1. The utility model provides an enterprise industry second grade trade multi-label classifier based on deep learning algorithm which characterized in that: the device comprises an acquisition module, a preprocessing module, a management module, a model establishing module, a training verification module, an input module and a display module, wherein:

the acquisition module is used for acquiring enterprise operation range information;

the preprocessing module is used for preprocessing the enterprise operation range information;

the management module is used for manually indexing the enterprise operation range information and making a training set, a verification set and a test set for multi-label classification training;

the model establishing module is used for establishing an Albert + TextCNN model by using a training set;

the training verification module is used for training the established Albert + TextCNN model and verifying the accuracy;

the input module is used for inputting the enterprise information to be predicted into the trained model;

the display module is used for displaying the multi-label classification result of the enterprise industry.

2. The deep learning algorithm-based enterprise industry secondary industry multi-label classifier according to claim 1, wherein: the manual indexing method comprises the following steps:

the method comprises the following steps: collecting enterprise operation range information;

step two: extracting the collected enterprise operation range information;

step three: and classifying and indexing the extracted enterprise operation range information.

3. The deep learning algorithm-based enterprise industry secondary industry multi-label classifier according to claim 1, wherein: the enterprise business management system further comprises a storage module, and the storage module is used for storing the enterprise business management range information.

4. The deep learning algorithm-based enterprise industry secondary industry multi-label classifier according to claim 1, wherein: and preprocessing the enterprise operation range information, wherein the preprocessing comprises data information noise removal, data information cleaning and feature extraction.

5. The deep learning algorithm-based enterprise industry secondary industry multi-label classifier according to claim 1, wherein: the multi-label classification system further comprises a visual display module which is used for visually displaying the multi-label classification result.

6. The deep learning algorithm-based enterprise industry secondary industry multi-label classifier according to claim 1, wherein: during prediction, an approximate range is determined by simply screening, then, prediction is carried out by using the models of the corresponding groups, and manual screening is carried out if the situation which cannot be determined occurs after prediction.

7. The deep learning algorithm-based enterprise industry secondary industry multi-label classifier according to claim 1, wherein: the multi-label classification method comprises the following steps:

the method comprises the following steps: collecting the operation range information of the enterprise, and preprocessing the operation range information of the enterprise;

step two: carrying out manual indexing on the enterprise operation range information, and establishing a training set by adopting a manual indexing mode;

step three: setting parameters needed to be used by the Albert + TextCNN model;

step four: converting the operation range information and the manually indexed label into word vectors according to the application rule of Albert, and sending the word vectors into an Albert model for learning;

step five: the results of Albert and the initially transformed word vectors are sent into a TextCNN model for training;

step six: constructing a Full connection layer (Full connection layer) by the result output by the TextCNN, the result weight (output weight) and the result bias value (output bias) together, and storing the model;

step seven: judging whether debugging is needed or not according to the accuracy of the trained model, if so, readjusting the parameters needed to be set in the step three, and repeating the steps three to six until the final accuracy of the model is satisfied;

step eight: and carrying out multi-label classification on the industry of any enterprise by using the trained model.

Background

NLP natural language processing refers to developing applications or services that are capable of understanding human language.

The prior art is mainly divided into three categories, namely, unsupervised learning methods and semi-supervised learning methods, traditional machine learning methods and deep learning methods; wherein: unsupervised learning methods and semi-supervised learning methods are generally implemented by manually establishing standards and extracting features, such as using word frequency (TF), inverse file frequency (IDF), Logistic Regression (Logistic Regression), Decision Tree (Decision Tree), Mutual Information (Mutual Information), K-adjacent value (K-adjacent value), adaptive enhancement (AdaBoost) and multi-node language model (N-GRAM); the non-machine learning method has very low accuracy, needs a large amount of manual participation and is very inconvenient; the traditional machine learning methods mainly comprise SVM, CNN, RNN, LSTM and BERT algorithms, most of the methods need a large amount of training time, but some of the algorithms are not high in accuracy due to the problem that the release time is early or the algorithms are not suitable for text analysis.

The prior art has poor support for enterprise secondary industry classification and multi-label classification, the accuracy of the method used by the prior art is low, the prior art is difficult to complete the multi-label classification task, and the prior art is tedious, complex, troublesome and difficult to maintain;

in order to realize multi-label classification of enterprise secondary industry and solve the problem that the training time required by the existing method is too long, a deep learning algorithm-based multi-label classifier of enterprise secondary industry is provided.

