CN108197337A - A kind of file classification method and device - Google Patents

Abstract

The present invention provides a text classification method and device. On the basis of the pre-established CNN classification model, the method of initializing the weight of the convolutional layer is improved. Specifically, the weight is initialized according to the Gaussian distribution. Compared with the existing CNN-based The classification model realizes the method of text classification, which improves the accuracy of classification results. And compared with Naive Bayesian, machine learning algorithms such as SVM also improve the accuracy of classification results.

Description

A text classification method and device

technical field

The present invention relates to the technical field of classification, in particular to a text classification method and device.

Background technique

The method for implementing text classification in the prior art is: extracting text features of the text to be classified, and classifying the text to be classified according to the text features.

Based on the text classification methods disclosed in the prior art, when classifying news, since the news is a long text, when extracting the text features of the news to be classified, it is often necessary to invest a lot of manpower and time to design effective text features. Helping with classification is time-consuming and labor-intensive.

Since deep learning can automatically learn text features, it can solve the problem of difficult text feature extraction when classifying long texts such as news. The convolutional neural network model (CNN) is commonly used in deep learning.

The method of using the CNN model to realize text classification includes: preprocessing the text to be classified to obtain several sentences; inputting each sentence into the convolutional layer and sampling layer of the trained CNN model, and inputting the output result of the sampling layer into the SVM In the classifier, the classification of text is realized.

However, the inventors found that the existing method for text classification based on the CNN model has a problem of low accuracy.

Contents of the invention

In view of this, the object of the present invention is to provide a text classification method and device to solve the problem of low accuracy in the prior art method for text classification based on CNN model.

The technical solution is as follows:

The present invention provides a text classification method, comprising:

Preprocess the text to be classified to obtain multiple sentences;

The sentence is input into the input layer of the pre-established CNN classification model;

extract the word2vec feature of the sentence, and obtain the input matrix;

The input matrix is input into the convolution layer, and features are extracted through convolution operations; wherein, the weights of the convolution layer are initialized with Gaussian distribution parameter values;

The features are input into the classifier for classification.

Preferably, the preprocessing of the text to be classified to obtain a plurality of sentences includes:

judging whether the length of the text to be classified is greater than a preset length;

Judging that the length of the text to be classified is greater than a preset length, truncating the text to be classified according to the preset length to obtain multiple sentences;

judging whether the length of the sentence is less than a preset length;

If it is judged that the length of the sentence is less than the preset length, the text to be classified is spliced after the sentence according to the order of the content included in the text until the length of the new sentence formed after splicing is equal to the preset length.

Preferably, the training method of the CNN classification model comprises:

Acquiring data sets and uncleaned data; wherein, the uncleaned data includes preset classification marks;

Utilize described data set, carry out initial training to described CNN classification model;

Using the CNN classification model after the initial training to classify and predict the uncleaned data, and obtain the predicted classification marks and predicted probabilities of the uncleaned data;

judging whether the predicted probability of the uncleaned data is greater than a preset probability value;

Judging that the predicted probability of the uncleaned data is greater than a preset probability value, then judging whether the predicted classification mark of the uncleaned data is the same as the preset classification mark of the uncleaned data;

Judging that the predicted classification mark of the uncleaned data is different from the preset classification mark of the uncleaned data, modifying the preset classification mark of the uncleaned data to the predicted classification mark to obtain the cleaned data;

Using the cleaned data to train the CNN classification model.

Preferably, the modifying the preset classification mark of the uncleaned data to the predicted classification mark comprises:

According to preset rules, select the uncleaned data to be processed from the uncleaned data whose predicted classification labels of the uncleaned data are judged to be different from the predicted classification labels of the uncleaned data;

modifying the preset classification label of the uncleaned data to be the predicted classification label.

Preferably, the training method of the CNN classification model also includes:

If the number of training samples is less than the preset number of samples, the pre-trained CNN classification model is used;

Using the training samples, the pre-trained CNN classification model is trained.

