基于卷积神经网络的生物医学实体标准化研究

Research on Biomedical Entity Standardization Based on Convolutional Neural Network

随着深度学习的崛起,越来越多的人使用深度学习的方法来研究实体标准化。基于神经网络的各种复杂模型都需要大量的标注数据来进行训练,当可用的训练数据较少时其性能急剧下降。文章使用结合基于卷积神经网络的模型来研究实体标准化。标准实体由向量空间模型处理成为标准向量,预标注文本中的通俗实体经由卷积神经网络提取其中的语义特征并转化成为特征向量。在新的特征空间中,词义相近的2个特征向量之间的余弦距离应当较小。文章使用完美匹配模块来提升模型准确率和训练效率,仅采用1个卷积层和2个全连接层的浅层网络结构极大降低模型的复杂程度。整合3个结构相同但卷积核大小不同的网络模型保证模型的可靠性。5-折交叉验证来被用来提升模型的泛化能力。得益于卷积神经网络,该模型能够很好地捕捉到词向量的特征并将其标准化。简单的网络结构使得模型在可用的训练数据较少时也能够有出色表现。

With the rise of deep learning,more and more people use the method of deep learning to study entity standardization.All kinds of complex models based on neural network need a lot of labeled data for training,and their performance degrades sharply when there is less available training data.In this paper,a model based on convolution neural network is used to study entity standardization.The standard entity is processed by the vector space model into the standard vector,and the popular entities in the pre-labeled text are extracted and transformed into feature vectors through the convolution neural network.In the new feature space,the cosine distance between two feature vectors with similar meanings should be small.In this paper,the perfect matching module is used to improve the accuracy and training efficiency of the model.The shallow network structure of only one convolution layer and two fully connected layers greatly reduces the complexity of the model.Three network models with the same structure but different convolution kernel sizes are integrated to ensure the reliability of the model,and the 5-fold cross validation is used to improve the generalization ability of the model.Thanks to the convolution neural network,this model can well capture the features of word vectors and standardize them.The simple network structure enables the model to perform well when there is less training data available.