Disclosure of Invention

The invention aims to provide an enterprise industry secondary industry multi-label classifier based on a deep learning algorithm, and aims to solve the problems that the prior art provided in the background art has poor support for enterprise secondary industry classification and multi-label classification, the method used in the prior art has low accuracy, the prior art is difficult to complete a multi-label classification task, and the prior art is complicated, troublesome and difficult to maintain.

In order to achieve the purpose, the invention provides the following technical scheme: the utility model provides an enterprise trade second grade trade multi-label classifier based on deep learning algorithm, comprises acquisition module, preprocessing module, management module, model building module, training verification module, input module, display module, wherein:

the acquisition module is used for acquiring enterprise operation range information;

the preprocessing module is used for preprocessing the enterprise operation range information;

the management module is used for manually indexing the enterprise operation range information and making a training set, a verification set and a test set for multi-label classification training;

the model establishing module is used for establishing an Albert + TextCNN model by using a training set;

the training verification module is used for training the established Albert + TextCNN model and verifying the accuracy;

the input module is used for inputting the enterprise information to be predicted into the trained model;

the display module is used for displaying the multi-label classification result of the enterprise industry.

As a preferred technical solution of the present invention, the manual indexing method comprises:

the method comprises the following steps: collecting enterprise operation range information;

step two: extracting the collected enterprise operation range information;

step three: and classifying and indexing the extracted enterprise operation range information.

The invention further comprises a storage module, wherein the storage module is used for storing the enterprise operation range information.

As a preferred technical scheme of the invention, the enterprise operation range information is preprocessed, and the preprocessing comprises data information noise removal, data information cleaning and feature extraction.

The visual display module is used for visually displaying the multi-label classification result.

As a preferred technical scheme of the invention, during prediction, an approximate range is determined by simply screening, then, the model of the corresponding group is used for prediction, and after prediction, manual screening is carried out if the situation which cannot be determined occurs.

As a preferred technical solution of the present invention, a multi-label classification method is as follows:

the method comprises the following steps: collecting the operation range information of the enterprise, and preprocessing the operation range information of the enterprise;

step two: carrying out manual indexing on the enterprise operation range information, and establishing a training set by adopting a manual indexing mode;

step three: setting parameters needed to be used by the Albert + TextCNN model;

step four: converting the operation range information and the manually indexed label into word vectors according to the application rule of Albert, and sending the word vectors into an Albert model for learning;

step five: the results of Albert and the initially transformed word vectors are sent into a TextCNN model for training;

step six: constructing a Full connection layer (Full connection layer) by the result output by the TextCNN, the result weight (output weight) and the result bias value (output bias) together, and storing the model;

step seven: judging whether debugging is needed or not according to the accuracy of the trained model, if so, readjusting the parameters needed to be set in the step three, and repeating the steps three to six until the final accuracy of the model is satisfied;

step eight: and carrying out multi-label classification on the industry of any enterprise by using the trained model.

Compared with the prior art, the invention has the beneficial effects that:

(1) the multi-label classification of the enterprise second-level industry is realized, and the problem that the training time required by the existing method is too long is solved;

(2) due to the modular design, single label classification can be performed, and the accuracy is greatly improved compared with the existing method.

Drawings

FIG. 1 is a schematic diagram of the present invention;

FIG. 2 is a flow chart of a method of multi-label classification in accordance with the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example 1

Referring to fig. 1 and fig. 2, the present invention provides a technical solution: the utility model provides an enterprise trade second grade trade multi-label classifier based on deep learning algorithm, comprises acquisition module, preprocessing module, management module, model building module, training verification module, input module, display module, wherein:

the acquisition module is used for acquiring enterprise operation range information;

the preprocessing module is used for preprocessing the enterprise operation range information;

the management module is used for manually indexing the enterprise operation range information and making a training set, a verification set and a test set for multi-label classification training;

the model establishing module is used for establishing an Albert + TextCNN model by using a training set;

the training verification module is used for training the established Albert + TextCNN model and verifying the accuracy;

the input module is used for inputting the enterprise information to be predicted into the trained model;

the display module is used for displaying the multi-label classification result of the enterprise industry.

The manual indexing method is as follows:

the method comprises the following steps: collecting enterprise operation range information;

step two: extracting the collected enterprise operation range information;

step three: and classifying and indexing the extracted enterprise operation range information.

In this embodiment, preferably, the enterprise operation range information is preprocessed, where the preprocessing includes removing data information noise, cleaning data information, and extracting features, and the influence of redundant data is removed.

In this embodiment, preferably, the multi-label classification system further includes a visual display module, and the visual display module is configured to visually display the multi-label classification result.

