CN111428028A - Information classification method based on deep learning and related equipment - Google Patents

Abstract

The present application relates to the technical field of data analysis, and in particular, to a deep learning-based information classification method and related equipment, including: obtaining the data quantity of the information to be identified, determining a clustering method of the information to be identified according to the data quantity, and applying the clustering method to treat the information to be identified. The identification information is preprocessed to obtain pre-classified data; the word vector conversion is performed on the pre-classified data to obtain the word vector of the pre-classified data; the word vector of the pre-classified data is input into the deep learning model for text feature extraction, and multiple Text features; classify each text feature to obtain the classification result of the text feature; use the voting mechanism to score the classification result, and determine the classification label of the information to be identified according to the scoring result. The present application effectively solves the problem of data imbalance, which leads to the problem that the content of the original information cannot be accurately reflected when the text feature extraction is performed by applying a deep learning model.

Description

Information classification method and related equipment based on deep learning

technical field

The present application relates to the technical field of data analysis, and in particular, to a deep learning-based information classification method and related equipment.

Background technique

Usually people can clearly understand multiple intentions expressed in the text, but it is difficult for the robot to understand all the intentions expressed in the text, which leads to incomplete answers given by the machine, and customers cannot obtain complete and satisfactory answers from the robot, even It is possible to return the wrong answer because it does not understand the multiple intentions of the customer, causing a very poor experience for the customer and reducing customer satisfaction. Therefore, it is an important task for the customer service robot to allow the machine to interpret the multiple intentions expressed by the customer.

At present, text classification is mainly used in multi-intent recognition. However, there is a problem of data imbalance in text classification, which makes it impossible to obtain the content that accurately reflects the original information when extracting text features by applying deep learning models.

SUMMARY OF THE INVENTION

Based on this, in view of the problem of data imbalance in the current text classification, the content of the original information cannot be accurately reflected when applying the deep learning model for text feature extraction, a deep learning-based information classification method is provided. and related equipment.

An information classification method based on deep learning, comprising the following steps:

Obtaining the data quantity of the information to be identified, determining a clustering method of the information to be identified according to the quantity of data, and applying the clustering method to preprocess the information to be identified to obtain pre-classified data;

performing word vector conversion on the pre-classified data to obtain a word vector of the pre-classified data;

Inputting the word vector of the pre-classified data into a preset deep learning model to perform text feature extraction to obtain multiple text features;

classifying the text features to obtain a classification result of the text features;

The classification result is scored by using a preset voting mechanism to obtain a scoring result, and the classification label of the information to be identified is determined according to the scoring result.

In one possible embodiment, the amount of data of the information to be identified is obtained, a clustering method of the information to be identified is determined according to the amount of data, and the information to be identified is preprocessed by applying the clustering method , get pre-classified data, including:

Compare the data quantity with a preset data quantity threshold, and if the data quantity is greater than the data quantity threshold, determine that the information to be identified is large sample data, otherwise determine that the information to be identified is small sample data ;

If the information to be identified is large sample data, after removing noise points and isolated points in the large sample data, apply a clustering algorithm to cluster the large sample data to obtain the pre-classified data;

If the to-be-identified information is small sample data, a clustering algorithm is used to cluster similar samples in the small sample data to generate multiple clusters, and the data in each cluster is processed by a genetic cross algorithm, respectively, The pre-classified data is obtained.

In one possible embodiment, performing word vector transformation on the pre-classified data to obtain a word vector of the pre-classified data, including:

Obtain a preset word vector embedding model, and divide the pre-classified data into multiple sentences according to the attributes of the word vector embedding model;

Inputting the sentence into the word vector embedding model for mapping to obtain an initial text word vector;

Calculate the eigenvalues of the initial text word vectors, delete the initial text word vectors whose eigenvalues are zero, summarize the remaining initial text word vectors, and obtain the word vectors of the pre-classified data.