The present invention also provides a text classification device, comprising:

A preprocessing unit is used to preprocess the text to be classified to obtain multiple sentences;

The input unit is used to input the sentence into the input layer of the pre-established CNN classification model;

The first processing unit is used to extract the word2vec feature of the sentence and obtain an input matrix;

The second processing unit is configured to input the input matrix into a convolution layer, and extract features through a convolution operation; wherein, the weight of the convolution layer is initialized with a Gaussian distribution parameter value;

A classification unit, configured to input the features into a classifier for classification.

Preferably, the preprocessing unit includes:

a first judging unit, configured to judge whether the length of the text to be classified is greater than a preset length;

A truncation unit, configured to truncate the text to be classified according to the preset length to obtain multiple sentences when the first judging unit judges that the length of the text to be classified is greater than a preset length;

The second judging unit is used to judge whether the length of the sentence is less than a preset length;

splicing unit, for when the second judging unit judges that the length of the sentence is less than the preset length, according to the order of content included in the text to be classified, it is spliced behind the sentence until a new sentence is formed after splicing The length of is equal to the preset length.

Preferably, it also includes:

An acquisition unit, configured to acquire data sets and uncleaned data; wherein, the uncleaned data includes preset classification marks;

A training unit, configured to use the data set to initially train the CNN classification model;

A prediction unit, configured to use the CNN classification model after initial training to classify and predict the uncleaned data, and obtain the predicted classification marks and prediction probabilities of the uncleaned data;

A third judging unit, configured to judge whether the predicted probability of the uncleaned data is greater than a preset probability value;

A fourth judging unit, configured to judge the predicted classification flag of the uncleaned data and the preset classification of the uncleaned data when the third judging unit judges that the predicted probability of the uncleaned data is greater than a preset probability value whether the marks are the same;

A modifying unit, configured to modify the preset classification mark of the uncleaned data to the The above prediction classification marks are obtained to obtain the cleaned data;

The training unit is further configured to use the cleaned data to train the CNN classification model.

Preferably, the modifying unit includes:

The selection subunit is used to select the uncleaned data to be processed from the uncleaned data whose predicted classification labels of the uncleaned data are judged to be different from the predicted classification labels of the uncleaned data according to preset rules;

The modifying subunit is configured to modify the preset classification mark of the uncleaned data to be the predicted classification mark.

Preferably, it also includes:

A multiplexing unit is used to reuse the pre-trained CNN classification model if the number of training samples is less than the preset number of samples;

The training unit is further configured to use the training samples to train the pre-trained CNN classification model.

Compared with the prior art, the above-mentioned technical solution provided by the present invention has the following advantages:

It can be seen from the above technical solution that, on the basis of the pre-established CNN classification model, the application improves the way of initializing the weight of the convolutional layer, specifically according to the Gaussian distribution of weight initialization, compared with the existing CNN-based classification model The method for implementing text classification improves the accuracy of classification results. And compared with Naive Bayesian, machine learning algorithms such as SVM also improve the accuracy of classification results.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are For some embodiments of the present invention, those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a flow chart of a text classification method provided by an embodiment of the present invention;

FIG. 2 is a flowchart of another text classification method provided by an embodiment of the present invention;

Fig. 3 is the flowchart of the training method of the CNN classification model that the embodiment of the present invention provides;

FIG. 4 is a schematic structural diagram of a text classification device provided by an embodiment of the present invention;

Fig. 5 is a schematic structural diagram of another text classification device provided by an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

This embodiment discloses a text classification method, which is applied in the scene of long text classification, for example, the scene of news classification, referring to Fig. 1, this embodiment includes the following steps:

S101. Preprocessing the text to be classified to obtain multiple sentences;

When the text to be classified is a long text, it is necessary to truncate the text to be classified into multiple texts of preset lengths, and each text of a specific length is a sentence to obtain multiple sentences. Wherein, the preset length can be set according to actual needs.

S102. Input the sentence into the input layer of the pre-established CNN classification model;

In this embodiment, a CNN classification model is pre-established and trained. The CNN classification model has two parallel input layers, and each input layer corresponds to an input channel.

