WO2021081837A1

WO2021081837A1 - Model construction method, classification method, apparatus, storage medium and electronic device

Info

Publication number: WO2021081837A1
Application number: PCT/CN2019/114473
Authority: WO
Inventors: 刘园林
Original assignee: 深圳市欢太科技有限公司; Oppo广东移动通信有限公司
Priority date: 2019-10-30
Filing date: 2019-10-30
Publication date: 2021-05-06
Also published as: CN114175017A

Abstract

A model construction method, a classification method, an apparatus, a storage medium and an electronic device. The model construction method comprises: determining a target loss value according to a first-level label prediction matrix, a first-level label reference matrix, a target matrix and a second-level label reference matrix; and once the training of a batch of data to be trained is completed, obtaining one target loss value, and each time one target loss value is obtained, passing the target loss value back to a model to be trained, so as to adjust parameters of the model to be trained until the model to be trained converges.

Description

Model construction method, classification method, device, storage medium and electronic equipment

Technical field

This application belongs to the field of electronic technology, and in particular relates to a model construction method, classification method, device, storage medium, and electronic equipment.

Background technique

With the continuous development of electronic technology, the number of applications (APP) installed in electronic devices such as smart phones or tablet computers is increasing. In the process of users using the above APP, information flow services are often involved. The information flow service refers to the service of pushing information flow data to the above-mentioned electronic devices. Among them, the information flow data is used to form pages such as a home page, a list page, or a content page.

Take information flow data as an example. Before pushing an article to an electronic device, the article needs to be marked with upper-level tags, second-level tags, and third-level tags for the information flow recommendation algorithm to use, so that it can be sent to different electronic devices. Push different articles.

Summary of the invention

The embodiments of the present application provide a model construction method, classification method, device, storage medium, and electronic equipment, which can improve the accuracy of joint prediction of primary tags and secondary tags.

In the first aspect, an embodiment of the present application provides a model construction method, including:

Obtain the data to be trained. The data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, a first-level label reference matrix composed of codes corresponding to a first-level label corresponding to each text, and each text The reference matrix of the secondary label formed by the codes corresponding to the corresponding secondary label;

Acquiring a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label;

Inputting the to-be-trained data and the preset relationship dependency matrix into the to-be-trained model to obtain a primary label prediction matrix and a secondary label prediction matrix;

Determine a target relationship dependence matrix according to the first-level label prediction matrix and the preset relationship dependence matrix;

Determine a target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, and the target matrix is based on the secondary label prediction matrix and the target relationship Dependent matrix determination;

Whenever a batch of training data is trained, a target loss value is obtained. For each target loss value obtained, the target loss value is returned to the model to be trained to adjust the parameters of the model to be trained until the model to be trained converges. Confirm that the model training is over and get the trained model.

In the second aspect, an embodiment of the present application provides a classification method, including:

Obtain the text to be classified;

Input the text to be classified into the trained model to obtain a first-level label probability matrix and a second-level label prediction probability matrix. Each element in the first-level label probability matrix corresponds to a first-level label, and the first-level label Each element in the probability matrix is a real number, each element in the secondary label prediction probability matrix corresponds to a primary and secondary label, and each element in the secondary label prediction probability matrix is a real number;

Determine the primary label corresponding to the text to be classified according to the primary label probability matrix, and the primary label corresponding to the element with the largest value in the primary label probability matrix is the primary label corresponding to the text to be classified;

Perform integerization processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix changes from a real number to an integer to obtain a first-level tag integerization matrix. The value of the element is 0 or 1;

Determine the first relationship dependence matrix according to the first-level label integerization matrix and the preset relationship dependence matrix;

Determining a secondary label probability matrix according to the first relationship dependency matrix and the secondary label prediction probability matrix, where each element in the secondary label probability matrix corresponds to a primary and secondary label;

According to the secondary label probability matrix, the secondary label corresponding to the text to be classified is determined, and the secondary label corresponding to the element with the largest value in the secondary label probability matrix is the secondary label corresponding to the text to be classified.

In the third aspect, an embodiment of the present application provides a model construction device, including:

The first acquisition module is used to acquire data to be trained. The data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, and a word segmentation matrix composed of codes corresponding to first-level tags corresponding to each text. A level-label reference matrix and a second-level label reference matrix composed of codes corresponding to the second-level labels corresponding to each text;

The second acquiring module is configured to acquire a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label;

The first training module is configured to input the data to be trained and the preset relationship dependency matrix into the model to be trained to obtain a primary label prediction matrix and a secondary label prediction matrix;

A first determining module, configured to determine a target relationship dependence matrix according to the first-level label prediction matrix and the preset relationship dependence matrix;

The second determining module is configured to determine a target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, and the target matrix is based on the secondary label The prediction matrix and the target relationship dependency matrix are determined;

The second training module is used to obtain a target loss value after the training of a batch of training data is completed, and for each target loss value obtained, the target loss value is transmitted back to the model to be trained to adjust the parameters of the model to be trained , Until the model to be trained converges, confirm the end of the model training, and get the trained model.

In a fourth aspect, an embodiment of the present application provides a classification device, including:

The third obtaining module is used to obtain the text to be classified;

The prediction module is used to input the text to be classified into the trained model to obtain a first-level label probability matrix and a second-level label prediction probability matrix, each element in the first-level label probability matrix corresponds to a first-level label, Each element in the primary label probability matrix is a real number, each element in the secondary label prediction probability matrix corresponds to a primary and secondary label, and each element in the secondary label prediction probability matrix is a real number;

The third determining module is configured to determine the primary label corresponding to the text to be classified according to the primary label probability matrix, and the primary label corresponding to the element with the largest value in the primary label probability matrix is the to be classified The first level label corresponding to the text;

The rounding module is used to perform integerization processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix is changed from a real number to an integer to obtain the first-level tag integerization matrix. The value of the element in the label integerization matrix is 0 or 1;

A fourth determining module, configured to determine a first relationship dependence matrix according to the first-level label integerization matrix and a preset relationship dependence matrix;

A fifth determining module, configured to determine a secondary label probability matrix according to the first relationship dependency matrix and the secondary label prediction probability matrix, and each element in the secondary label probability matrix corresponds to a primary and secondary label;

The sixth determining module is configured to determine the secondary label corresponding to the text to be classified according to the secondary label probability matrix, and the secondary label corresponding to the element with the largest value in the secondary label probability matrix is the to be classified The secondary label corresponding to the text.

In a fifth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, wherein, when the computer program is executed on a computer, the computer is caused to execute the model construction method or classification method provided in this embodiment .

