计算机应用 ›› 2017, Vol. 37 ›› Issue (10): 2861-2865.DOI: 10.11772/j.issn.1001-9081.2017.10.2861

• 人工智能 • 上一篇    下一篇

基于深度学习的问答匹配方法

荣光辉, 黄震华   

  1. 同济大学 计算机科学与技术系, 上海 201800
  • 收稿日期:2017-05-03 修回日期:2017-07-09 出版日期:2017-10-10 发布日期:2017-10-16
  • 通讯作者: 黄震华(1980-),男,福建泉州人,教授,博士,CCF会员,主要研究方向:数据分析、数据挖掘、机器学习,E-mail:huangzhenhua@tongji.edu.cn
  • 作者简介:荣光辉(1992-),男,安徽六安人,硕士研究生,主要研究方向:深度学习、自然语言处理;黄震华(1980-),男,福建泉州人,教授,博士,CCF会员,主要研究方向:数据分析、数据挖掘、机器学习.
  • 基金资助:
    中央高校基本科研业务费专项资金资助项目(1600219256);上海市青年科技启明星计划项目(15QA1403900);上海市自然科学基金资助项目(17ZR1445900);霍英东教育基金会高等院校青年教师基金资助项目(142002)。

Question answer matching method based on deep learning

RONG Guanghui, HUANG Zhenhua   

  1. Department of Computer Science and Technology, Tongji University, Shanghai 201800, China
  • Received:2017-05-03 Revised:2017-07-09 Online:2017-10-10 Published:2017-10-16
  • Supported by:
    This work is partially supported by the Fundamental Research Funds for the Central Universities (1600219256), the Sponsored by Shanghai Rising-Star Program (15QA1403900), the Shanghai Natural Science Foundation (17ZR1445900), the Fok Ying-Tong Education Foundation for Young Teachers in the Higher Education Institutions of China (142002).

摘要: 面向中文问答匹配任务,提出基于深度学习的问答匹配方法,以解决机器学习模型因人工构造特征而导致的特征不足和准确率偏低的问题。在该方法中,主要有三种不同的模型。首先应用组合式的循环神经网络(RNN)与卷积神经网络(CNN)模型去学习句子中的深层语义特征,并计算特征向量的相似度距离。在此模型的基础上,加入两种不同的注意力机制,根据问题构造答案的特征表示去学习问答对中细致的语义匹配关系。实验结果表明,基于组合式的深度神经网络模型的实验效果要明显优于基于特征构造的机器学习方法,而基于注意力机制的混合模型可以进一步提高匹配准确率,其结果最高在平均倒数排序(MRR)和Top-1 accuray评测指标上分别可以达到80.05%和68.73%。

关键词: 问答匹配, 深度学习, 循环神经网络, 卷积神经网络, 注意力机制, 机器学习

Abstract: For Chinese question answer matching tasks, a question answer matching method based on deep learning was proposed to solve the problem of lack of features and low accuracy due to artificial structural feature in machine learning. This method mainly includes 3 different models. The first model is the combination of Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN), which is used to learn the deep semantic features in the sentence and calculate the similarity distance of feature vectors. Moreover, adding two different attention mechanism into this model, the feature representation of answer was constructed according to the question to learn the detailed semantic matching relation of them. Experimental results show that the combined deep nerual network model is superior to the method of feature construction based on machine learning, and the hybrid model based on attention mechanism can further improve the matching accuracy where the best results can reach 80.05% and 68.73% in the standard evaluation of Mean Reciprocal Rank (MRR) and Top-1 accuracy respectively.

Key words: question answer matching, deep learning, Recurrent Neural Network (RNN), Convolution Neural Network (CNN), attention mechanism, machine learning

中图分类号: