A method for information classification of a government intensive platform

Abstract

The application belongs to the technical field of information classification method, and particularly relates to a government affair intensive platform information classification method, which comprises the following steps: through a crawler, government affair disclosure information of each government affair website is crawled and saved in a database of an intensive government affair platform; government affair disclosure data is used to train a government affair information recognition AI model; the government affair information AI model is used to recognize government affair big data, and label information of government affair information is output; the label information is stored in a relational database of the government affair intensive platform, and the platform classifies and displays government affair information according to the label information, and realizes data retrieval. The application adopts a deep neural network, constructs a government affair information recognition model with strong robustness and generalization ability, and compared with a traditional machine learning model, the application does not need experts to perform artificial feature engineering on data features in a government affair information classification task, and has excellent recognition effect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of information classification method, and particularly relates to a method for classifying information of a government intensive platform. BACKGROUND

[0002] In order to promote efficient and intensive integration and unified fusion, fuse N personalized services and operation and maintenance of government websites, realize the ecological operation environment of 1+N government websites, and realize the intensive platform of government websites, the intensive platform of government websites is first required to classify the information of each government website quickly and accurately, and different labels are assigned to government information, so that the information can be quickly displayed, and the search efficiency of the information in the website can be improved. In the existing intensive government platform, the government information classification method is mainly based on manual auditing, which is extremely low in efficiency, and the classification result is easily affected by subjective factors, which is prone to classification errors. SUMMARY

[0003] In view of the above technical problems that the existing intensive government platform is extremely low in efficiency, and the classification result is easily affected by subjective factors, the application provides a method for classifying information of a government intensive platform, which is high in efficiency, good in recognition effect and small in error.

[0004] In order to solve the above technical problems, the technical scheme adopted by the application is as follows:

[0005] A method for classifying information of a government intensive platform, comprising the following steps:

[0006] S1, through a crawler, government disclosure information of each government website is crawled and saved in a database of the intensive government platform;

[0007] S2, using government disclosure data to train a government information recognition AI model;

[0008] S3, using the government information AI model to recognize government big data, and outputting label information of the government information;

[0009] S4, storing the label information in a relational database of the government intensive platform, and classifying and displaying the government information according to the label information, and realizing data search.

[0010] The method for training the government information recognition AI model in S2 comprises the following steps:

[0011] S2.1, constructing a data set: manually labeling the crawled government information data, and the labeling categories n are customized according to the website requirements, and the data set will be used for model training; each piece of data is labeled based on Q1, Q2, Q3, Q4,..., Qn different labels, respectively representing different types of government data; the data set needs to contain more than 2000 pieces of government data;

[0012] S2.2, reading data;

[0013] S2.3, normalizing the data;

[0014] S2.4, label reconstruction;

[0015] S2.5, data set division;

[0016] S2.6, constructing a government information recognition AI model;

[0017] S2.7, network training of the government information recognition AI model;

[0018] S2.8, model verification;

[0019] S2.9, model deployment: deploying the trained model in the AI support platform of the government intensive platform.

[0020] The method of reading data in S2.2 is to read all the data of the three data sets, and expand all the data to a length equal to the longest data length with 10000 bytes as the standard, and the expansion method is to fill 0 at the end of the data. After that, all the data is read into a numpy matrix and randomly shuffled.

[0021] The method of normalizing the data in S2.3 is to normalize the data by min-max Prevent the model from learning wrong features and overfitting, and accelerate the model training to some extent.

[0022] The method of label reconstruction in S2.4 is to convert the government data category label into One-Hot form, that is, to use a 1*n matrix form, and to use 0 and 1 to label the corresponding positions of n labels, Q1 is [1, 0, 0, 0,..., 0], Q2 is [0, 1, 0, 0,..., 0], Q3 is [0, 0, 1, 0,..., 0], and Qn is [0, 0, 0,..., n], which is used for network training.

[0023] The method of data set division in S2.5 is to divide the data into a training set, a validation set and a test set in a ratio of 7:1:2, the training set is used for model parameter training to obtain an identification model, the validation set is used to confirm that the model has reached the best identification state, and the test set is used for model evaluation.

[0024] The method of constructing the government information identification AI model in S2.6 is to use a multi-layer feature fusion CNN combined with an LSTM to construct, perform multi-scale feature extraction through the CNN network, and perform time domain analysis of the features using the LSTM, and finally obtain the government category analysis result.

[0025] The method of using a multi-layer feature fusion CNN combined with an LSTM to construct is that after the government data is input into the network, it enters the CNN structure, first uses a convolutional layer with a convolution kernel size of 5 and a step size of 2 to extract features, then uses a ReLU function to activate and performs batch normalization operation; after the first feature extraction is completed, the LSTM network is used to analyze the features extracted by the CNN network in the time domain, and the output features are input into three kinds of deep neural network structures, the first kind is a 2-layer CNN+MAX POOL network, which uses a 1*1 convolution kernel for convolution operation, and then uses a pooling layer with a pooling kernel of 3 to pool the features; the second kind is a 3-layer CNN network, the first layer is a 1*1 convolution, and the second and third layers are convolution kernels with a size of 3; the third kind is a 3-layer dilated convolution network, the first layer is a 1*1 ordinary convolution, and the second and third layers are dilated convolutions with a kernel size of 3 and an expansion scale of 2; the features extracted by the three networks are fused using the ADD method, and then input into a BI-LSTM network for final classification of government information. The final government information classification result is input into the BI-LSTM network, and the classification result is output by softmax, and the maximum value of the output matrix is taken as the recognized data label.

[0026] The method of network training of the government information identification AI model in S2.7 is to train the network parameters using the training set data after the network is built, use Adam as the optimizer, the initial learning rate is 0.01, the learning rate is reduced by 50% every 100 epochs, the batch size is 32, the loss function uses center loss, and the training is set to 300 epochs. If the model loss value does not decrease for 20 consecutive epochs, stop training and save the model.

[0027] The method of model validation in S2.8 is to use the validation set data to retrain the trained data model for 100 epochs, if the model loss does not decrease, save the model; if the model loss decreases, retrain the model using the training set until the model loss is stable.

[0028] Compared with the prior art, the present application has the beneficial effects that:

[0029] The present application uses a multi-layer feature fusion CNN and LSTM to jointly construct a government information classification model, performs multi-scale feature extraction through a CNN network, performs time domain analysis of features using an LSTM, and finally obtains a government category analysis result. The present application uses a deep neural network to construct a government information recognition model with strong robustness and generalization ability. Compared with traditional machine learning models, the present application does not need experts to perform artificial feature engineering on data features in the government information classification task, and has excellent recognition effect. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and those skilled in the art can also obtain other implementation drawings according to the provided drawings without creative labor.

[0031] The structures, proportions, sizes, etc. shown in the specification are only used to cooperate with the content disclosed in the specification, to be understood and read by those skilled in the art, and do not define the limiting conditions for the implementation of the present application, so they do not have technical substantive significance. Any modification of structure, change of proportion relationship or adjustment of size, without affecting the effects and purposes that can be achieved by the present application, should still fall within the scope of the technical content disclosed by the present application.

[0032] Figure 1 Structure diagram of the government information recognition AI model of the present application;

[0033] Figure 2 Step flow chart of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. These descriptions are only for further illustrating the features and advantages of the present application, not for limiting the claims of the present application; based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0035] The specific embodiments of the present application are described in further detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application but are not intended to limit the scope of the present application.

[0036] In this embodiment, as shown in Figure 1 、 Figure 2 , the AI technology is used to classify government information, and the corresponding categories of government information are stored in the data storage structure of the government intensive platform. The construction of the system is divided into three steps:

[0037] Step 1: Through the crawler, the government information of each government website is crawled and saved in the database of the intensive government platform;

[0038] Step 2: Use government disclosure data to train government information recognition AI model;

[0039] The process of training the AI model is as follows:

[0040] Data set construction: manually label the crawled government disclosure data, and the label category n is customized according to the website requirements. The data set will be used for model training. Each data is labeled based on Q1, Q2, Q3, Q4,..., Qn different labels, respectively representing different types of government data. The data set needs to contain more than 2000 government data.

[0041] Data reading: read all data of three data sets, take 10000 bytes as standard, expand all data to make its length equal to the longest data length, and expand the data at the end of the data. 0. Then read all the data into the numpy matrix and randomly shuffle.

[0042] Data normalization: min-max normalization of data Prevent the model from learning incorrect features and overfitting, and accelerate the model training to some extent.

[0043] Label reconstruction: convert the government data category label to One-Hot form, that is, use a 1*n matrix form to label the corresponding position with 0 and 1. Q1 is [1, 0, 0, 0,..., 0], Q2 is [0, 1, 0, 0,..., 0], Q3 is [0, 0, 1, 0,..., 0], and Qn is [0, 0, 0,..., n]. Used for network training.

[0044] Data set division: divide the data into training set, validation set and test set according to the ratio of 7:1:2. The training set is used for model parameter training to obtain the recognition model, the validation set is used to confirm that the model has reached the best recognition state, and the test set is used for model evaluation.

[0045] Model structure: The model in this scheme uses a multi-layer feature fusion CNN combined with LSTM to build. Through the CNN network, multi-scale feature extraction is performed, and LSTM is used for time domain analysis of the features. Finally, the analysis result of the government affairs category is obtained. The detailed structure of the network is as follows:

[0046] After the government data is input into the network, it enters the CNN structure. First, a convolutional layer with a kernel size of 5 and a step size of 2 is used for feature extraction. Then, a ReLU function is used for activation, and a batch normalization operation is performed. After the first feature extraction, the LSTM network is used to analyze the features extracted by the CNN network in the time domain. The output features are input into three types of deep neural networks. The first type is a 2-layer CNN+MAX POOL network, which uses a 1*1 convolution kernel for convolution operation, and then uses a pooling kernel of 3 for pooling. The second type is a 3-layer CNN network, where the first layer is a 1*1 convolution, and the second and third layers are convolution kernels with a size of 3. The third type is a 3-layer dilated convolution network, where the first layer is a 1*1 ordinary convolution, and the second and third layers are dilated convolution with a kernel size of 3 and an expansion scale of 2. The features extracted by the three networks are fused using the ADD method, and then input into a BI-LSTM network for the final classification of government information. The classification result is output by the softmax, and the maximum value of the output matrix is taken as the recognized data label.

[0047] Network training: After the network is built, the training set data is used to train the network parameters. Adam is used as the optimizer, the initial learning rate is 0.01, the learning rate is reduced by 50% every 100 epochs, the batch size is 32, and the loss function uses center loss. Set the training to 300 epochs, stop training if the model loss value does not decrease for 20 consecutive epochs, and save the model.

[0048] Model verification: The data model obtained by training is trained for 100 epochs using the validation set data. If the model loss does not decrease, save the model. If the model loss decreases, redivide the training set to train the model until the model loss is stable.

[0049] Model deployment: Deploy the trained model in the AI support platform of the government intensive platform.

[0050] Step 3: Use the government information AI model to recognize government big data and output the label information of government information.

[0051] The fourth step is to store the label information in the relational database of the government affair intensive platform, and the platform classifies and displays the government affair information according to the label information, and realizes data retrieval.

[0052] The above only describes the preferred embodiments of the present application in detail, but the present application is not limited to the above-mentioned embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the present application, and all the changes shall be included in the protection scope of the present application.