In a sixth aspect, an embodiment of the present application provides an electronic device, including a memory and a processor, the memory stores a computer program, and the processor invokes the computer program stored in the memory to execute:

In a seventh aspect, an embodiment of the present application provides an electronic device, including a memory and a processor, the memory stores a computer program, and the processor invokes the computer program stored in the memory to execute:

Obtain the text to be classified;

Description of the drawings

In the following, with reference to the accompanying drawings, the technical solutions of the present application and its beneficial effects will be apparent through a detailed description of the specific implementations of the present application.

FIG. 1 is a schematic diagram of the first flow of a model construction method provided by an embodiment of the present application.

FIG. 2 is a schematic diagram of the second flow of the model construction method provided by the embodiment of the present application.

FIG. 3 is a schematic diagram of a preset relationship dependency matrix M0 provided by an embodiment of the present application.

FIG. 4 is a schematic diagram of the primary label reference matrix y1 provided by an embodiment of the present application.

Fig. 5 is a schematic diagram of a secondary label reference matrix y2 provided by an embodiment of the present application.

FIG. 6 is a schematic diagram of the primary label prediction matrix P1 provided by an embodiment of the present application.

FIG. 7 is a schematic diagram of a secondary label prediction matrix P2 provided by an embodiment of the present application.

Fig. 8 is a schematic diagram of a 0-1 integer matrix P1-1 provided by an embodiment of the present application.

FIG. 9 is a schematic diagram of a target relationship dependency matrix M provided by an embodiment of the present application.

Fig. 10 is a schematic diagram of a dictionary provided by an embodiment of the present application.

FIG. 11 is a schematic flowchart of a classification method provided by an embodiment of the present application.

FIG. 12 is a schematic diagram of a scene of a classification method provided by an embodiment of the present application.

Fig. 13 is a schematic structural diagram of a model construction device provided by an embodiment of the present application.

Fig. 14 is a schematic structural diagram of a classification device provided by an embodiment of the present application.

FIG. 15 is a schematic diagram of a first structure of an electronic device provided by an embodiment of the present application.

FIG. 16 is a schematic diagram of a second structure of an electronic device provided by an embodiment of the present application.

Detailed ways

Please refer to the drawings, in which the same component symbols represent the same components, and the principle of the present application is implemented in an appropriate computing environment as an example. The following description is based on the exemplified specific embodiments of the present application, which should not be construed as limiting other specific embodiments of the present application that are not described in detail herein.

Please refer to FIG. 1. FIG. 1 is a schematic diagram of the first process of the model construction method provided by an embodiment of the present application. The process of the model construction method can include:

101. Obtain data to be trained. The data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, a first-level label reference matrix composed of codes corresponding to a first-level label corresponding to each text, and each The secondary label reference matrix composed of the codes corresponding to the secondary labels corresponding to the text.

For example, first, the electronic device may obtain multiple pieces of text, and each piece of text in the multiple pieces of text is marked with a primary label and a secondary label. Then, the electronic device can segment each text and determine the code corresponding to the segmentation corresponding to each text. Finally, the electronic device can form a word segmentation matrix according to the code corresponding to the word segmentation corresponding to each text. Among them, the first dimension (row) of the word segmentation matrix represents each text, and the second dimension (column) represents the code corresponding to the word segmentation corresponding to each text. For example, the i-th row and j-th column of the word segmentation matrix represents the code corresponding to the j-th word segmentation corresponding to the i-th text.

Subsequently, the electronic device can encode the primary label corresponding to each text, and then compose the primary label reference matrix. Among them, the first dimension (row) of the primary label reference matrix represents each text, and the second dimension (column) represents the code corresponding to the primary label corresponding to each text. For example, the i-th row of the primary label reference matrix represents the code corresponding to the primary label corresponding to the i-th text. The electronic device can encode the secondary label corresponding to each text, and then compose the secondary label reference matrix. Among them, the first dimension (row) of the secondary label reference matrix is each text, and the second dimension (column) is the code corresponding to the secondary label corresponding to each text. For example, the i-th row of the secondary label reference matrix represents the code corresponding to the secondary label corresponding to the i-th text.

The word segmentation matrix, the primary label reference matrix and the secondary label reference matrix constitute the data to be trained.

102. Obtain a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label.

For example, multiple primary tags and multiple secondary tags can be obtained from a database or collected by a user in advance, and the hierarchical relationship between each primary tag and each secondary tag can be determined. In other words, determine which secondary labels correspond to each primary label. For example, suppose the first-level label is: TV series, and the second-level labels may be: ancient costume, fantasy, modern, and so on. For another example, suppose the first-level label is: sports, and the second-level labels may be: football, basketball, volleyball, table tennis, and so on. Then, the electronic device can establish a preset relationship dependency matrix according to the hierarchical relationship between the primary label and the secondary label. Wherein, the first dimension (row) of the preset relationship dependence matrix represents each first-level label, and the second dimension (column) represents each second-level label. If the j-th second-level label is included under the i-th first-level label, then the preset relationship dependency matrix is 1 in the i-th row and j-th column, otherwise it is 0.

It should be noted that, in the embodiment of the present application, after the preset relationship dependency matrix is established, when the electronic device needs to use the preset relationship dependency matrix, it can directly obtain the preset relationship dependency matrix for use without the need. Perform the establishment of the preset relationship dependency matrix again before using it. That is, create it once and use it many times.

In the embodiment of the present application, the electronic device can obtain the preset relationship dependency matrix.

It should be noted that when obtaining multiple texts, the user can also obtain multiple texts according to the collected multiple primary tags and multiple secondary tags, and input them into the electronic device, and the electronic device can obtain the multiple texts. . That is to say, among the multiple texts obtained by the electronic device, the primary label corresponding to each text is one of the multiple primary labels collected by the user; the secondary label corresponding to each text It is one of the second-level tags among multiple second-level tags collected by the user.

It is understandable that, in order to improve the accuracy of model prediction, the primary label and secondary label corresponding to the text obtained by the electronic device can fully represent the text. For example, a manual labeling method can be used to mark the upper level label and the second level label for the text. Then, the electronic device can obtain these manually labeled texts.

103. Input the to-be-trained data and the preset relationship dependency matrix into the to-be-trained model to obtain a primary label prediction matrix and a secondary label prediction matrix.

For example, after obtaining the data to be trained and the preset relationship dependency matrix, the electronic device may input the data to be trained and the preset relationship dependency matrix into the model to be trained to obtain the primary label prediction matrix and the secondary label prediction matrix.

Among them, in the first-level label prediction matrix, the number of rows of the matrix is the number of texts included in the data to be trained, and the number of columns of the matrix is the number of first-level labels. For example, if the number of texts is 64 and the number of first-level labels is 10, then the first-level label prediction matrix is a 64*10 matrix. The first dimension (row) of the primary label prediction matrix represents each text, and the second dimension (column) represents the probability that the primary label corresponding to each text predicted by the model to be trained is each primary label among multiple primary labels . For example, the i-th row and j-th column of the primary label prediction matrix indicates the probability that the primary label corresponding to the i-th text is the j-th primary label among multiple primary labels.

In the secondary label prediction matrix, the number of rows of the matrix is the number of texts included in the data to be trained, and the number of columns of the matrix is the number of secondary labels. For example, if the number of texts is 64 and the number of secondary labels is 200, then the secondary label prediction matrix is a 64*200 matrix. The first dimension (row) of the secondary label prediction matrix represents each text, and the second dimension (column) represents the probability that the secondary label corresponding to each text predicted by the model to be trained is each secondary label among multiple secondary labels . For example, the i-th row and j-th column of the secondary label prediction matrix indicates the probability that the secondary label corresponding to the i-th text is the jth secondary label among multiple secondary labels.

In this embodiment of the present application, the model to be trained may first initialize its parameters. Then, the electronic device can input the data to be trained and the preset relationship dependency matrix into the model to be trained, and the output result is obtained through forward propagation of the convolutional layer, down-sampling layer, fully connected layer, etc., that is, the first-level label prediction is obtained Matrix and secondary label prediction matrix.

104. Determine the target relationship dependence matrix according to the first-level label prediction matrix and the preset relationship dependence matrix.

105. Determine the target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, the target matrix being determined according to the secondary label prediction matrix and the target relationship dependency matrix.

For example, after obtaining the first-level label prediction matrix, the electronic device may determine the target relationship dependency matrix according to the first-level label prediction matrix and the preset relationship dependency matrix. After obtaining the target relationship dependence matrix, the electronic device can determine the target matrix according to the target relationship dependence matrix and the secondary label prediction matrix. Subsequently, the electronic device can determine the target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix.

106. Whenever a batch of training data is trained, a target loss value is obtained, and each target loss value is obtained, the target loss value is returned to the model to be trained to adjust the parameters of the model to be trained until the model to be trained Convergence, confirm the end of model training, and get the trained model.

It is understandable that after a target loss value is obtained, the target loss value can be transmitted back to each layer of the model to be trained, so as to adjust the parameters of the model to be trained. Subsequently, the electronic device can continue to obtain the data to be trained and input the adjusted parameters into the model to be trained to continue training the model to be trained. The data to be trained and the data to be trained acquired in the process 101 are two different batches of data, and the acquisition process of the data to be trained can refer to the acquisition process of the process 101. After a target loss value is obtained this time, the target loss value can still be transmitted back to the model to be trained, so as to adjust the parameters again until the model to be trained converges, confirm the end of the model training, and obtain the trained model. Among them, the target loss value gradually approaches a certain value, or fluctuates around a certain value, and when the loss change is less than a small positive number, it can be confirmed that the model to be trained has converged.

It can be understood that, in this embodiment, the target relationship dependence matrix determined by the preset relationship dependence matrix and the primary label prediction matrix can be used to perform enhanced training on the secondary label, and the secondary label can be trained according to the primary label prediction matrix and the primary label benchmark. The target loss value determined by the matrix, the target matrix and the secondary label reference matrix can make the model to be trained reach the overall optimum, thereby improving the accuracy of the joint prediction of the primary label and the secondary label by the trained model.

Please refer to FIG. 2, which is a schematic diagram of the second flow of the model construction method provided by the embodiment of the application. The model building method can include:

201. The electronic device obtains multiple first-level tags.

202. The electronic device obtains multiple secondary tags.

203. The electronic device determines the hierarchical relationship between each primary label and each secondary label.

204. The electronic device establishes a preset relationship dependency matrix according to the hierarchical relationship.

For example, 201, 202, 203, and 204 can be:

The user can collect multiple primary tags and multiple secondary tags in advance. Then, the user can input the collected multiple primary tags and multiple secondary tags into the electronic device, and the electronic device can obtain the multiple primary tags and multiple secondary tags.

Then, the electronic device can determine the hierarchical relationship between the obtained multiple primary labels and secondary labels, that is, which secondary labels are under each primary label. It is understandable that this process can be analyzed and classified by the user. For example, the user can determine which secondary labels are under each primary label, and then mark the secondary labels with the mark of the primary label to which they belong. The user can input these primary labels and secondary labels marked with the primary labels to which they belong to the electronic device, and the electronic device can determine which secondary label belongs to which primary label according to the mark on the secondary label.

After determining the hierarchical relationship between each primary label and each secondary label, the electronic device can establish a preset relationship dependency matrix according to the hierarchical relationship between each primary label and each secondary label. Wherein, the first dimension (row) of the preset relationship dependency matrix represents each primary label in the plurality of primary labels, and the second dimension (column) represents each secondary label in the multiple secondary labels. If the i-th first-level label contains the j-th second-level label, then the preset relationship dependence matrix is 1 in the jth row and jth column, otherwise it is 0.

For example, suppose there are 5 primary labels, namely: L1, L2, L3, L4, L5, and 15 secondary labels, respectively: S1, S2, S3, S4, S5, S6, S7, S8, S9 , S10, S11, S12, S13, S14, S15. Among them, the primary label L1 contains secondary labels S1, S2, S4; the primary label L2 contains secondary labels S3, S5; the primary label L3 contains secondary labels S5, S6, S7; the primary label L4 contains There are secondary labels S9, S11, S12, S15; the primary label L5 includes secondary labels S10, S13, S14. Then, according to 5 primary labels L1, L2, L3, L4, L5 and 15 secondary labels S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14 , S15, the established preset relationship depends on the 0th, 1st, and 3rd columns in the 0th row of the matrix M0 as 1, the others are 0; the 1st row, the 2nd and 4th columns are 1, and the others are 0; the 2nd row The 5th, 6th, and 7th columns are 1, and the others are 0; the 8th, 10th, 11th, and 14th columns of the 3rd row are 1, and the others are 0; the 9th, 12th, and 13th columns of the 4th row are 1, and the others are Is 0; that is, the preset relationship dependency matrix M0 is shown in Figure 3.

205. The electronic device obtains multiple pieces of text, the primary label corresponding to each text, and the secondary label corresponding to each text.

206. The electronic device performs word segmentation processing on each text to obtain word segmentation corresponding to each text.

207. The electronic device determines the code corresponding to the word segmentation corresponding to each text.

208. The electronic device determines the word segmentation matrix according to the code corresponding to the word segmentation corresponding to each text.

209. The electronic device performs one-hot encoding processing on the first-level label corresponding to each text to obtain the code corresponding to the first-level label corresponding to each text.

210. The electronic device determines the primary label reference matrix according to the code corresponding to the primary label corresponding to each text.

211. The electronic device performs one-hot encoding processing on the secondary label corresponding to each text to obtain the code corresponding to the secondary label corresponding to each text.

212. The electronic device determines a reference matrix of the secondary label according to the code corresponding to the secondary label corresponding to each text.

213. The electronic device determines the data to be trained according to the word segmentation matrix, the primary label reference matrix, and the secondary label reference matrix.

For example, 205 to 213 can be:

The user can collect multiple pieces of text marked with the primary label and the secondary label according to the multiple primary tags and multiple secondary tags collected by the user. Among them, the first-level label marked by each text is one of the multiple first-level labels collected by the user; the second-level label marked by each text is one of the multiple second-level labels collected by the user One of the secondary labels. The user can input multiple collected texts marked with primary labels and secondary labels into the electronic device, and the electronic device will obtain multiple texts and the primary labels corresponding to each text, that is, the primary labels and primary labels marked by each text. The secondary label corresponding to each text, that is, the secondary label marked by each text.

In some embodiments, in order to improve the accuracy of the model prediction after training, among the collected multiple pieces of text marked with primary tags and secondary tags, the secondary label marked by each piece of text is subordinate to each piece of text marked. The first level label. It can also be said that the first-level label marked by each text contains the second-level label marked by each text. That is to say, if the second-level label marked by a certain piece of text is not subordinate to the first-level label marked by the text, the text may not be collected when collecting the text.

Subsequently, the electronic device can perform word segmentation processing on each text in the multiple texts to obtain the word segmentation corresponding to each text. For example, the electronic device may use a jieba tokenizer to perform word segmentation processing on each text.

In some embodiments, the electronic device may filter the invalid and infrequently used special characters in each text, and then use a jieba tokenizer to perform word segmentation processing on each text.

After the word segmentation corresponding to each text is obtained, the electronic device can determine the code corresponding to the word segmentation corresponding to each text. Then, the electronic device can determine the word segmentation matrix according to the code corresponding to the word segmentation corresponding to each text. Among them, the first dimension (row) in the word segmentation matrix represents each text in a plurality of texts, and the second dimension (column) represents the code corresponding to the word segmentation corresponding to each text. For example, the j-th column of the i-th row of the word segmentation matrix represents the code corresponding to the j-th word segmentation corresponding to the i-th text.

Then, the electronic device can perform one-hot encoding processing on the primary label corresponding to each text to obtain the code corresponding to the primary label corresponding to each text. After obtaining the code corresponding to the primary label corresponding to each text, the electronic device can determine the primary label reference matrix according to the code corresponding to the primary label corresponding to each text. Among them, the first dimension (row) of the first-level label matrix represents each text in the multiple texts, and the second dimension (column) represents the code corresponding to the first-level label corresponding to each text in the multiple texts. For example, the i-th row of the primary label reference matrix represents the code corresponding to the primary label corresponding to the i-th text.

For example, suppose there are 5 first-level tags, namely L1, L2, L3, L4, and L5. When a piece of text contains L1, the corresponding code of the first-level tag corresponding to the text is: 10000; when a piece of text When the text contains L2, the corresponding code of the first-level label of the text is: 01000; when a text contains L3, the corresponding code of the first-level label of the text is: 00100; when a text contains When there is L4, the corresponding code of the first-level label corresponding to the text is: 00010; when a piece of text contains L5, the corresponding code of the first-level label corresponding to the text is: 00001.

For example, assuming there are 15 texts, the first text corresponding to the first level label is L2, the second text corresponding to the first level label is L3, the third text corresponding to the first level label is L1, and the fourth text corresponds to The first-level label is L2, the first-level label corresponding to the text of Article 5 is L5, the first-level label corresponding to the text of Article 6 is L4, the first-level label corresponding to the text of Article 7 is L1, and the first-level label corresponding to the text of Article 8 is L1. The label is L5, the first-level label corresponding to the text in Article 9 is L1, the first-level label corresponding to the text in Article 10 is L3, the first-level label corresponding to the text in Article 11 is L4, and the first-level label corresponding to the text in Article 12 is L4, the first-level label corresponding to the text of Article 13 is L5, the first-level label corresponding to the text of Article 14 is L3, and the first-level label corresponding to the text of Article 15 is L4.

It can be determined that the corresponding code of the first-level label corresponding to the first text is: 01000, the corresponding code of the first-level label corresponding to the second text is: 00100, and the corresponding code of the first-level label corresponding to the third text is: 10000 , The code corresponding to the first level label corresponding to the text in Article 4 is: 01000, the code corresponding to the first level label corresponding to the text in Article 5 is: 00001, the code corresponding to the first level label corresponding to the text in Article 6 is: 00010, The code corresponding to the first-level label corresponding to the 7 texts is: 10000, the code corresponding to the first-level label corresponding to the 8th text is: 00001, the code corresponding to the first-level label corresponding to the 9th text is: 10000, Article 10 The code corresponding to the first-level label corresponding to the text is: 00100, the corresponding code corresponding to the first-level label corresponding to the text of Article 11 is: 00010, the corresponding code corresponding to the first-level label corresponding to the text of Article 12 is: 00010, and the text corresponding to Article 13 is corresponding to The corresponding code of the first-level label is: 00001, the corresponding code of the first-level label corresponding to the text of Article 14 is: 00100, and the corresponding code of the first-level label corresponding to the text of Article 15 is: 00010. Then, the first-level label reference matrix y1 composed of the codes corresponding to the first-level labels corresponding to each of the above 15 texts is shown in FIG. 4.

For example, the electronic device may perform one-hot encoding processing on the secondary label corresponding to each text to obtain the code corresponding to the secondary label corresponding to each text. After obtaining the code corresponding to the secondary label corresponding to each text, the electronic device can determine the secondary label reference matrix according to the code corresponding to the secondary label corresponding to each text. Among them, the first dimension (row) of the secondary label reference matrix is each text in the multiple texts, and the second dimension (column) is the code corresponding to the secondary label corresponding to each text in the multiple texts. For example, the i-th row of the secondary label reference matrix represents the code corresponding to the secondary label corresponding to the i-th text.

For example, suppose there are 15 secondary labels, namely S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, when a piece of text contains S1 When, the corresponding code of the secondary label corresponding to the text is: 100000000000000; when a piece of text contains S2, the corresponding code of the secondary label corresponding to the text is: 010000000000000; when a piece of text contains S3, The code corresponding to the secondary label of the text is: 001000000000000; when a piece of text contains S4, the code corresponding to the secondary label of the text is: 000100000000000; when a piece of text contains S5, the text The corresponding code of the corresponding secondary label is: 000010000000000; when a piece of text contains S6, the corresponding code of the secondary label of the text is: 000001000000000; when a piece of text contains S7, the text corresponds to The code corresponding to the secondary label is: 000000100000000; when a piece of text contains S8, the code corresponding to the secondary label of the text is: 000000010000000; when a piece of text contains S9, the corresponding secondary label of the text The code corresponding to the label is: 000000001000000; when a piece of text contains S10, the corresponding code of the secondary label corresponding to the text is: 000000000100000; when a piece of text contains S11, the secondary label corresponding to the text corresponds to The code of is: 000000000010000; when a piece of text contains S12, the code corresponding to the secondary label of the text is: 000000000001000; when a piece of text contains S13, the code corresponding to the secondary label of the text It is: 000000000000100; when a piece of text contains S14, the corresponding code of the secondary label of the text is: 000000000000010; when a piece of text contains S15, the corresponding code of the secondary label of the text is: 000000000000001.

