【摘要】 为了有效获取交通运输信息标准中的一致性条款,简化标准测试方法,针对现有文本分类方法中卷积神经网络存在的缺少上下文含义和循环神经网络存在的梯度消失及梯度弥散等问题,提出一种基于BLSTM的文本增强表示方法和基于CNN网络的语句分类相结合的方法进行一致性条款分类.其核心思想是将BLSTM前向和后向过程产生的向量相加,然后与原文本向量拼接作为文本的向量表示,将文本向量作为CNN网络的输入进行文本分类.为验证所提模型的有效性,设置了与传统机器模型TF-IDF+SVM、单CNN、BLSTM神经网络模型及经典混合模型的对比试验.通过构造的交通运输信息标准条款数据集测试表明,基于改进的BLSTM和CNN的链式混合神经网络模型准确率达到93.77%.更多还原

【Abstract】 To effectively obtain the conformance clauses in the transportation information standards and simplify the standard methods, combining the bidirectional long short-term memory(BLSTM)based text enhancement representation with the CNN based sentence classification, the classification method was proposed to classify the conformance clauses for solving the problems of lack of context meaning in convolution neural network and gradient disappearance and gradient dispersion in cyclic neural network in the existing text classification methods. The core idea was to add the vectors generated by the forward and backward processes of BLSTM, and the added vectors were spliced with the original vector as vector representation of the text. The text was classified as the input of CNN network. To verify the proposed model, the comparative test with traditional TF-IDF+SVM machine model, single CNN, BLSTM neural network model and classic hybrid model was set up. According to the test of the data set of standard terms of transportation information, the accuracy of the chain-mixed neural network model based on the improved BLSTM and CNN reaches 93.77%.更多还原