
软件学报 ISSN 1000-9825, CODEN RUXUEW E-mail: jos@iscas.ac.cn
Journal of Software,2019,30(11):3313−3325 [doi: 10.13 328/j.cnki.jos.005862] http://www.jos.org.cn
©中国科学院软件研究所版权所有. Tel: +86-10-62562563
面向人机对话意图分类的混合神经网络模型
∗
周俊佐
,
朱宗奎
,
何正球
,
陈文亮
,
张
民
(苏州大学 计算机科学与技术学院 人工智能研究院,江苏 苏州 215008)
通讯作者: 陈文亮, E-mail: wlchen@suda.edu.cn
摘 要: 随着人机对话的不断发展,让计算机能够准确地理解用户查询意图,对整个人机对话领域都有着重要意
义.意图分类的主要目标是在人机对话的过程中判断用户的意图,提升人机对话系统的准确度与自然度.首先分析多
个分类模型在意图分类任务上的优缺点.在此基础上,提出一种混合神经网络模型,综合利用多个深度网络模型的多
样性输出.在输入特征预处理上,采用语言模型词向量,将语言模型拥有的语义挖掘能力应用到混合网络中,可以进
一步提升模型的表达能力.所提出的混合神经网络模型相对于最好的基准模型在两份数据集上分别取得了 2.95%
和 3.85%的性能提升.新模型在该数据上取得了最优的性能.
关键词: 混合模型;意图分类;语言模型;注意力机制;胶囊网络
中图法分类号: TP18
中文引用格式: 周俊佐,朱宗奎,何正球,陈文亮,张民.面向人机对话意图分类的混合神经网络模型.软件学报,2019,30(11):
3313−3325. http ://www.jos.org.cn/1000-9825/5862.htm
英文引用格式: Zhou JZ, Zhu ZK, He ZQ, Chen WL, Zhang M. Hybrid neural network models for human-machine dialogue
intention classification. Ruan Jian Xue Bao/Journal of Software, 2019,30(11):3313−3325 (in Chinese). http://www.jos.org.cn/
1000-9825/5862.ht m
Hybrid Neural Network Models for Human-machine Dialogue Intention Classi ficatio n
ZHOU Jun-Zuo, ZHU Zong-Kui, HE Zheng-Qiu, CHEN Wen-Liang, ZHANG Min
(Institute of Artificial In telligence, S chool of Computer Science and Technology, Soochow Universit y, Suzhou 215008, China)
Abstra ct : With the development of human-machine dialogue, it is of great significance for the computer to accurately understand the
user’s query intention in human-machine dialogue systems. Intention classifi cation aims at judging th e user ’s intention in human machine
dialogue and improves the accuracy and naturalness of the hu man machine dialogue system. This s tudy first analyzes the advantages and
disadvantages of multiple classification models in the intention classification task. On this basis, this study proposes a hybrid neural
network model to comprehensively utilize the diversity outputs of multiple deep network models. To further improve the perfoance, the
language model embedding is used in the input feature preprocessing and the semantic mining ability possessed for the hybrid network
which can effectively improve the expression ability of the model. The proposed model achieves 2.95% and 3.85% performance
improvement on the two data sets respectively compared to the best benchmark model. The proposed model also achieves the top
performance in a shared task.
Key words: hybrid model; intention classification; language model; attention mechanism; capsule network
近年来,随着数字虚拟人技术、移动终端和语音识别处理的快速发展,人与计算机直接对话与问答的人机
交互形式变得越来越重要.目前,许多数字虚拟人能逼真地朗读出用户给定的内容、理解用户的查询意图、回
答用户信息查询信息以及购物问题等等
[1]
.国内外各个机构和组织纷纷开发了自己的人机对话系统,例如 2011
∗ 基金项目: 国家自然科学基金(61876115, 61572338, 61525205); 江苏高校优势学科建设工程(PAPD)
Foundation item: National Natural Science Foundation of China (61876115, 61572338, 61525205); Project Funded by the Priority
Academic Program Develop ment of J iangsu High er Edu cation Ins titutions
收稿时间: 2019-0 1-15; 修改时间: 2019-03-12; 采用时间: 2019-04-04
评论