For example, suppose there are 15 texts, the first text corresponding to the secondary label is S3, the second text corresponding to the secondary label is S7, the third text corresponding to the secondary label is S4, and the fourth text corresponds to The second-level label is S5, the second-level label corresponding to the text in Article 5 is S10, the second-level label corresponding to the text in Article 6 is S12, the second-level label corresponding to the text in Article 7 is S1, and the second-level label corresponding to the text in Article 8 is S1. The label is S14, the second-level label corresponding to the text in Article 9 is S2, the second-level label corresponding to the text in Article 10 is S6, the second-level label corresponding to the text in Article 11 is S11, and the second-level label corresponding to the text in Article 12 is S15, the second-level label corresponding to the text of Article 13 is S13, the second-level label corresponding to the text of Article 14 is S8, and the second-level label corresponding to the text of Article 15 is S9.

It can be determined that the corresponding code of the secondary label corresponding to the first text is: 001000000000000, the corresponding code of the secondary label corresponding to the second text is: 000000100000000, and the corresponding code of the secondary label corresponding to the third text is: 000100000000000 , The code corresponding to the secondary label corresponding to the text in Article 4 is: 000010000000000, the code corresponding to the secondary label corresponding to the text in Article 5 is: 000000000100000, and the code corresponding to the secondary label corresponding to the text in Article 6 is: 000000000001000. The code corresponding to the secondary label corresponding to the 7 texts is: 100000000000000, the code corresponding to the secondary label corresponding to the 8th text is: 000000000000010, and the code corresponding to the secondary label corresponding to the 9th text is: 010000000000000, Article 10. The code corresponding to the second-level label of the text is: 000001000000000, the corresponding code of the second-level label corresponding to the text of Article 11 is: 000000000010000, the code corresponding to the second-level label corresponding to the text of Article 12 is: 000000000000001, and the corresponding code of the text of Article 13 is: The code corresponding to the second-level label is: 000000000000100, the code corresponding to the second-level label corresponding to the text of Article 14 is: 000000010000000, and the code corresponding to the second-level label corresponding to the text of Article 15 is: 000000001000000. Then, the secondary label reference matrix y2 formed according to the code corresponding to the secondary label corresponding to each text in the above 15 texts is shown in FIG. 5.

Finally, the electronic device can determine the data to be trained according to the word segmentation matrix, the primary label reference matrix y1, and the secondary label reference matrix y2.

For example, the form of the data to be trained in the input model can be (x, y), then, in this embodiment of the application, the word segmentation matrix can be used as the x input model, the primary label reference matrix y1 and the secondary label reference matrix y2 Can be input into the model as y.

214. The electronic device obtains a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label.

For example, the electronic device can obtain the preset relationship dependency matrix M0 as shown in FIG. 3.

215. The electronic device inputs the data to be trained and the preset relationship dependency matrix into the model to be trained to obtain a primary label prediction matrix and a secondary label prediction matrix, each element in the primary label prediction matrix is a real number.

For example, the electronic device may input the data to be trained (x, y) and the preset relationship dependency matrix M0 into the model to be trained for training, so that the primary label prediction matrix and the secondary label prediction matrix can be obtained. Among them, x is the word segmentation matrix, and y is the primary label reference matrix y1 and the secondary label reference matrix y2.

It is understandable that the network structure of the model to be trained can adopt CNN, or any one of RNN, such as LSTM, GRU, Bi-LSTM, and so on.

The elements of the primary label prediction matrix and the secondary label prediction matrix obtained after the training of the model to be trained are all real numbers. Among them, the first dimension (row) of the first-level label prediction matrix represents each text in multiple texts, and the second dimension (column) represents that the first-level label corresponding to each text in the multiple texts predicted by the model to be trained is The probability of each first-level tag among multiple first-level tags. For example, the i-th row and the j-th column indicate the probability that the first-level label corresponding to the i-th text is the j-th first-level label. For example, the first-level label prediction matrix P1 may be as shown in FIG. 6.

In Figure 6, row 0 and column 0 indicate the probability that the first-level label corresponding to the first text is L1, and the first row and first column indicate the probability that the first-level label corresponding to the second text is L2. Row 2 and column 2 indicate the probability that the primary label corresponding to the third text is L3, row 3 and column 3, indicate the probability that the primary label corresponding to the fourth text is L4, row 4 and column 4, Indicates the probability that the first level label corresponding to the fifth text is L5...the 10th row, column 0, represents the probability that the first level label corresponding to the 11th text is L1, the 11th line, the first column, represents the 12th text The corresponding first-level label is the probability of L2, the 12th row and the second column, indicate the probability that the first-level label corresponding to the 13th text is L3, the 13th row and the third column, the first-level label corresponding to the 14th text is The probability of L4, the 14th row and the 4th column, indicates the probability that the first-level label corresponding to the 15th text is L5.

The first dimension (row) of the second-level label prediction matrix represents each text in multiple texts, and the second dimension (column) represents that the second-level label corresponding to each text in the multiple texts predicted by the model to be trained is multiple The probability of each second-level tag in the second-level tag. For example, the i-th row and the j-th column indicate the probability that the secondary label corresponding to the i-th text is the j-th secondary label. For example, the secondary label prediction matrix P2 may be as shown in FIG. 7.

In this figure 7, row 0 and column 0 indicate the probability that the second-level label corresponding to the first text is S1, and the first row and first column indicate the probability that the second-level label corresponding to the second text is S2. Row 2 and column 2 indicate the probability that the secondary label corresponding to the third text is S3, row 3 and column 3, indicate the probability that the secondary label corresponding to the fourth text is S4, row 4, column 4, Indicates the probability that the second-level label corresponding to the fifth text is S5, the fifth row and fifth column indicates the probability that the second-level label corresponding to the sixth text is S6, and the sixth row and sixth column indicate that the seventh text corresponds to The second-level label of is the probability of S7. The seventh row and seventh column indicates the probability that the second-level label corresponding to the eighth text is S8, and the eighth row and eighth column indicates that the second-level label corresponding to the ninth text is S9 The probability of, the 9th row and 9th column, indicates the probability that the secondary label corresponding to the 10th text is S10, the 10th row and 10th column, the probability that the secondary label corresponding to the 11th text is S11, the 11th row The 11th column indicates the probability that the secondary label corresponding to the 12th text is S12, the 12th row and the 12th column indicates the probability that the secondary label corresponding to the 13th text is S13, and the 13th column and 13th column indicate the probability of S13. The probability that the secondary label corresponding to the 14th text is S14, the 14th row and the 14th column, indicates the probability that the secondary label corresponding to the 15th text is S15, and so on.