In one possible embodiment, the word vector of the pre-classified data is input into a preset deep learning model to perform text feature extraction to obtain multiple text features, including:

Input the preset standard word vector into the input layer in the preset cyclic neural network model, and perform probability prediction on the word vector processed by the input layer through the hidden layer in the cyclic neural network model to obtain the probability prediction result , and applying the output layer in the cyclic neural network model to convert the probability prediction result to obtain prediction keywords;

Compare the predicted keywords with the keywords corresponding to the standard word vectors, and if they are consistent, input the word vectors of the pre-classified data into the cyclic neural network model for feature extraction, otherwise change the The parameters in the hidden layer are re-predicted until the predicted keyword is consistent with the keyword corresponding to the standard word vector.

In one possible embodiment, the classification of the text features to obtain a classification result of the text features includes:

Obtain classifiers of different categories, and establish a classifier subtree according to the hierarchical relationship between the classifiers;

The text feature is input into the root node of the classifier subtree, first classification is performed, the first classification result is obtained, and the first classification result is input into the next-level leaf node of the root node;

Continue to classify with the next-level leaf node as a new root node, until the next-level leaf node is the smallest leaf node;

Summarize the classification results of the minimum leaf nodes to obtain the classification results of the text features.

In one possible embodiment, the classification result is scored by applying a preset voting mechanism to obtain a scoring result, and the classification label of the to-be-identified information is determined according to the scoring result, including:

Obtain the classification accuracy rate of the end classifier corresponding to each of the minimum leaf nodes, and use the classification accuracy rate as the weight of the end classifier;

Using the weight as an auxiliary parameter, applying the voting mechanism to vote and score the classification label output by the end classifier;

A classification label with a voting score greater than a score threshold is extracted as the classification label of the information to be identified.

In one possible embodiment, the classification continues with the next-level leaf node as a new root node until the next-level leaf node is the smallest leaf node, including:

Obtain the similarity between the output results of all leaf nodes at any level, and extract the target leaf nodes corresponding to multiple output results whose similarity is greater than the similarity threshold as the root node of the next level classification;

The node label corresponding to the target leaf node is acquired, and the node label is input into the next-level classifier for classification, until the root node of the next-level classification is the minimum leaf node.

An information classification device based on deep learning, comprising the following modules:

The pre-classification module is configured to obtain the data quantity of the information to be identified, determine the clustering method of the information to be identified according to the data quantity, and apply the clustering method to preprocess the information to be identified to obtain pre-classified data ;

A word vector module, configured to perform word vector conversion on the pre-classified data to obtain a word vector of the pre-classified data;

A feature extraction module, configured to input the word vectors of the pre-classified data into a preset deep learning model to perform text feature extraction to obtain multiple text features;

A result generation module, configured to classify each text feature, and obtain a classification result of the text feature;

The label generation module is configured to use a preset voting mechanism to score the classification result, obtain a score result, and determine the classification label of the to-be-identified information according to the score result.

A computer device, comprising a memory and a processor, wherein computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, the processor is made to execute the above-mentioned deep learning-based information classification method A step of.

A storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the above-mentioned deep learning-based information classification method.

Compared with the existing mechanism, the present application obtains the data quantity of the information to be identified, determines the clustering method of the to-be-identified information according to the data quantity, and applies the clustering method to preprocess the to-be-identified information, Obtain pre-classified data; perform word vector transformation on the pre-classified data to obtain a word vector of the pre-classified data; input the word vector of the pre-classified data into a preset deep learning model to perform text feature extraction, and obtain multiple text features; classify each text feature to obtain a classification result of the text feature; apply a preset voting mechanism to score the classification result to obtain a scoring result, and determine the information to be identified according to the scoring result Category labels. It can effectively solve the problem of data imbalance, which leads to the problem that the content of the original information cannot be accurately reflected when applying the deep learning model for text feature extraction.

Description of drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustrating preferred embodiments only and are not to be considered limiting of the application.

1 is an overall flowchart of a deep learning-based information classification method in an embodiment of the present application;

2 is a schematic diagram of a pre-classification process in a deep learning-based information classification method in an embodiment of the present application;

3 is a schematic diagram of a result generation process in a deep learning-based information classification method in an embodiment of the present application;

FIG. 4 is a structural diagram of an apparatus for classifying information based on deep learning in an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

It will be understood by those skilled in the art that the singular forms "a", "an", "the" and "the" as used herein can include the plural forms as well, unless expressly stated otherwise. It should be further understood that the word "comprising" used in the specification of this application refers to the presence of stated features, integers, steps, operations, elements and/or components, but does not preclude the presence or addition of one or more other features, Integers, steps, operations, elements, components and/or groups thereof.

Fig. 1 is the overall flow chart of a kind of information classification method based on deep learning in one embodiment of the application, a kind of information classification method based on deep learning, comprises the following steps:

S1. Obtain the data quantity of the information to be identified, determine a clustering method of the information to be identified according to the quantity of data, and apply the clustering method to preprocess the information to be identified to obtain pre-classified data;

Specifically, when performing information identification, it is necessary to determine whether the information to be identified is single-intent or multi-intent. Among them, single-intent means that the text sentence contains only one intention, such as: "I want to listen to Jay Chou's song", this sentence intends It can be attributed to musical intent, while multi-intent means that the text utterance can contain multiple intents, such as: "I want to buy apples", the intent of this sentence can be attributed to fruit-like intent, I want to buy fruit, and it can also be attributed to For the electronic intent, I want to buy an iPhone, and then based on the user's previous search records or contextual information, determine which intent the sentence belongs to in this conversation, and return the best answer first. , when the data volume of the information to be recognized is counted, the natural language recognition algorithm is required to classify the single-sentence intent and the multi-sentence intent in the to-be-recognized information. The single sentence corresponding to each single-sentence intent is regarded as one data, and each Sentences together as 1 data. When clustering, the clustering algorithms used are mainly K-mean algorithm and agglomerative hierarchical clustering algorithm.

S2, performing word vector conversion on the pre-classified data to obtain a word vector of the pre-classified data;

Specifically, Wordvec2 is usually used for word vector conversion. In this step, the BERT model is used as the word vector embedding model. When using the BERT model, the text can be converted into word vectors in the following ways: 1. Install the BERT model: ( 1) Install BERT on the server side; (2) Install BERT on the client side; 2. Start the service: execute the following code bert-serving-start-model_dir/tmp/engli sh_L-12_H-768_A-12-num_worker=4, where /tmp/english_L-12_H-768_A-12/ is the path of the downloaded model 3. Use python script for text vectorization. The executed code is: from bert _serving.client import BertClient bc=BertClient() bc.encode(['Which insurance is good', 'What kind of insurance is good for children'])

S3. The word vector of the pre-classified data is input into a preset deep learning model to perform text feature extraction to obtain multiple text features;

Specifically, the deep learning model selected for text extraction is usually a convolutional neural network model or a recurrent neural network model. When extracting text features from word vectors, the above models must be trained first. After the text feature accuracy rate of the deep learning model is greater than a preset threshold, the deep learning model can be used to extract text features from the pre-classified word vectors. In this step, the word frequency-inverse document frequency algorithm (TF-IDF) can be used to extract text features.

S4, classify each text feature to obtain a classification result of the text feature;

Specifically, the text feature classification uses a computer to automatically classify and mark the text set (or other entities or objects) according to a certain classification system or standard. Commonly used text feature classification methods are Naive Bayes classification method, decision tree method and SVM support vector machine and so on. Classifiers need to be trained when classifying text features, and only the results classified by the validated classifiers can be used as reliable results.

Among them, the decision tree method is to use the decision tree model to classify text features, first input the text features to the root node of the decision tree model for the first classification, and then use the first leaf node for the second classification, and so on until The smallest leaf node of a decision tree model. The decision tree model can classify text features from coarse to fine, so as to obtain more accurate classification results. If the text feature is: orange, the root node in the decision tree model is "biology", the first leaf node is "plant", and the smallest leaf node is "fruit".

S5. Use a preset voting mechanism to score the classification result, obtain a scoring result, and determine a classification label of the information to be identified according to the scoring result.

Among them, the voting mechanism (voting) is a combination strategy for classification problems in ensemble learning. The basic idea is to select the class with the most output among all machine learning algorithms. There are two types of outputs of machine learning classification algorithms: one is to output class labels directly, and the other is to output class probabilities. Using the former for voting is called Majority/Hard voting, and using the latter for classification is called soft voting ( Soft voting). In this step, the soft voting mechanism is adopted, and the weight is added as an auxiliary parameter in the voting, so that the classification label can be better obtained.

For the soft voting mechanism, the following steps can be taken: first obtain the class probability output by the machine learning algorithm used, such as 50% for class A, 30% for class B, and 20% for class C, and then obtain the weight value of each class , such as class A is 0.3, class B is 0.5, class C is 0.2, the weighted average of each class is calculated, and the class with a larger value will be selected.

In this embodiment, by pre-classifying the information that needs to be used for intent recognition, and using different processing methods according to different types, the problem of data imbalance can be effectively solved, resulting in that the text feature extraction cannot be accurately obtained by applying a deep learning model. Questions that reflect the content of the original information.

2 is a schematic diagram of a pre-classification process in an information classification method based on deep learning in an embodiment of the present application. As shown in the figure, in S1, the data quantity of the information to be identified is obtained, which is determined according to the data quantity The clustering method of the to-be-identified information, using the clustering method to preprocess the to-be-identified information to obtain pre-classified data, including:

S11. Compare the data quantity with a preset data quantity threshold. If the data quantity is greater than the data quantity threshold, determine that the to-be-identified information is large sample data, otherwise, determine that the to-be-identified information is small sample;

Among them, usually for different data sets, the data volume thresholds for large and small samples are set differently. Generally, indicators such as median, mode, and average can be used as data volume thresholds. If the sample data volume is greater than the threshold, it is determined as a large sample, and vice versa. judged to be a small sample.

S12. If the information to be identified is large sample data, apply a clustering algorithm to cluster the large sample data after removing noise points and isolated points in the large sample data to obtain the pre-classified data;

Specifically, large sample data is data that conforms to a normal distribution. When clustering large sample data, it is necessary to remove non-text noise points such as special symbols ";" and "." in the large sample data, and if If there are points that do not conform to the normal distribution in the large sample data, then these points need to be removed as outliers before clustering.

S13. If the to-be-identified information is small sample data, use a clustering algorithm to cluster similar samples in the small sample data to generate multiple clusters, and use a genetic cross algorithm for the data in each cluster. processing to obtain the pre-classified data.

Specifically, since the small sample data contains less feature information than the large sample data, the machine learns relatively less feature information from the small sample, and the calculated joint probability is relatively small. Therefore, clustering algorithms such as K-mean cannot be directly used for classification and statistics, which will seriously affect the classification accuracy. When applying the genetic crossover algorithm for processing, the used crossover operator is the uniform crossover operator. Two crossover points are randomly generated in the two ligands A and B, and then three integers of 0, 1, and 2 are randomly generated for gene exchange, thereby forming two new individuals and completing the crossover.

In this embodiment, different clustering algorithms are used for sample classification of the data to be analyzed, thereby avoiding inaccurate information classification caused by data imbalance.

In one embodiment, performing word vector conversion on the pre-classified data to obtain a word vector of the pre-classified data, including:

Obtain a preset word vector embedding model, and divide the pre-classified data into multiple sentences according to the attributes of the word vector embedding model;

Specifically, different word vector embedding models have different restrictions on the character length for word vector conversion.

Inputting the sentence into the word vector embedding model for mapping to obtain an initial text word vector;

Calculate the eigenvalues of the initial text word vectors, delete the initial text word vectors whose eigenvalues are zero, summarize the remaining initial text word vectors, and obtain the word vectors of the pre-classified data.

In this embodiment, BERT is used for word vector embedding. The model adopts the Transformer sequence model, which has a bidirectional function, and can obtain sentence-level semantic representation higher than words, and has the advantages of strong universality and good effect.

In one embodiment, the word vector of the pre-classified data is input into a preset deep learning model to perform text feature extraction to obtain multiple text features, including:

Input the preset standard word vector into the input layer in the preset cyclic neural network model, and perform probability prediction on the word vector processed by the input layer through the hidden layer in the cyclic neural network model to obtain the probability prediction result , and applying the output layer in the cyclic neural network model to convert the probability prediction result to obtain prediction keywords;

Among them, the recurrent neural network model includes an input layer, a hidden layer and an output layer, wherein the input layer is used to receive data, the hidden layer is used to process the data, and the output layer is used to output the results. Among them, a series of processing will be performed on the data in the hidden layer, mainly including: gradient truncation, regularization, gating, etc. Data is efficiently processed through the hidden layer.

Compare the predicted keywords with the keywords corresponding to the standard word vectors, and if they are consistent, input the word vectors of the pre-classified data into the cyclic neural network model for feature extraction, otherwise change the The parameters in the hidden layer are re-predicted until the predicted keyword is consistent with the keyword corresponding to the standard word vector.

In this embodiment, the text feature extraction is performed on the word vector through the deep learning model, which ensures the accuracy of the text feature.

FIG. 3 is a schematic diagram of a result generation process in an information classification method based on deep learning in an embodiment of the present application. As shown in the figure, the step S4 is to classify the text features to obtain the classification of the text features. Results, including:

S41, obtain classifiers of different categories, and establish a classifier subtree according to the hierarchical relationship between the classifiers;

Specifically, a probability calculation method can be used when performing hierarchical classification, as follows:

P(n _k |d _i )=∏p(n _j |d _i )xa _k

where P(n _k |d _i ) represents the probability of document d _i classification finally reaching node n _k , P(n _j |d _i ) represents the probability that document d _i passes through ancestor node n _j before reaching node n _k , a _k represents the height penalty factor.

S42, input the text feature into the root node of the classifier subtree, perform first classification, obtain the first classification result, and input the first classification result into the next-level leaf node of the root node;

S43, continue to classify with the next-level leaf node as a new root node, until the next-level leaf node is the smallest leaf node;

Specifically, the similarity between the output results of all leaf nodes at any level is obtained, and the target leaf nodes corresponding to multiple output results whose similarity is greater than the similarity threshold are extracted as the root node of the next level classification;

The node label corresponding to the target leaf node is acquired, and the node label is input into the next-level classifier for classification, until the root node of the next-level classification is the minimum leaf node.

Among them, the similarity threshold is set according to the type of the classifier, such as: judging the probability that the classifier belongs to task type, small talk, and FAQ, if the calculated probabilities are 0.8, 0.3, and 0.9, the set threshold is 0.5, Then it can be judged that the text data belongs to the task type and chat, and the chat belongs to the leaf node, then the judgment will not continue, and then the next-level classifier is used to calculate the probability that the text belongs to the customer service task type and the customer service task type. If the calculated probability are 0.88 and 0.21 respectively, then the final judgment results are customer service task type and small talk.

S44. Summarize the classification results of the minimum leaf nodes to obtain the classification results of the text features.

In this embodiment, by adopting the improved hierarchical classification method, the problem that all subsequent classifications are erroneous due to classification errors occurring in the above-level classifications is avoided.

In one embodiment, the classification result is scored by applying a preset voting mechanism to obtain a scoring result, and the classification label of the to-be-identified information is determined according to the scoring result, including:

Obtain the classification accuracy rate of the end classifier corresponding to each of the minimum leaf nodes, and use the classification accuracy rate as the weight of the end classifier;

Among them, the calculation formula of classification accuracy is:

In the formula, TP is the true example, TN is the true negative example, FP = false positive example, and FN = false negative example.

Using the weight as an auxiliary parameter, applying the voting mechanism to vote and score the classification label output by the end classifier;

A classification label with a voting score greater than a score threshold is extracted as the classification label of the information to be identified.

In this embodiment, the voting mechanism is used to effectively score the classification results, thereby greatly improving the accuracy of the classification labels.

The technical features mentioned in any of the above-mentioned corresponding embodiments or implementation manners are also applicable to the embodiment corresponding to FIG. 4 in the present application, and the similarities will not be repeated hereafter.

A method for classifying information based on deep learning in the present application is described above, and the following describes an apparatus for performing the above-mentioned information classification based on deep learning.

As shown in FIG. 4 , a structure diagram of an information classification device based on deep learning can be applied to information classification based on deep learning. The apparatus for classifying information based on deep learning in this embodiment of the present application can implement steps corresponding to the method for classifying information based on deep learning performed in the embodiment corresponding to FIG. 1 . The functions implemented by the deep learning-based information classification apparatus may be implemented by hardware, or by executing corresponding software in hardware. The hardware or software includes one or more modules corresponding to the above functions, and the modules may be software and/or hardware.

In one embodiment, an information classification device based on deep learning is proposed, as shown in FIG. 4 , including the following modules:

The pre-classification module 10 is configured to obtain the data quantity of the information to be identified, determine the clustering method of the to-be-identified information according to the data quantity, and apply the clustering method to pre-process the to-be-identified information to obtain a pre-classification data;

The word vector module 20 is configured to perform word vector conversion on the pre-classified data to obtain the word vector of the pre-classified data;

The feature extraction module 30 is configured to input the word vector of the pre-classified data into a preset deep learning model to perform text feature extraction to obtain a plurality of text features;

The result generation module 40 is configured to classify the text features to obtain the classification results of the text features;

The label generation module 50 is configured to use a preset voting mechanism to score the classification result, obtain a score result, and determine the classification label of the to-be-identified information according to the score result.

In one embodiment, a computer device is proposed. The computer device includes a memory and a processor, and computer-readable instructions are stored in the memory. When the computer-readable instructions are executed by the processor, the processor causes the processor to execute the foregoing embodiments. The steps in the deep learning-based information classification method.

In one embodiment, a storage medium storing computer-readable instructions is provided, and when the computer-readable instructions are executed by one or more processors, the one or more processors cause the one or more processors to execute all of the above-mentioned embodiments. Describe the steps of a deep learning-based information classification method. The storage medium may be a non-volatile storage medium or a volatile storage medium, which is not specifically limited in this application.

Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium can include: Read Only Memory (ROM, Read Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk, etc.

The technical features of the above-described embodiments can be combined arbitrarily. In order to simplify the description, all possible combinations of the technical features in the above-described embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, all It should be considered as the range described in this specification.

The above-mentioned embodiments only represent some exemplary embodiments of the present application, and the descriptions are relatively specific and detailed, but should not be construed as limiting the patent scope of the present application. It should be pointed out that for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the scope of protection of the patent of the present application shall be subject to the appended claims.

Claims (10)

1. An information classification method based on deep learning is characterized by comprising the following steps:

acquiring the data quantity of information to be identified, determining the clustering mode of the information to be identified according to the data quantity, and preprocessing the information to be identified by applying the clustering mode to obtain pre-classified data;

performing word vector conversion on the pre-classified data to obtain word vectors of the pre-classified data;

inputting the word vectors of the pre-classified data into a preset deep learning model for text feature extraction to obtain a plurality of text features;

classifying the text features to obtain a classification result of the text features;

and scoring the classification result by using a preset voting mechanism to obtain a scoring result, and determining the classification label of the information to be identified according to the scoring result.

2. The information classification method based on deep learning of claim 1, wherein the obtaining of the data quantity of the information to be identified, the determining of the clustering mode of the information to be identified according to the data quantity, and the preprocessing of the information to be identified by applying the clustering mode to obtain pre-classification data comprises:

comparing the data quantity with a preset data quantity threshold, if the data quantity is greater than the data quantity threshold, determining that the information to be identified is large sample data, otherwise, determining that the information to be identified is small sample data;

if the information to be identified is large sample data, clustering the large sample data by applying a clustering algorithm after removing noise points and isolated points in the large sample data to obtain pre-classified data;

if the information to be identified is small sample data, clustering similar samples in the small sample data by using a clustering algorithm to generate a plurality of clusters, and processing the data in each cluster by respectively adopting a genetic crossover algorithm to obtain the pre-classification data.

