHCA)两大关键模组组成。LFFI-MHCA通过LFFI提取序列中有效的意图和槽位标签信息,对两者进行融合和交互;利用MHCA对交互过程中产生的不同子空间信息进行提炼,获得更为精确的意图和槽位标签特征。该模型在ATIS和SNIPS数据集上进行实验,句子准确率分别为88.1%和89.0%,验证了该模型的有效性。In view of the current joint learning of intent detection and slot filling,the effective extraction and fusion of label feature information before interaction is not fully considered,and the extraction of label features after interaction is lacking,a spoken language understanding was proposed by fusing label features and capsule attention.The method mainly consisted of two key modules,intent and slot label feature fusion interactive(LFFI)and multi-head capsule attention(MHCA).LFFI-MHCA was used to extract the effective intent and slot label information in the sequence through LFFI and further the extraction was fused and interacted.To better obtain more accurate intent and slot label features,MHCA was applied to refine the different subspace information generated during the interaction process.The model was tested on the ATIS and SNIPS datasets,and the sentence accuracy rates are 88.1%and 89.0%,respectively,verifying the effectiveness of the model.

关 键 词：口语理解 意图检测 槽位填充 标签特征融合交互 多头胶囊注意力机制 深度学习 自然语言处理 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]