216. The electronic device performs integerization processing on the first-level label prediction matrix, so that each element in the first-level label prediction matrix is changed from a real number to an integer to obtain a first-level label integer matrix. The value is 0 or 1.

For example, the electronic device can perform integerization processing on the first-level label prediction matrix, so that each element in the first-level label prediction matrix is changed from a real number to an integer, to obtain a first-level label integer matrix, and each element in the first-level label integer matrix The value is 0 or 1.

For example, the electronic device may perform integerization processing on the first-level label prediction matrix P1, and convert the first-level label prediction matrix P1 into a 0-1 integer matrix P1-1. In the 0-1 integer matrix P1-1, only the largest The probability is 1 and the rest is 0. That is, the 0-1 integer matrix P1-1 is as shown in FIG. 8.

217. The electronic device cross-multiplies the first-level label integer matrix with the preset relationship dependence matrix to obtain the target relationship dependence matrix.

Among them, the matrix A cross multiplies the matrix B to calculate the product of the matrix A and the matrix B.

For example, the electronic device may calculate the product of the first-level label integer matrix and the preset relationship dependence matrix, and determine the product as the target relationship dependence matrix. The target relationship dependence matrix M = 0-1 integer matrix P1-1 × M0. For example, the target relationship dependency matrix M is shown in FIG. 9.

218. The electronic device determines the first loss value according to the primary label prediction matrix and the primary label reference matrix.

For example, the electronic device can use the first-level label prediction matrix P1 and the first-level label reference matrix y1 as parameters, and input them into the specified loss function, so as to calculate the difference between the first-level label prediction matrix P1 and the first-level label reference matrix y1. The loss value between time, the loss value is the first loss value.

For example, the specified loss function may be: H(P1, y1)=-Σ _n y1(n)·logP1(n). Among them, n represents the nth column element of the matrix.

In the embodiment of the present application, the loss function is usually used to estimate the degree of inconsistency between the predicted value of the model (such as the first-level label prediction matrix P1) and the true value (such as the first-level label reference matrix). In general, the smaller the loss function, the better the robustness of the model. The loss function can be set according to actual needs, and this application does not limit it.

219. The electronic device determines the target matrix according to the secondary label prediction matrix and the target relationship dependency matrix.

220. The electronic device determines the second loss value according to the target matrix and the secondary label reference matrix.

For example, the electronic device may determine the target matrix P3 according to the secondary label prediction matrix P2 and the target relationship dependency matrix M.

The electronic device can take the target matrix P3 and the secondary label reference matrix y2 as parameters and input them into the specified loss function, so that the loss value between the target matrix P3 and the secondary label reference matrix y2 can be calculated. That is the second loss value.

For example, the specified loss function may be: H(P3, y2)=-Σ _n y2(n)·logP3(n). Among them, n represents the nth column element of the matrix.

In order to avoid log(0) when taking the logarithm, the electronic device may add the target relationship dependence matrix M to a positive number e approaching 0 to obtain the first matrix M1. Then, the second-level label prediction matrix P2 is dot-multiplied by the first matrix M1 to obtain the target matrix P3. Among them, matrix A is multiplied by matrix B to calculate the Hadamard product of matrix A and matrix B.

221. The electronic device determines a target loss value according to the first loss value and the second loss value.

For example, after obtaining the first loss value and the second loss value, the electronic device can determine the target loss value according to the first loss value and the second loss value. The target loss value is the loss value corresponding to the model to be trained, and the target loss value is used to characterize whether the model to be trained is optimal.

222. Whenever a batch of training data is trained, a target loss value is obtained, and each target loss value is obtained, the target loss value is returned to the model to be trained to adjust the parameters of the model to be trained until the model to be trained Convergence, confirm the end of model training, and get the trained model.

It is understandable that after a target loss value is obtained, the target loss value can be transmitted back to each layer of the model to be trained, so as to adjust the parameters of the model to be trained. Subsequently, the electronic device can continue to obtain another batch of texts marked with a primary label and a secondary label to obtain another to-be-trained data, and input it into the to-be-trained model after adjusting the parameters to continue the training of the to-be-trained model. Wherein, the other data to be trained and the data to be trained obtained in the process 213 are two different batches of data, and the determination process of the other data to be trained can refer to the process 205 to the process 213. When a target loss value is obtained this time, the target loss value can still be transmitted back to the layers of the model to be trained, so as to adjust the parameters again until the model to be trained converges, confirm the end of the model training, and obtain the trained model . Among them, the target loss value gradually approaches a certain value, or fluctuates around a certain value, and when the loss change is less than a small positive number, the convergence of the model to be trained can be confirmed.

In other embodiments, after each batch of data to be trained is input and the model to be trained is trained, a trained model can be obtained, and the electronic device can obtain a batch of verification data from the verification set and input it into the trained model to verify The accuracy of the trained model. When the accuracy rate obtained this time is greater than the accuracy rate obtained last time, the electronic device may save the trained model this time. When the accuracy rate obtained this time is less than the accuracy rate obtained last time, the electronic device may not save the trained model this time. When the accuracy of the trained model obtained multiple times does not increase, for example, when the accuracy of the trained model obtained multiple times is 87%, 86.9%, 86.7%, and 86.8%, the electronic device can confirm The model training is over.

In some embodiments, the process 221 may include:

The electronic device multiplies the first loss value by the first weight value to obtain the third loss value;

The electronic device multiplies the second loss value by the second weight value to obtain a fourth loss value, where the second weight value is less than the first weight value;

The electronic device determines the target loss value based on the third loss value and the fourth loss value.

In order to improve the accuracy of the model predicting the first-level label, after obtaining the first loss value and the second loss value, the electronic device can multiply the first loss value by a larger weight value to obtain the third loss value; The second loss value is multiplied by a smaller weight value to obtain the fourth loss value. Then, the electronic device can add the third loss value and the fourth loss value to obtain the target loss value. The first weight value and the second weight value can be set according to actual conditions. For example, the first weight value can be 0.6, and the second weight value can be 0.4.

In some embodiments, the process 219 may include:

The electronic device adds the target relationship dependence matrix to the preset value to obtain the first matrix;

The electronic device multiplies the second-level label prediction matrix by the first matrix to obtain the target matrix.