In this embodiment, preferably, the multi-label classification method is as follows:

the method comprises the following steps: collecting the operation range information of the enterprise, and preprocessing the operation range information of the enterprise;

step two: carrying out manual indexing on the enterprise operation range information, and establishing a training set by adopting a manual indexing mode;

step three: setting parameters needed to be used by the Albert + TextCNN model;

step four: converting the operation range information and the manually indexed label into word vectors according to the application rule of Albert, and sending the word vectors into an Albert model for learning;

step five: the results of Albert and the initially transformed word vectors are sent into a TextCNN model for training;

step six: constructing a Full connection layer (Full connection layer) by the result output by the TextCNN, the result weight (output weight) and the result bias value (output bias) together, and storing the model;

step seven: judging whether debugging is needed or not according to the accuracy of the trained model, if so, readjusting the parameters needed to be set in the step three, and repeating the steps three to six until the final accuracy of the model is satisfied;

step eight: and carrying out multi-label classification on the industry of any enterprise by using the trained model.

The national economic industry classification GB/T4754-2017 standard divides an enterprise into four levels of a gate class, a major class, a middle class and a minor class, and divides the enterprise into 20 gate classes and 97 major classes in total, wherein each major class corresponds to a plurality of middle classes, and each middle class corresponds to a single or a plurality of minor classes.

Since the range of the main classes under each gate class is relatively close and the difference is relatively large, the training set cannot be established by taking the gate class as the range. Similarly, although the ranges of the middle classes under the large classes are very close, since the similarity exists between the large classes, the business of the same enterprise often covers multiple middle classes under multiple large classes at the same time, and it is not suitable to use the large classes as the ranges to establish the training set.

Therefore, in the first step, the large categories of the industry are classified by taking similarity as a standard according to the items in the national economy industry classification; for example, the 13-agroindustrial food processing industry, 14-food manufacturing industry, 15-wine, beverage and fine tea manufacturing industry and 16-tobacco product industry are collectively classified as a food group; the 17-textile industry, the 18-textile garment industry, the clothing industry and the 19-leather, fur, feather and products thereof and the shoe industry are uniformly classified into clothing groups; according to the standard, 97 large classes under all 20 gate classes are grouped, enterprises with similar related ranges are classified into one group, and other grouping standard groups can be established according to requirements so as to establish a training set later;

then, according to the divided groups, enterprises are simply screened firstly, public data can be easily obtained from enterprises marked by a large class or a middle class, 2000 pieces of enterprise data are collected in each group, after the data are collected, the enterprises are manually indexed in the group by taking the middle class or the small class as a standard, the indexed variable is the middle class or the small class of the enterprise to be marked in the group, and the marking range can be set according to requirements; in addition, an error index is additionally added for marking an abnormal observed value; the observed value is the operation range or other related information of each enterprise, if the operation range or information of the enterprise can judge that the enterprise accords with a certain middle class or subclass, the middle class or the subclass which accords with the operation range or information of the enterprise is marked with 1, and the middle class or the subclass which does not accord with the operation range or information of the enterprise is marked with 0; if no one is matched, marking 1 at the error index and marking 0 under all other middle classes or subclasses; thus constructing a training set, and simultaneously constructing a verification set with the number of observation values being 200 according to the same standard;

the observed value in the training set can be other judgment standards formed by characters, such as corresponding product names, besides the business range and other information of the enterprise;

after the training set and the test set are manufactured, the training set after manual indexing is trained by using the algorithm described in the invention, and the model achieves the expected effect by adjusting parameters, and the difference of the most suitable parameters among different groups is possibly very large, so that the model is debugged according to the difference of the groups, and finally the trained model is stored according to the groups;

and finally, forecasting the enterprise by using the trained model, determining an approximate range by simply screening during forecasting, forecasting by using the correspondingly grouped models, and manually screening if an undeterminable condition occurs after forecasting.

Example 2

Referring to fig. 1 and fig. 2, the present invention provides a technical solution: the utility model provides an enterprise trade second grade trade multi-label classifier based on deep learning algorithm, comprises acquisition module, preprocessing module, management module, model building module, training verification module, input module, display module, wherein:

the acquisition module is used for acquiring enterprise operation range information;

the preprocessing module is used for preprocessing the enterprise operation range information;

the management module is used for manually indexing the enterprise operation range information and making a training set, a verification set and a test set for multi-label classification training;

the model establishing module is used for establishing an Albert + TextCNN model by using a training set;

the training verification module is used for training the established Albert + TextCNN model and verifying the accuracy;

the input module is used for inputting the enterprise information to be predicted into the trained model;

the display module is used for displaying the multi-label classification result of the enterprise industry.

