基于LCN的医疗知识问答模型

An LCN-Based Medical Knowledge Base Question Answering Model

中文医疗领域分词比较困难,导致现有算法对于医疗问题特征提取不充分,针对中文分词的特点,提出基于LCN(Lattice CNN,格子卷积神经网络)的医疗知识问答模型.首先,利用某三甲医院提供的15000份电子住院记录,基于电子住院记录利用Glove模型训练医学词向量.其次,通过各大医疗网站获得大量医学名词及名词间的关系,构建医学知识图谱,并提取知识图谱中的关系词,结合已训练的词向量获取关系向量.最终,以医学词向量作为模型输入端并利用LCN神经网络提取医疗问题特征,计算问题特征与关系向量的相似度,进而训练医疗知识问答模型.实验表明,LCN模型准确率可达89.0%,与同类问答模型比较,提高了2%.

Word segmentation in the Chinese medical field is difficult,resulting in inadequate extraction of medical problem features by the existing algorithms.This paper proposes a medical knowledge base question answering model based on LCN(Lattice Convolutional Neural Network)for the characteristics of Chinese word segmentation.First,15000 electronic medical records provided by a first-class hospital at Grade 3 are used to train medical word vectors with the Glove model.Then,a large number of medical nouns and their intra-relations are obtained through major medical websites to construct a medical knowledge map.The relationship words in the knowledge graph are extracted and,combined with the trained word vectors,the relationship vectors are obtained.Finally,the medical word vector is used as the model input and the LCN neural network is used to extract the medical problem features.The model is trained by calculating the similarity of the question features and relationship vectors.Experiments show that the accuracy rate of the LCN model is as high as 89.0%,which is an improvement of 2%compared with similar question answering models.