In order to avoid log(0) when taking the logarithm, the electronic device may add the target relationship dependence matrix to a preset value, for example, to a positive number e approaching 0 to obtain the first matrix M1. Then, the second-level label prediction matrix P2 is dot-multiplied by the first matrix M1 to obtain the target matrix P3. Among them, matrix A is multiplied by matrix B to calculate the Hadamard product of matrix A and matrix B.

In some embodiments, the process 207 may include:

The electronic device constructs a dictionary based on the word segmentation corresponding to each text, and the dictionary includes multiple word segmentation and their corresponding codes;

The electronic device determines the code corresponding to the word segmentation corresponding to each text according to the word segmentation and dictionary corresponding to each text.

For example, the electronic device can sort the word segmentation corresponding to multiple pieces of text, and sort out the different word segmentation in the word segmentation corresponding to the multiple pieces of text. Then, the electronic device can encode these different word segmentation, and construct a dictionary according to the different word segmentation and the codes corresponding to the different word segmentation respectively.

For example, suppose the word segmentation corresponding to a certain text is: we, 的, the motherland, is, China. The corresponding participles of the other text are: we, the, the motherland, is, and South Korea. Then, a dictionary constructed based on the word segmentation corresponding to the text and the word segmentation corresponding to another text can be as shown in FIG. 10.

Then, the electronic device can determine the code corresponding to the word segmentation corresponding to each text according to the word segmentation and dictionary corresponding to each text.

For example, if a text is: our home country is China, then the corresponding code of the word segmentation corresponding to the text is: 1, 2, 3, 4, 5.

Please refer to FIG. 11, which is a schematic flowchart of a classification method provided by an embodiment of the present application. The process of the classification method can include:

301. Obtain the text to be classified.

For example, the electronic device can obtain a text that needs to be marked with the upper level label and the second level label, that is, the text to be classified.

302. Input the text to be classified into the trained model to obtain a first-level label probability matrix and a second-level label prediction probability matrix. Each element in the first-level label probability matrix corresponds to a first-level label, and the first-level label probability matrix Each element in the secondary label prediction probability matrix is a real number, each element in the secondary label prediction probability matrix corresponds to a secondary label, and each element in the secondary label prediction probability matrix is a real number.

303. Determine the primary label corresponding to the text to be classified according to the primary label probability matrix, and the primary label corresponding to the element with the largest value in the primary label probability matrix is the primary label corresponding to the text to be classified.

After obtaining the text to be classified, the electronic device can input the text to be classified into the trained model to obtain the first-level label probability matrix. Wherein, each element in the first-level tag probability matrix corresponds to a first-level tag. The value of each element in the first-level label probability matrix represents the probability that the first-level label corresponding to the text to be classified is each of the multiple first-level labels. For example, as shown in FIG. 12, suppose that multiple first-level labels are respectively: L1, L2, L3, L4, L5, and the first-level label probability matrix may be P4. In the first-level label probability matrix P4, column 0 represents the probability that the first-level label corresponding to the text to be classified is L1, and the first column represents the probability that the first-level label corresponding to the text to be classified is L2...The fourth column represents The probability that the first-level label corresponding to the text to be classified is L5. It can be seen that in the first-level label probability matrix P4, the probability of the 0th column is the largest. Therefore, the first-level label corresponding to the text to be classified is L1.

In the embodiment of the present application, the trained model may be generated using the model construction method described in the foregoing embodiment. For the specific generation process, reference may be made to the relevant description of the foregoing embodiment, which will not be repeated here.

304. Perform integerization processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix is changed from a real number to an integer to obtain the first-level tag integerization matrix, and the value of the element in the first-level tag integerization matrix is obtained It is 0 or 1.

For example, after obtaining the first-level tag probability matrix, the electronic device may perform integerization processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix is changed from a real number to an integer to obtain the first-level tag integerization matrix. Wherein, the value of the element in the first-level label integerization matrix is 0 or 1.

For example, as shown in FIG. 12, the electronic device can perform integerization processing on the first-level tag probability matrix P4, and convert the first-level tag probability matrix P4 into a 0-1 integer matrix P5. In the 0-1 integer matrix P5, Only the maximum probability is 1 and the rest are 0.

305. Determine the first relationship dependence matrix according to the first-level label integerization matrix and the preset relationship dependence matrix.

For example, as shown in FIG. 12, the electronic device may cross-multiply the first-level label integerization matrix P5 by the preset relationship dependency matrix M0 shown in FIG. 3 to obtain the first relationship dependency matrix M2. Among them, matrix A cross multiplies matrix B to calculate the product of matrix A and matrix B.

306. Determine a secondary label probability matrix according to the first relationship dependency matrix and the secondary label prediction probability matrix, where each element in the secondary label probability matrix corresponds to a primary and secondary label.

307. Determine the secondary label corresponding to the text to be classified according to the secondary label probability matrix, and the secondary label corresponding to the element with the largest value in the secondary label probability matrix is the secondary label corresponding to the text to be classified.

For example, after acquiring the text to be classified, the electronic device may input the text to be classified into the trained model to obtain the secondary label prediction probability matrix. Wherein, each element in the second-level label prediction probability matrix corresponds to a first-level and second-level label. The value of each element in the secondary label prediction probability matrix represents the probability that the secondary label corresponding to the text to be classified is each secondary label among multiple secondary labels. For example, as shown in Figure 12, suppose that multiple secondary labels are: S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15. The label prediction probability matrix can be P6. In the second-level label prediction probability matrix P6, column 0 indicates the probability that the second-level label corresponding to the text to be classified is S1, and the first column indicates the probability that the second-level label corresponding to the text to be classified is S2... Column 14 Indicates the probability that the secondary label corresponding to the text to be classified is S15.

Then, the electronic device can multiply the second-level label prediction probability matrix P6 by the first relationship dependency matrix M2 to obtain the second-level label probability matrix P7. It can be seen that in the first relationship dependency matrix M2, only the 0th, 1st, and 3rd columns are 1, and the others are all 0. Therefore, only the 0th, 1st, and There is a non-zero value in the third column, and all others are 0. The value of each element in the secondary label probability matrix P7 represents the probability that the secondary label corresponding to the text to be classified is each secondary label among the multiple secondary labels. Then, it is only necessary to determine the maximum value from column 0, column 1, and column 2 of the secondary label probability matrix P7, and determine the secondary label corresponding to the maximum value as the secondary label corresponding to the text to be trained That's it. Compared with the related technology that needs to determine the maximum value from the values corresponding to all elements in the secondary label probability matrix, and the secondary label corresponding to the maximum value is determined as the secondary label corresponding to the text to be trained, this application The classification method provided by the embodiment has a higher accuracy rate.

As shown in FIG. 12, in the second-level label probability matrix P7, the probability of the element in the first column is the largest, so the second-level label corresponding to the text to be classified is S2.

