计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (18): 233-240.DOI: 10.3778/j.issn.1002-8331.2102-0172

• 模式识别与人工智能 • 上一篇    下一篇

语义及句法特征多注意力交互的医疗自动问答

张华丽,康晓东,李小军,刘汉卿,王笑天   

  1. 1.天津医科大学 医学影像学院,天津 300203
    2.重庆市黔江中心医院,重庆 409099
  • 出版日期:2022-09-15 发布日期:2022-09-15

Semantic and Syntactic Features with Multi-Attentive Interaction for Medical Question Answering

ZHANG Huali, KANG Xiaodong, LI Xiaojun, LIU Hanqing, WANG Xiaotian   

  1. 1.School of Medical Imaging, Tianjin Medical University, Tianjin 300203, China
    2.Chongqing Qianjiang Central Hospital, Chongqing 409099, China
  • Online:2022-09-15 Published:2022-09-15

摘要: 针对中文医疗自动问答任务,为了捕捉问答句中重要的句法信息和语义信息,提出引入图卷积神经网络捕捉句法信息,并添加多注意力池化模块实现问答句的语序特征和句法特征联合学习的方法。在BERT模型学习问答句的高阶语义特征基础上,利用双向门控循环单元描述句子的全局语义特征,以及引入图卷积神经网络编码句子的语法结构信息,以与双向门控循环单元所获取的序列特征呈现互补关系;通过多注意力池化模块对问答对的不同语义空间上的编码向量进行两两交互,并着重突出问答对的共现特征;通过衡量问答对的匹配分数,找出最佳答案。实验结果表明,在cMedQA v1.0和cMedQA v2.0数据集上,相比于主流的深度学习方法,所提方法的ACC@1有所提高。实验证明引入图卷积神经网络和多注意力池化模块的集成算法能有效提升自动问答模型的性能。

关键词: 自动问答, 双向门循环单元, 图卷积神经网络, 句法信息, 多注意力池化

Abstract: For Chinese medical question answering task, in order to capture important syntactic and semantic information in question and answer sentences, it is proposed to introduce graph convolutional network to capture syntactic information, and add a multi-attentive pooling module to realize joint learning of word order and syntactic features of question and answer sentences. On the basis of the BERT model learning the high-order semantic features of question and answer sentences, the bi-directional gated recurrent unit is used to describe the global semantic features of the sentence, and the introduction of the graph convolutional neural network to encode the grammatical structure information of the sentence to present a complementary relationship with the sequence features obtained by the bi-directional gated recurrent unit. Through the multi-attentive pooling module, the encoding vectors in different semantic spaces of the question and answer pairs are interacted in pairs, and the co-occurrence characteristics of the question and answer pairs are emphasized. By measuring matching score of the question and answer pairs to find the best answer. The experimental results show that on the cMedQA v1.0 and cMedQA v2.0 data sets, the ACC@1 of the proposed method is improved compared with mainstream deep learning methods. Experiments prove that the introduction of the integrated algorithm of graph convolutional network and multi-attentive pooling module can effectively improve the performance of question answering model.

Key words: question answering, bi-directional gated recurrent unit, graph convolutional network, syntactic information, multi-attentive pooling