Claims

1. A method for information classification of government intensive platform, characterized in that: The method comprises the following steps: S1, by crawler, crawling the government information of each government website, and saving it in the database of the intensive government platform; S2, using government information AI model to identify government data; S2.1, constructing a data set: manually labeling the crawled government information, labeling categories n are customized according to website requirements, and the data set will be used for model training; Each data is respectively based on Q1, Q2, Q3, Q4,..., Qn different labels, respectively representing different types of government data; The data set needs to contain more than 2000 government data; S2.2, read data; S2.3, normalize data; S2.4, reconstruct labels; S2.5, data set division; S2.6, constructing government information recognition AI model; using multi-layer feature fusion CNN and LSTM to jointly construct, through CNN network for multi-scale feature extraction, using LSTM for time domain analysis of features, and finally obtaining government category analysis results; The method for using multi-layer feature fusion CNN and LSTM to jointly construct is: after the government data is input into the network, it enters the CNN structure, first uses the convolution kernel with the size of 5 and the step of 2 to extract features, then uses the ReLU function to activate, and carries out batch normalization operation; After the first feature extraction is completed, the LSTM network is used to analyze the time domain of the features extracted by the CNN network, and the output features are input into three kinds of deep neural network structures, the first kind is a 2-layer CNN+MAX POOL network, using a 1*1 convolution kernel for convolution operation, and then using a 3-pooling kernel for pooling layer to pool the features; The second kind is a 3-layer CNN network, the first layer is 1*1 convolution, and the second layer and the third layer are convolution kernels with the size of 3; The third kind is a 3-layer dilated convolution network, the first layer is a 1*1 ordinary convolution, the second layer and the third layer are convolution kernels with the size of 3 and the dilated scale of 2; The features extracted by the three networks are fused by ADD method, and then input into a BI-LSTM network for final classification of government information. The final government information classification result is input into the BI-LSTM network, and the classification result is output by softmax. The maximum value of the output matrix is taken as the recognized data label; S2.7, network training of government information recognition AI model; S2.8, model verification; S2.9, model deployment: deploying the trained model in the AI support platform of the government intensive platform; S3, using government information AI model to identify government big data, and outputting the label information of government information; S4, storing the label information in the relational database of the government intensive platform, and classifying and displaying the government information according to the label information, and realizing data retrieval. 2.The method of information classification of a government affair intensive platform according to claim 1, characterized in that: The method for reading data in S2.2 is to read all data of the data set, expand all data to a length equal to the longest data length by filling 0 at the end of the data, and then read all data into a numpy matrix and randomly shuffle.

3. The method of claim 1, wherein: The method for normalizing data in S2.3 is min-max normalization of data , prevent the model from learning error features to overfit, and accelerate model training.

4. The method of claim 1, wherein: The method for label reconstruction in S2.4 is to convert the government data category label into One-Hot form, that is, to use a 1*n matrix form, and to use 0 and 1 to mark the corresponding positions for n labels. Q1 is [1, 0, 0, 0,..., 0], Q2 is [0, 1, 0, 0,..., 0], Q3 is [0, 0, 1, 0,..., 0], and Qn is [0, 0, 0,..., n], which is used for network training.

5. The method of claim 1, wherein: The method for data set division in S2.5 is to divide the data into training set, validation set and test set according to the ratio of 7:1:

2. The training set is used for model parameter training to obtain an identification model. The validation set is used to confirm that the model has reached the best identification state. The test set is used for model evaluation.

6. The method of claim 1, wherein: The method for network training of the government information identification AI model in S2.7 is to train the network parameters using the training set data after the network is built. Adam is used as the optimizer, the initial learning rate is 0.01, the learning rate is reduced by 50% every 100 epochs, the batch size is 32, the loss function uses center loss, 300 epochs are set for training, and the model is stopped training if the model loss value does not decrease for 20 consecutive epochs. The model is saved.

7. The method of claim 1, wherein: The method for model validation in S2.8 is to use the validation set data to perform secondary training of the data model obtained by training for 100 epochs. If the model loss does not decrease, the model is saved. If the model loss decreases, the training set is re-divided to train the model until the model loss is stable.

Images