Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (10): 1735-1743.DOI: 10.3778/j.issn.1673-9418.1912058

Previous Articles     Next Articles

Research of Short Text Multi-intent Detection with Capsule Network

LIU Jiao, LI Yanling, LIN Min   

  1. College of Computer Science and Technology, Inner Mongolia Normal University, Hohhot 010022, China
  • Online:2020-10-01 Published:2020-10-12

胶囊网络用于短文本多意图识别的研究

刘娇李艳玲林民   

  1. 内蒙古师范大学 计算机科学技术学院,呼和浩特 010022

Abstract:

Intent detection is a key sub-task of spoken language understanding in human-machine dialogue system. Considering the problem of user??s multi-intent expressed, a multi-intent classifier based on single-intent marker is constructed by using capsule network to identify user??s multiple intents expressed. In order to ensure the feature quality of the intent text, the deep semantic information of intent text is extracted by adding convolution capsule layer in capsule network, at the same time feature capsules are dynamically allocated to intent capsule category by using dynamic routing in capsule network. The probability of multiple intents is determined by setting threshold value, thus completing the task of multi-intent detection. Experimental results show that the capsule network is better than the convolutional neural network in the multi-intent detection task, the capsule network with convolution capsule layer can improve the performance of multi-intent detection, and the macro average F1 values on Chinese and English datasets reach 77.3% and 94.7% respectively.

Key words: multi-intent detection, deep learning, spoken language understanding, dialogue system

摘要:

意图识别是人机对话系统中口语理解的关键子任务。考虑到当前用户表达存在多个意图的问题,主要采用胶囊网络构造基于单意图标记的多意图分类器对用户表达的多种意图进行识别。为了保证意图文本的特征质量,通过在胶囊网络中增加卷积胶囊层提取意图文本的深层次语义信息,同时利用胶囊网络中的动态路由将特征胶囊动态分配到意图胶囊类别中,通过设置阈值大小判别多种意图存在的概率,从而完成多意图识别任务。实验结果表明在多意图识别任务中,胶囊网络优于卷积神经网络,而增加卷积胶囊层的胶囊网络可以提升多意图识别的性能效果,在中文和英文数据集上的宏平均F1值分别达到77.3%和94.7%。

关键词: 多意图识别, 深度学习, 口语理解, 对话系统