基于字符级卷积神经网络的数学运算符识别

发布时间：2018-03-16 02:28

本文选题：深度学习　切入点：卷积神经网络　出处：《华中师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,在教育信息化背景下,人工智能、自然语言处理等技术日渐成熟,越来越多的数学题目以电子文档的形式呈现的迫切需求,探索教育信息化环境下学生自主创新学习的新模式,辅助学生提高数学问题的解决能力,为学习者提供智能化的教育服务已成为必然趋势。四则运算在数学计算中起着关键性作用,实现四则运算的智能解答是机器自动求解数学题目的基础,可以为教育信息化环境下学生的自主学习提供个性化辅导。构建四则运算题目的智能解答模型,可以为数学运算、合并同类项、求解方程式等题目的机器自动解答提供技术支持和参考。而四则运算题目的智能解答的核心在于四则运算式中运算符的识别,本项研究针对智能解答系统中的四则运算符识别问题,以深度学习为切入点,采用基于字符级的编码方式,训练CNN,形成了四则运算符识别的网络模型,并通过实验验证了其有效性。本项研究主要包括四个方面:第一,算法架构的搭建,构建了神经网络和卷积神经网络。并设置好卷积神经网络模型的卷积、池化等提取特征方法,以及卷积神经网络中隐藏层的层数及每个隐藏层对应的节点数。第二,设置编码和字符量化的规则。根据字符量化规则,使用One-hot编码方法对实验数据进行编码处理,将输入数据转换为一个一维向量。第三,生成实验数据,从加减乘除四个维度分别生成三种类型的数学四则运算式数据集,包括训练集、验证集、测试集。第四,将编码后向量化的数据集分别输入到设计好的两个网络中,从训练中学习运算符识别的“经验”。并使用测试数据集和预测数据检验了识别效果。目前,将深度学习技术与机器自动求解相融合的大量的科研工作均是基于词的文本分类研究,而本项研究采用的是字符级的字符编码识别,通过实验验证了网络的有效性,生成了四则运算符识别的网络模型,相比传统神经网络的接近100%的识别率,深度卷积网络的正确识别率达到了 100%。并用随机生成的预测集进行了预测,预测完全正确,达到了数学四则运算符的识别目的。
[Abstract]:In recent years, under the background of educational information, artificial intelligence, natural language processing and other technologies are becoming more and more mature, and more and more mathematical problems are presented in the form of electronic documents. To explore a new model of students' independent and innovative learning under the environment of educational information, to assist students to improve their ability to solve mathematical problems, It has become an inevitable trend to provide intelligent educational services for learners. Four principle operations play a key role in mathematical calculation, and the intelligent solution of the four principles is the basis for the machine to solve mathematical problems automatically. It can provide individualized tutoring for students' autonomous learning in the information environment of education. The machine automatic solution to the problems such as equations provides technical support and reference. The core of the intelligent solution of the four principle operation problems lies in the identification of operators in the four principles. In order to solve the problem of the recognition of the four principle operators in the intelligent solution system, this research takes the deep learning as the breakthrough point, uses the coding method based on the character level, trains the CNN, and forms a network model of the recognition of the four arithmetic operators. This research mainly includes four aspects: firstly, the structure of the algorithm, the neural network and the convolution neural network are constructed, and the convolution neural network model is set up. In addition, the number of hidden layers and the number of nodes corresponding to each hidden layer in the convolutional neural network. Secondly, the rules of encoding and quantization are set, and the rules of character quantization are used according to the rules of character quantization. The experimental data is encoded by One-hot coding method, and the input data is converted into a one-dimensional vector. Thirdly, the experimental data is generated, and three types of mathematical four-principle expression data sets are generated from the four dimensions of addition, subtraction, multiplication and division, respectively. Including the training set, the verification set, the test set. 4th, the encoded vectorized data sets are respectively input into the two networks designed. Learn the "experience" of operator recognition from the training. And test the recognition effect using test data sets and predictive data. A great deal of research work that combines depth learning technology with machine automatic solution is based on text classification of words, and this research adopts character level character coding recognition, and the effectiveness of the network is verified by experiments. The network model of four arithmetic operators is generated. Compared with the traditional neural network, the recognition rate of the deep convolution network is close to 100%, and the correct recognition rate of the deep convolution network is 100. The prediction is completely correct by using the randomly generated prediction set. The recognition purpose of mathematical four arithmetic operators is achieved.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：G623.5;G434

【参考文献】