基于CNN的陆空通话语义识别方法

发布时间：2018-01-21 20:25

本文关键词： 陆空通话卷积神经网络语义识别自然语言处理分类器　出处：《中国民航大学》2017年硕士论文　论文类型：学位论文

【摘要】：在陆空通话过程中,飞行员对管制员的指令进行复诵是保障民航飞行安全的重要措施之一。虽然国际民航组织(International Civil Aviation Organization,ICAO)不断完善陆空通话标准,但指令复诵失误依然不可避免。因此,利用计算机辅助判断复诵过程中是否存在失误,对提高复诵准确性具有重要价值。近年来,在自然语言处理中,卷积神经网络(Convolutional Neural Network,CNN)取得了突出的效果,所以探索基于CNN模型的陆空通话语义分析与识别方法具有重要的研究意义。主要工作概括如下:(1)建立陆空通话语义分析文本库。以原始陆空通话录音为语料,在专业管制工作人员的指导下,建立陆空通话语义文本数据库。(2)研究了基于传统CNN的陆空通话语义识别方法。对分词后的语料库分别采用one-hot向量和word2vec向量表示每个词的词义,然后利用传统CNN对陆空通话双方的语句进行处理,得到对应的语义向量。根据语义向量相似度,利用分类器识别其语义一致性。(3)对传统CNN模型进行改进,提出了两种模型。通过改变卷积核的尺寸,提出了第一种改进的CNN模型;将管制员指令语句和飞行员复诵语句转换为一个矩阵作为CNN的输入,提出了第二种改进的CNN模型。在两种改进CNN模型上对陆空通话进行语义识别,分别得到语义识别结果。实验结果表明,CNN对陆空通话的语义具有较好的识别效果,并且第二种改进的CNN模型的识别结果最好,其平均识别精度最高达到84.50%。
[Abstract]:During land-to-air calls. The pilot's repetition of the controller's instructions is one of the most important measures to ensure the safety of civil aviation. Although the International Civil Aviation Organization (ICAO). International Civil Aviation Organization. ICAO) improves the standard of land-to-air communication, but the error of command recitation is still inevitable. Therefore, the use of computer aid to judge whether there are errors in the recitation process. In recent years, in natural language processing, convolution neural networks have been used to improve the volume Neural Network. CNN) has achieved outstanding results. Therefore, it is of great significance to explore the semantic analysis and recognition method of land-to-air communication based on CNN model. The main work is summarized as follows: 1). The semantic analysis database of land-to-air communication is established. The original land-to-air voice recording is used as the corpus. Under the guidance of professional control staff. Establishment of land and air conversation semantic text database. In this paper, the traditional CNN based land-to-air communication semantic recognition method is studied. The one-hot vector and the word2vec vector are used to represent the word meaning of each word in the corpus after word segmentation. Then the traditional CNN is used to deal with the statements of both sides of the land-to-air call, and the corresponding semantic vector is obtained, according to the similarity of the semantic vector. The traditional CNN model is improved by classifier recognition. Two models are proposed. By changing the size of convolution kernel, the first improved CNN model is proposed. The controller command statement and the pilot recitation statement are converted into a matrix as the input of CNN, and a second improved CNN model is proposed. The two improved CNN models are used to identify the land and air conversation semantics. The experimental results show that CNNs have a good recognition effect on the semantics of land-to-air calls, and the second improved CNN model has the best recognition results. The average recognition accuracy is up to 84.50.
【学位授予单位】：中国民航大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：V328.3;TP391.1

【参考文献】