智能问诊中基于深度神经网络的反问生成方法

在智能问诊中，为了让医生快速提出合理的反问以提高医患对话效率，提出了基于深度神经网络的反问生成方法。首先获取大量医患对话文本并进行标注；然后使用文本循环神经网络（TextRNN）、文本卷积神经网络（TextCNN）二种分类模型分别对医生的陈述进行分类；再利用双向文本循环神经网络（TextRNN-B）、双向变形编码器（BERT）分类模型进行问题触发；设计六种不同的问答选取方式来模拟医疗咨询领域情景，采用开源神经机器翻译（OpenNMT）模型进行反问生成；最后对已生成的反问进行综合评估。实验结果表明，使用TextRNN进行分类优于TextCNN，利用BERT模型进行问题触发优于TextRNN-B，采用OpenNMT模型在Window-top方式下实现反问生成时，使用双语评估替补（BLEU）和困惑度（PPL）指标进行评价的结果最好。所提方法验证了深度神经网络技术在反问生成中的有效性，可以有效解决智能问诊中医生反问生成的问题。

Method of generating rhetorical questions based on deep neural network in intelligent consultation

In order to improve the efficiency of doctor-patient dialogue by enabling doctors to quickly propose reasonable rhetorical questions in intelligent consultation， a method of rhetorical question generation based on deep neural network was proposed. Firstly， a large number of doctor-patient dialogue texts were obtained and labeled. Then， two classification models， Text Recurrent Neural Network （TextRNN） and Text Convolutional Neural Network （TextCNN）， were used to classify doctor’s statements respectively. Then， Text Recurrent Neural Network-Bidirectional Long Short-Term Memory （TextRNN-B） and Bidirectional Encoder Representations from Transformers （BERT） classification models were used to trigger questions. Six different Q&A selection methods were designed to simulate the situations in the field of medical consultation. Then， Open-Source Neural Machine Translation （OpenNMT） model was used to generate rhetorical questions. Finally， the generated rhetorical questions were evaluated comprehensively. Experimental results show that TextRNN is better than TextCNN in classification， and BERT model is better than TextRNN-B in question triggering； when OpenNMT model is used to realize rhetorical question generation in Window-top mode， the best results are obtained by using two evaluation indexes： Bilingual Evaluation Understudy （BLEU） and Perplexity （PPL）. The proposed method verifies the effectiveness of deep neural network technology in the generation of rhetorical questions， which can effectively solve the problem of doctor-patient question generation.