3. The method for information classification based on deep learning of claim 1, wherein the performing word vector transformation on the pre-classified data to obtain a word vector of the pre-classified data comprises:

acquiring a preset word vector embedding model, and dividing the pre-classified data into a plurality of sentences according to the attribute of the word vector embedding model;

inputting the sentence into the word vector embedding model for mapping to obtain an initial text word vector;

and calculating the characteristic value of the initial text word vector, deleting the initial text word vector with the characteristic value of zero, summarizing the rest initial text word vectors, and obtaining the word vector of the pre-classified data.

4. The information classification method based on deep learning of claim 1, wherein the step of inputting the word vector of the pre-classification data into a preset deep learning model for text feature extraction to obtain a plurality of text features comprises:

inputting a preset standard word vector into an input layer in a preset cyclic neural network model, performing probability prediction on the word vector processed by the input layer through a hidden layer in the cyclic neural network model to obtain a probability prediction result, and converting the probability prediction result by using an output layer in the cyclic neural network model to obtain a prediction keyword;

and comparing the predicted key words with the keywords corresponding to the standard word vectors, if the predicted key words are consistent with the keywords corresponding to the standard word vectors, adding the word vectors of the pre-classified data into the cyclic neural network model for feature extraction, and otherwise, changing the parameters in the hidden layer for re-prediction until the predicted key words are consistent with the keywords corresponding to the standard word vectors.

5. The information classification method based on deep learning according to any one of claims 1 to 4, wherein the classifying the text features to obtain a classification result of the text features includes:

obtaining classifiers of different categories, and establishing a classifier sub-tree according to the hierarchical relation among the classifiers;

inputting the text features into a root node of the classifier subtree, performing primary classification to obtain a primary classification result, and inputting the primary classification result into a next-level leaf node of the root node;

taking the next-level leaf node as a new root node to continue classifying until the next-level leaf node is the minimum leaf node;

and summarizing the classification result of the minimum leaf node to obtain the classification result of the text features.

6. The information classification method based on deep learning of claim 5, wherein the applying a preset voting mechanism to score the classification result to obtain a scoring result, and determining the classification label of the information to be identified according to the scoring result comprises:

obtaining the classification accuracy of the end classifier corresponding to each minimum leaf node, and taking the classification accuracy as the weight of the end classifier;

voting and scoring the classification labels output by the end classifier by applying the voting mechanism by taking the weights as auxiliary parameters;

and extracting the classification label with the voting score larger than a score threshold value as the classification label of the information to be identified.

7. The method for classifying information based on deep learning according to claim 5, wherein the classifying continues with the next-level leaf node as a new root node until the next-level leaf node is a minimum leaf node, including:

acquiring the similarity among the output results of all leaf nodes at any level, and extracting target leaf nodes corresponding to a plurality of output results with the similarity larger than a similarity threshold value as root nodes of next-level classification;

and acquiring a node label corresponding to the target leaf node, and inputting the node label into a next-level classifier for classification until a root node of the next-level classification is the minimum leaf node.

8. An information classification device based on deep learning is characterized by comprising the following modules:

the pre-classification module is used for acquiring the data quantity of the information to be identified, determining the clustering mode of the information to be identified according to the data quantity, and preprocessing the information to be identified by applying the clustering mode to obtain pre-classification data;

the word vector module is used for carrying out word vector conversion on the pre-classified data to obtain word vectors of the pre-classified data;

the feature extraction module is used for inputting the word vectors of the pre-classified data into a preset deep learning model to extract text features so as to obtain a plurality of text features;

the result generation module is used for classifying the text features to obtain the classification result of the text features;

and the label generation module is set to score the classification result by applying a preset voting mechanism to obtain a scoring result, and determine the classification label of the information to be identified according to the scoring result.

9. A computer device comprising a memory and a processor, the memory having stored therein computer-readable instructions, which, when executed by the processor, cause the processor to carry out the steps of the deep learning based information classification method according to any one of claims 1 to 7.

10. A storage medium storing computer readable instructions which, when executed by one or more processors, cause the one or more processors to perform the steps of the deep learning based information classification method according to any one of claims 1 to 7.

Images