The manual indexing method is as follows:

the method comprises the following steps: collecting enterprise operation range information;

step two: extracting the collected enterprise operation range information;

step three: and classifying and indexing the extracted enterprise operation range information.

In this embodiment, preferably, the system further includes a storage module, and the storage module is configured to store the enterprise operation range information.

In this embodiment, preferably, the enterprise operation range information is preprocessed, and the preprocessing includes data information noise removal, data information cleaning, and feature extraction.

In this embodiment, preferably, the multi-label classification system further includes a visual display module, and the visual display module is configured to visually display the multi-label classification result.

In this embodiment, preferably, during prediction, an approximate range is determined by simply screening, then prediction is performed by using a model of a corresponding group, and manual screening is performed if an indeterminable condition occurs after prediction.

In this embodiment, preferably, the multi-label classification method is as follows:

the method comprises the following steps: collecting the operation range information of the enterprise, and preprocessing the operation range information of the enterprise;

step two: carrying out manual indexing on the enterprise operation range information, and establishing a training set by adopting a manual indexing mode;

step three: setting parameters needed to be used by the Albert + TextCNN model;

step four: converting the operation range information and the manually indexed label into word vectors according to the application rule of Albert, and sending the word vectors into an Albert model for learning;

step five: the results of Albert and the initially transformed word vectors are sent into a TextCNN model for training;

step six: constructing a Full connection layer (Full connection layer) by the result output by the TextCNN, the result weight (output weight) and the result bias value (output bias) together, and storing the model;

step seven: judging whether debugging is needed or not according to the accuracy of the trained model, if so, readjusting the parameters needed to be set in the step three, and repeating the steps three to six until the final accuracy of the model is satisfied;

step eight: and carrying out multi-label classification on the industry of any enterprise by using the trained model.

The national economic industry classification GB/T4754-2017 standard divides an enterprise into four levels of a gate class, a major class, a middle class and a minor class, and divides the enterprise into 20 gate classes and 97 major classes in total, wherein each major class corresponds to a plurality of middle classes, and each middle class corresponds to a single or a plurality of minor classes.

Since the range of the main classes under each gate class is relatively close and the difference is relatively large, the training set cannot be established by taking the gate class as the range. Similarly, although the ranges of the middle classes under the large classes are very close, since the similarity exists between the large classes, the business of the same enterprise often covers multiple middle classes under multiple large classes at the same time, and it is not suitable to use the large classes as the ranges to establish the training set.

Therefore, in the first step, the large categories of the industry are classified by taking similarity as a standard according to the items in the national economy industry classification; for example, the 13-agroindustrial food processing industry, 14-food manufacturing industry, 15-wine, beverage and fine tea manufacturing industry and 16-tobacco product industry are collectively classified as a food group; the 17-textile industry, the 18-textile garment industry, the clothing industry and the 19-leather, fur, feather and products thereof and the shoe industry are uniformly classified into clothing groups; according to the standard, 97 large classes under all 20 gate classes are grouped, enterprises with similar related ranges are classified into one group, and other grouping standard groups can be established according to requirements so as to establish a training set later;

then, according to the divided groups, enterprises are simply screened firstly, public data can be easily obtained from enterprises marked by a large class or a middle class, 2000 pieces of enterprise data are collected in each group, after the data are collected, the enterprises are manually indexed in the group by taking the middle class or the small class as a standard, the indexed variable is the middle class or the small class of the enterprise to be marked in the group, and the marking range can be set according to requirements; in addition, an error index is additionally added for marking an abnormal observed value; the observed value is the operation range or other related information of each enterprise, if the operation range or information of the enterprise can judge that the enterprise accords with a certain middle class or subclass, the middle class or the subclass which accords with the operation range or information of the enterprise is marked with 1, and the middle class or the subclass which does not accord with the operation range or information of the enterprise is marked with 0; if no one is matched, marking 1 at the error index and marking 0 under all other middle classes or subclasses; thus constructing a training set, and simultaneously constructing a verification set with the number of observation values being 200 according to the same standard;

the observed value in the training set can be other judgment standards formed by characters, such as corresponding product names, besides the business range and other information of the enterprise;

after the training set and the test set are manufactured, the training set after manual indexing is trained by using the algorithm described in the invention, and the model achieves the expected effect by adjusting parameters, and the difference of the most suitable parameters among different groups is possibly very large, so that the model is debugged according to the difference of the groups, and finally the trained model is stored according to the groups;

and finally, forecasting the enterprise by using the trained model, determining an approximate range by simply screening during forecasting, forecasting by using the correspondingly grouped models, and manually screening if an undeterminable condition occurs after forecasting.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.