The input layer corresponding to the first input channel is initialized with the word2vec algorithm, and does not participate in the training of the CNN classification model, but classifies the content input by the first input channel; the input layer corresponding to the second input channel is initialized randomly, not only for the second The content input by the input channel is classified and participates in the training of the CNN classification model.

The sentence is input into the input layer of the CNN classification model through the input channel.

S103, extracting the word2vec feature of the sentence, and obtaining an input matrix;

S104. The input matrix is input into a convolution layer, and features are extracted through a convolution operation; wherein, the weight of the convolution layer is initialized with a Gaussian distribution parameter value;

Choose to initialize the weights of the convolutional layers with a Gaussian distribution with zero mean and small standard deviation.

S105. Input the features into a classifier for classification.

The feature dimension extracted in this embodiment is 256 dimensions, and the classifier is a softmax classifier. Through the external softmax classifier, it is classified according to the extracted 256-dimensional features.

It can be seen from the above technical solution that in this embodiment, on the basis of the pre-established CNN classification model, the method of initializing the weight of the convolutional layer is improved, specifically, the weight is initialized according to the Gaussian distribution. Compared with the existing classification based on CNN The model realizes the method of text classification, which improves the accuracy of classification results. And compared with Naive Bayesian, machine learning algorithms such as SVM also improve the accuracy of classification results.

This embodiment discloses another method for text classification, and introduces in detail the method for preprocessing the text to be classified. Referring to Fig. 2, this embodiment includes the following steps:

S201. Determine whether the length of the text to be classified is greater than a preset length; wherein, the text to be classified includes a title and a text;

If it is judged that the length of the text to be classified is greater than the preset length, step S202 is executed;

S202. Truncating the text to be classified according to the preset length to obtain multiple sentences;

For example, if the text to be classified is "shenqingwenjian" and the preset length is 4, "shenqingwenjian" is truncated, and the truncated results are "shen", "qing", "wenj", and "ian"; each truncated result is a Sentence, by performing a truncation operation on the file to be classified, multiple sentences are obtained.

S203. Determine whether the length of the sentence is less than a preset length;

If it is judged that the length of the sentence is less than the preset length, step S204 is executed;

Otherwise, execute step S205;

Multiple sentences can be obtained by performing a truncation operation on the file to be classified, wherein, since the length of the text to be classified may not be an integer multiple of the preset length, the length of the last sentence (“ian”) obtained by the truncation operation may not be Equal to the preset length, and less than the preset length, then execute step S204 on the last sentence ("ian");

In this embodiment, since the truncation of the text to be classified is truncated according to the preset length, among the multiple sentences obtained, only the length of the last sentence may not be equal to the preset length, so it is judged whether the length of the sentence is less than the preset length. In the step of setting the length, it may only be judged whether the length of the sentence is less than the preset length.

S204. Splicing behind the sentence according to the order of the content included in the text to be classified, until the length of the new sentence formed after splicing is equal to the preset length;

The length of "ian" is less than the preset length 4, therefore, the splicing operation needs to be performed. For "ian" splicing, the splicing operation can be performed using the content of the text to be classified "shenqingwenjian", and the resulting splicing result is "ians". Of course, the splicing operation can also be performed using its own content, and the resulting splicing result is "iani". The length of is equal to the preset length.

The object of the splicing operation in this step is the sentence obtained after performing the truncation operation on the text to be classified. However, in this embodiment, when it is judged by step S201 that the length of the text to be classified is not greater than the preset length, there is the length of the text to be classified itself. If the length is less than the preset length, the truncation operation does not need to be performed, but the splicing operation needs to be performed. Wherein, the specific implementation manner of the splicing operation is similar to the implementation manner of performing the splicing operation on the truncated result.

Specifically, when the length of the text to be classified is less than the preset length, the text to be classified is spliced into a new text to be classified whose length is equal to the preset length through a splicing operation.

For example, if the text to be classified is "shenqing", the length of the text to be classified is 8, and the preset length is 10, then it will be spliced after "shenqing" according to the order of the content contained in the text to be classified.

The order of "shenqing" is s, h, e, n, q, i, n, g. The splicing method is to splice "s" after "shenqing", and the new text to be classified after splicing is "shenqings". The length after splicing is 9, which is still less than the preset length, and continue splicing; splicing "h" behind "shenqings", the new text to be classified after splicing is "shenqingsh", and the length after splicing is 10, which is the same as the preset The length is the same, and the splicing operation is completed. The final new text to be classified is "shenqingsh".

When the preset length is 20, the new text to be classified obtained according to the above splicing method is "shenqingshenqingshen".

In addition, in this embodiment, when it is judged by step S201 that the length of the text to be classified is not greater than the preset length, there is also a case where the length of the text to be classified is equal to the preset length, then not only does it not need to perform a truncation operation on the text to be classified, but also There is no need to perform concatenation operations on the text to be classified.

S205. Input the sentence into the input layer of the pre-established CNN classification model;

S206, extracting the word2vec feature of the sentence, and obtaining an input matrix;

S207. The input matrix is input into a convolution layer, and features are extracted through a convolution operation; wherein, the weights of the convolution layer are initialized with Gaussian distribution parameter values;

S208. Input the features into a classifier for classification.

In this embodiment, the implementation manner of step S205-step S208 is similar to the implementation manner of step S102-step S105 in the previous embodiment, and will not be repeated here.

It can be seen from the above technical solution that in this embodiment, on the basis of the pre-established CNN classification model, the input content input into the CNN classification model is preprocessed, specifically, the text to be classified whose length exceeds the preset length is processed according to the preset length. Set the length to be truncated, and the text to be classified whose length is less than the preset length is cyclically filled with the content of the text to be classified, and the preprocessed content is input into the CNN classification model, and the convolution in the CNN classification model is improved. The weight initialization method of the layer is specifically to initialize the weight according to the Gaussian distribution, which improves the accuracy of the classification result compared with the existing method of text classification based on the CNN classification model.

The text classification method disclosed in the foregoing embodiment is realized based on the CNN classification model that is pre-established and trained. The training method of the CNN classification model is described in detail below. Referring to FIG. 3, the training method of the CNN classification model includes the following steps:

S301. Obtain a data set and uncleaned data; wherein, the uncleaned data includes preset classification marks;

The data set is composed of three types of data: original training data, manually labeled data, and outstation data. Among them, different types of data account for different proportions of the total data set. Preferably, the number of original training data accounts for 10% of the total data set 80%, the amount of manually labeled data accounts for 5% of the total dataset, and the amount of external data accounts for 15% of the total dataset. The total amount of the dataset is set to be around 300,000.

In this embodiment, the original training data refers to the stored text of the user website; for example, the stored news of Sohu itself, which is stored in the distributed cluster.

Manually annotated data refers to data that has been manually edited and labeled for classification.

External site data refers to text captured from other websites or official accounts that are different from the user's website.

In this embodiment, the uncleaned data may be directly sampled from the acquired data set. Of course, it may also be data obtained from a different data set. However, no matter where the uncleaned data is obtained, it is necessary to ensure that the uncleaned data already contains preset classification marks.

In practical applications, when there is a large amount of uncleaned data, filter rules can be used to filter the uncleaned data, for example, set a blacklist of keywords, select some uncleaned data through this filter rule, and filter The selected uncleaned data are subjected to subsequent cleaning steps.

S302. Using the data set, perform initial training on the CNN classification model;

S303. Use the CNN classification model after initial training to perform classification prediction on the uncleaned data, and obtain the predicted classification label and prediction probability of the uncleaned data;

Taking the news classification as an example, the obtained predicted classification mark is the label corresponding to the category of the news to be classified, wherein the label of the news category can be set according to the actual business scenario, and the label corresponding to the news category set in this embodiment is from 1 to 101, that is, different labels correspond to different news categories. For example, sports news corresponds to label 1, financial news corresponds to label 2, and entertainment news corresponds to label 3.

The prediction probability is the possibility that the CNN classification model classifies and predicts the uncleaned data correctly. For example, after classifying and predicting news A, the data to be cleaned, the predicted classification mark is 1, that is, news A is predicted to be sports news, and the predicted probability is 0.9, which means that the CNN classification model judges that news A belongs to sports news The probability is 90%.

S304. Determine whether the predicted probability of the uncleaned data is greater than a preset probability value;

If it is judged that the predicted probability of the uncleaned data is greater than the preset probability value, step S305 is executed;

Judging that the predicted probability of the uncleaned data is not greater than the preset probability value, no processing is performed, which means that the result of this prediction is not credible;

The preset probability value is 99.5%, that is, when the prediction probability obtained by the CNN classification model is greater than 99.5%, it means that the prediction result of the CNN classification model is credible, and step S305 is executed;

If the prediction probability obtained by the CNN classification model is not greater than 99.5%, it means that the prediction result of the CNN classification model is not credible, and there is no need to pay attention to the predicted classification marks obtained by the CNN classification model.

S305. Judging whether the predicted classification mark of the uncleaned data is the same as the preset classification mark of the uncleaned data;

Judging that the predicted classification mark of the uncleaned data is different from the preset classification mark of the uncleaned data, execute step S306;

Judging that the predicted classification flag of the uncleaned data is the same as the preset classification flag of the uncleaned data, the preset classification flag of the uncleaned data is not modified.

When the result predicted by the CNN classification model is credible, it is judged whether the predicted classification mark of the uncleaned data is the same as the preset classification mark of the uncleaned data;

For the uncleaned data, if the predicted classification labels are the same as the preset classification labels, it means that the initial preset classification labels of the uncleaned data are correct and do not need to be modified; if the predicted classification labels are different from the preset classification labels, it means that there is no The initial preset classification mark of cleaning data is incorrect, execute step S306;

S306. Modify the preset classification mark of the uncleaned data to the predicted classification mark to obtain the cleaned data;

Modify the classification labels of the uncleaned data to predicted classification labels, complete the cleaning step, and obtain the cleaned data.

For example, if the predicted classification mark of news B is 2, and the prediction probability is greater than 99.5%, then the classification prediction of news B by the CNN classification model is credible, and the classification mark of news B should be 2. However, the preset classification mark of news B is 1, which is different from the predicted classification mark 2, which means that the preset classification mark of news B is noisy, which is wrong, so the preset classification mark of news B is changed to the predicted classification mark, The modified category mark for News B will be 2.

S307. Using the cleaned data, train the CNN classification model.

Use the data that has completed the cleaning step to update the training set of the CNN classification model to increase the training samples, and extract data in the same proportion from the training set before the update, put it into the test set, and use the CNN classification model for classification prediction. That is, by using the initial CNN classification model to clean the data, and using the cleaned data to assist the training of the further CNN classification model, the data is cleaned in a progressive manner layer by layer.

In this embodiment, in order to further improve the classification accuracy, it is not necessary to modify all the news classification marks whose predicted probability is greater than 99.5%, and the predicted classification marks are different from the preset classification marks, but according to the preset rules, from Among the uncleaned data whose prediction probability is greater than 99.5%, and whose predicted classification labels are different from the preset classification labels, select the uncleaned data to be processed; and modify the classification labels of the selected uncleaned data to be processed.

Preferably, the preset rule is uniform distribution.

Using the training method of the above-mentioned CNN classification model, when a large number of training samples are used to train the CNN classification model, it is possible to judge whether the classification marks of a large number of training samples are correct, and to modify incorrect classification marks to correct classification marks. The CNN classification model is trained by using training samples with correct classification marks to obtain a CNN classification model with high classification accuracy.

When actually training the CNN classification model, there will still be a small number of training samples. When the number of training samples is small, directly using small-scale training samples to train the CNN classification model will easily lead to overfitting and cause errors in classification results. .

Based on this, in this embodiment, if it is judged that the number of training samples is less than the preset number of samples, the pre-trained CNN classification model is reused; and then the training samples are used to train the pre-trained CNN classification model.

Wherein, the pre-trained CNN classification model is a trained CNN classification model in a scene similar to the training sample.

For example, in this embodiment, the long text of news needs to be classified, and a CNN classification model capable of classifying news needs to be trained, but at this time, the number of news with accurate classification marks is small. However, the CNN classification model used to classify long texts such as novels has been trained. Since both news and novels belong to long texts, the main structure and feature representation of the two are not much different, so the pre-trained user can be reused. A CNN classification model for classifying novels; then train a CNN classification model for classifying novels using news with accurate classification labels. By reusing the CNN classification model for classifying novels, there is no need to relearn all the parameters in the CNN classification model, and only need to use small-scale news samples to adjust some parameters in the reused CNN classification model in the iterative process, namely A CNN classification model for news classification can be trained. Solved the overfitting problem caused by model training due to insufficient number of samples.

Corresponding to the above text classification method, this embodiment also discloses a text classification device. Please refer to FIG. 4 for a schematic structural diagram of the text classification device. In this embodiment, the text classification device includes:

A preprocessing unit 401, an input unit 402, a first processing unit 403, a second processing unit 404 and a classification unit 405;

A preprocessing unit 401, configured to preprocess the text to be classified to obtain multiple sentences;

Wherein, the preprocessing unit 401 includes:

a first judging unit, a truncation unit, a second judging unit and a splicing unit;

The first judging unit is configured to judge whether the length of the text to be classified is greater than a preset length;

The truncation unit is configured to truncate the text to be classified according to the preset length to obtain multiple sentences when the first judging unit judges that the length of the text to be classified is greater than a preset length;

The second judging unit is used to judge whether the length of the sentence is less than a preset length;

The splicing unit is configured to, when the second judging unit judges that the length of the sentence is less than a preset length, splice the sentence after the sentence according to the order of the content contained in the text to be classified until the spliced The length of the new sentence is equal to the preset length.

The input unit 402 is used to input the sentence into the input layer of the pre-established CNN classification model;

The first processing unit 403 is used to extract the word2vec feature of the sentence and obtain the input matrix;

The second processing unit 404 is configured to input the input matrix into a convolutional layer, and extract features through a convolution operation; wherein, the weights of the convolutional layer are initialized with Gaussian distribution parameter values;

A classification unit 405, configured to input the features into a classifier for classification.

It can be seen from the above technical solution that in this embodiment, on the basis of the pre-established CNN classification model, the input content input into the CNN classification model is preprocessed, specifically, the text to be classified whose length exceeds the preset length is processed according to the preset length. Set the length to be truncated, and the text to be classified whose length is less than the preset length is cyclically filled with the content of the text to be classified, and the preprocessed content is input into the CNN classification model, and the convolution in the CNN classification model is improved. The weight initialization method of the layer is specifically to initialize the weight according to the Gaussian distribution, which improves the accuracy of the classification result compared with the existing method of text classification based on the CNN classification model.

On the basis of the text classification device disclosed in the previous embodiment, this embodiment also discloses another text classification device. The structural diagram of the text classification device is shown in Figure 5. In this embodiment, the text classification device also include:

Acquiring unit 501, training unit 502, predicting unit 503, third judging unit 504, fourth judging unit 505, modifying unit 506 and multiplexing unit 507;

An acquisition unit 501, configured to acquire a data set and uncleaned data; wherein, the uncleaned data includes preset classification marks;

A training unit 502, configured to use the data set to perform initial training on the CNN classification model;

The prediction unit 503 is configured to use the CNN classification model after initial training to perform classification prediction on the uncleaned data, and obtain the predicted classification label and prediction probability of the uncleaned data;

A third judging unit 504, configured to judge whether the predicted probability of the uncleaned data is greater than a preset probability value;

The fourth judging unit 505 is configured to judge the predicted classification flag of the uncleaned data and the preset Whether the classification marks are the same;

The modifying unit 506 is configured to modify the preset classification mark of the uncleaned data to The predicted classification marks are obtained by cleaning the data;

Judging that the predicted classification mark of the uncleaned data is the same as the preset classification mark of the uncleaned data, the preset classification mark of the uncleaned data is not modified;

The training unit 502 is further configured to use the cleaned data to train the CNN classification model;

Wherein, the modification unit 506 includes:

select subunits and modify subunits;

The selection subunit is used to select the uncleaned data to be processed from the uncleaned data whose predicted classification labels of the uncleaned data are judged to be different from the predicted classification labels of the uncleaned data according to preset rules;

A modifying subunit, configured to modify the preset classification mark of the uncleaned data to be the predicted classification mark;

The multiplexing unit 507 is used to reuse the pre-trained CNN classification model if the number of training samples is less than the preset sample number;

The training unit 502 is further configured to use the training samples to train the pre-trained CNN classification model.

After training the CNN classification model, you can use the trained CNN classification model in text classification.

It can be seen from the above technical solution that in this embodiment, on the basis of the pre-established CNN classification model, the input content input into the CNN classification model is preprocessed, specifically, the text to be classified whose length exceeds the preset length is processed according to the preset length. Set the length to be truncated, and the text to be classified whose length is less than the preset length is cyclically filled with the content of the text to be classified, and the preprocessed content is input into the CNN classification model, and the convolution in the CNN classification model is improved. The weight initialization method of the layer is specifically to initialize the weight according to the Gaussian distribution, which improves the accuracy of the classification result compared with the existing method of text classification based on the CNN classification model. Moreover, when training the CNN classification model, a large number of training samples are first cleaned, and then the CNN classification model is trained using the cleaned training samples, which can be trained to obtain an accurate CNN classification model. At the same time, when the number of training samples is small, Reusing other CNN classification models that have been trained, and then using training samples to train the reused CNN classification model can avoid the occurrence of over-fitting problems.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts of each embodiment can be referred to each other. As for the device provided in the embodiment, since it corresponds to the method provided in the embodiment, the description is relatively simple, and for the relevant details, please refer to the description of the method part.

It should be noted that, in this document, the terms "comprising", "comprising" or any other variation thereof are intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements, It also includes other elements not expressly listed, or elements inherent in the process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

The above is only a preferred embodiment of the present invention, it should be pointed out that, for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (10)

1. a kind of file classification method, which is characterized in that including：

It treats classifying text to be pre-processed, obtains multiple sentences；

The input layer of CNN disaggregated models that the sentence inputting is pre-established；

The word2vec features of the sentence are extracted, obtain input matrix；

The input matrix is inputted into convolutional layer, feature is extracted by convolution operation；Wherein, the weight of the convolutional layer is using high This distributed constant value is initialized；

Feature input grader is classified.

2. file classification method according to claim 1, which is characterized in that the classifying text for the treatment of is pre-processed, Multiple sentences are obtained to include：

Judge whether the length of the text to be sorted is more than preset length；

Judge the text to be sorted length be more than preset length, then by the text to be sorted according to the preset length into Row blocks, and obtains multiple sentences；

Judge whether the length of the sentence is less than preset length；

Judge that the length of the sentence is less than preset length, then the content order splicing included according to the text to be sorted exists Behind the sentence, until the length of new sentence formed after splicing is equal to the preset length.

3. according to claims 1 or 2 any one of them file classification method, which is characterized in that the instruction of the CNN disaggregated models Practice method to include：

It obtains data set and does not clean data；Wherein, the data of not cleaning include default classification marker；

Using the data set, initial training is carried out to the CNN disaggregated models；

Classification prediction is carried out to the data of not cleaning using the CNN disaggregated models after initial training, obtains described not cleaning number According to prediction classification marker and prediction probability；

Whether the prediction probability for not cleaning data described in judging is more than predetermined probabilities value；

The prediction probability for not cleaning data described in judging is more than predetermined probabilities value, then does not clean the prediction classification of data described in judgement Whether label is identical with the default classification marker for not cleaning data；

The prediction classification marker for not cleaning data described in judging is different from the default classification marker for not cleaning data, then by institute It states and does not clean the default classification markers of data and be revised as the prediction classification marker, the data after being cleaned；

Utilize the data after the cleaning, the training CNN disaggregated models.

4. file classification method according to claim 3, which is characterized in that described by do not clean data default point Class label is revised as the prediction classification marker and includes：

According to preset rules, the prediction classification marker for not cleaning data and the prediction for not cleaning data are obtained from judgement Classification marker is different not to clean in data, selects pending not clean data；

The pending default classification marker for not cleaning data is revised as the prediction classification marker.

5. file classification method according to claim 3, which is characterized in that the training method of the CNN disaggregated models is also Including：

If the quantity of training sample is less than default sample size, CNN disaggregated models trained in advance are utilized；

Using the training sample, the CNN disaggregated models trained in advance are trained.

6. a kind of document sorting apparatus, which is characterized in that including：

Pretreatment unit is pre-processed for treating classifying text, obtains multiple sentences；

Input unit, for the input layer of CNN disaggregated models for pre-establishing the sentence inputting；

First processing units for extracting the word2vec features of the sentence, obtain input matrix；

For the input matrix to be inputted convolutional layer, feature is extracted by convolution operation for second processing unit；Wherein, it is described The weight of convolutional layer is initialized using Gaussian Distribution Parameters value；

Taxon, for feature input grader to be classified.

7. document sorting apparatus according to claim 6, which is characterized in that the pretreatment unit includes：

First judging unit, for judging whether the length of the text to be sorted is more than preset length；

Block unit, for judge when first judging unit length of the text to be sorted be more than preset length when, will The text to be sorted is blocked according to the preset length, obtains multiple sentences；

Second judgment unit, for judging whether the length of the sentence is less than preset length；

Concatenation unit, for judge when the second judgment unit length of the sentence be less than preset length when, according to described The content order that text to be sorted includes splices behind the sentence, until length of new sentence formed after splicing etc. In the preset length.

8. according to claim 6 or 7 any one of them document sorting apparatus, which is characterized in that further include：

Acquiring unit, for obtaining data set and not cleaning data；Wherein, the data of not cleaning include default contingency table Note；

For utilizing the data set, initial training is carried out to the CNN disaggregated models for training unit；

Predicting unit for carrying out classification prediction to the data of not cleaning using the CNN disaggregated models after initial training, obtains The prediction classification marker for not cleaning data and prediction probability；

Third judging unit, for judging whether the prediction probability for not cleaning data is more than predetermined probabilities value；

4th judging unit, for work as the third judging unit judge the prediction probability for not cleaning data be more than it is default generally During rate value, judge described in do not clean the prediction classification markers of data and the default classification marker for not cleaning data whether phase Together；

Change unit, for work as the 4th judging unit judge the prediction classification marker for not cleaning data with it is described not clearly When washing the default classification marker difference of data, the default classification marker for not cleaning data is revised as the prediction contingency table Note, the data after being cleaned；

The training unit is additionally operable to utilize the data after the cleaning, the training CNN disaggregated models.

9. document sorting apparatus according to claim 8, which is characterized in that the modification unit includes：

Select subelement, for according to preset rules, from judge to obtain the prediction classification marker for not cleaning data with it is described Do not clean data prediction classification marker it is different do not clean in data, select pending not clean data；

Subelement is changed, for the pending default classification marker for not cleaning data to be revised as the prediction contingency table Note.

10. document sorting apparatus according to claim 8, which is characterized in that further include：

Multiplexing Unit if the quantity for training sample is less than default sample size, is multiplexed CNN classification moulds trained in advance Type；

The training unit is additionally operable to using the training sample, and the CNN disaggregated models trained in advance are trained.