After determining the primary label and secondary label corresponding to the text to be classified, the electronic device can mark the primary label and secondary label on the text to be classified. Subsequently, the electronic device can classify the text to be classified into corresponding categories. For example, if the first-level label is "TV drama" and the second-level label is "ancient costume", then the electronic device can classify the text to be classified into the ancient costume category under the TV drama category.

It can be understood that, in this embodiment of the present application, the electronic device can execute the process 301 to the process 307 in sequence, so as to mark different texts to be classified with corresponding primary labels and secondary labels. When it is necessary to recommend text to a user, the electronic device can obtain the label corresponding to the user, and then select the corresponding text from the text marked with the primary label and the secondary label according to the corresponding label of the user, and recommend the text to user. Wherein, the tag corresponding to the user can be determined by the electronic device according to the user's preference for browsing articles. For example, if a user frequently browses articles with a first-level label L1 and a second-level label S2, then the corresponding tags for the user are L1 and S2, then when pushing articles to the user, he can push the articles marked with L1 and S2 Article.

It is understandable that if it is necessary to enable the trained model to provide services such as labeling tertiary labels and four-level labels, the electronic device can determine the hierarchical relationship between the tertiary label and the secondary label, and establish the tertiary label according to the hierarchical relationship The relationship dependence matrix with the secondary label can be established by referring to the method for establishing the relationship dependence matrix between the primary label and the secondary label. In the same way, the electronic device can also determine the hierarchical relationship between the fourth-level label and the third-level label, and establish the relationship dependency matrix between the fourth-level label and the third-level label according to the hierarchical relationship. The establishment method can also refer to the first-level label and the second-level label. The establishment method of the relationship dependence matrix. The electronic device can encode the third-level label and the fourth-level label of multiple texts according to the same coding method as the first-level label, thereby forming a three-level label reference matrix and a four-level label reference matrix. Then, the electronic device can also input the relationship dependency matrix between the third-level label and the second-level label, the relationship dependency matrix between the fourth-level label and the third-level label, the third-level label reference matrix, and the fourth-level label reference matrix into the model to be trained, so as to finally Train a model that can provide services for labeling primary, secondary, tertiary, and fourth-level labels.

Please refer to FIG. 13, which is a schematic structural diagram of a model construction device provided by an embodiment of the application. The model construction device may include: a first acquisition module 401, a second acquisition module 402, a first training module 403, a first determination module 404, a second determination module 405, and a second training module 406.

The first acquisition module 401 is configured to acquire data to be trained. The data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, and a word segmentation matrix composed of codes corresponding to first-level tags corresponding to each text. The first-level label reference matrix and the second-level label reference matrix composed of the codes corresponding to the second-level labels corresponding to each text;

The second acquiring module 402 is configured to acquire a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label;

The first training module 403 is configured to input the data to be trained and the preset relationship dependency matrix into the model to be trained to obtain a primary label prediction matrix and a secondary label prediction matrix;

The first determining module 404 is configured to determine a target relationship dependence matrix according to the first-level label prediction matrix and the preset relationship dependence matrix;

The second determining module 405 is configured to determine a target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, the target matrix according to the secondary label The label prediction matrix and the target relationship dependency matrix are determined;

The second training module 406 is used to obtain a target loss value after the training of a batch of training data is completed, and to return the target loss value to the model to be trained to perform the parameters of the model to be trained. Adjust until the model to be trained converges, confirm the end of the model training, and get the trained model.

In some embodiments, the second determining module 405 may be configured to: determine the first loss value according to the first-level label prediction matrix and the first-level label reference matrix; and according to the second-level label prediction matrix and the The target relationship dependence matrix is used to determine the target matrix; the second loss value is determined according to the target matrix and the secondary label reference matrix; the target loss value is determined according to the first loss value and the second loss value.

In some embodiments, the second determining module 405 may be used to: multiply the first loss value by a first weight value to obtain a third loss value; and multiply the second loss value by a second weight value, Obtain a fourth loss value, where the second weight value is less than the first weight value; and determine a target loss value according to the third loss value and the fourth loss value.

In some embodiments, the second determining module 405 may be used to: add the target relationship dependency matrix to a preset value to obtain a first matrix; and multiply the second-level label prediction matrix by the first matrix. , Get the target matrix.

In some embodiments, the first determining module 404 may be used to: perform integer processing on the first-level label prediction matrix, so that each element in the first-level label prediction matrix is changed from a real number to an integer to obtain a The first-level label integer matrix, each element in the first-level label integer matrix has a value of 0 or 1, and the first-level label integer matrix is cross-multiplied by the preset relationship dependency matrix to obtain a target relationship dependency matrix.

In some embodiments, the first obtaining module 401 may be used to: obtain multiple pieces of text, the first-level label corresponding to each text, and the second-level label corresponding to each text; perform word segmentation processing on each text to obtain the word segmentation corresponding to each text Determine the code corresponding to the word segmentation corresponding to each text; determine the word segmentation matrix according to the code corresponding to the word segmentation corresponding to each text; perform one-hot encoding processing on the first-level label corresponding to each text to obtain the first-level corresponding to each text The code corresponding to the label; determine the primary label reference matrix according to the code corresponding to the primary label corresponding to each text; perform one-hot encoding processing on the secondary label corresponding to each text to obtain the corresponding secondary label corresponding to each text Encoding; Determine the secondary label reference matrix according to the encoding corresponding to the secondary label corresponding to each text; Determine the data to be trained according to the word segmentation matrix, the primary label reference matrix and the secondary label reference matrix.

In some embodiments, the first acquisition module 401 may be used to: construct a dictionary according to the word segmentation corresponding to each text, the dictionary including a plurality of word segmentation and their corresponding codes; according to the word segmentation corresponding to each text and The dictionary determines the code corresponding to the word segmentation corresponding to each text.

In some embodiments, the first obtaining module 401 may be used to: obtain multiple primary tags; obtain multiple secondary tags; determine the hierarchical relationship between each primary tag and each secondary tag; according to the hierarchical relationship, Establish a default relationship dependency matrix.

Please refer to FIG. 14, which is a schematic structural diagram of a classification device provided by an embodiment of the application. The classification device may include: a third acquisition module 501, a prediction module 502, a third determination module 503, a rounding module 504, a fourth determination module 505, a fifth determination module 506, and a sixth determination module 507.

The third obtaining module 501 is configured to obtain the text to be classified;

The prediction module 502 is used to input the text to be classified into the trained model to obtain a first-level label probability matrix and a second-level label prediction probability matrix, each element in the first-level label probability matrix corresponds to a first-level label Each element in the primary label probability matrix is a real number, each element in the secondary label prediction probability matrix corresponds to a primary and secondary label, and each element in the secondary label prediction probability matrix is a real number ；

The third determining module 503 is configured to determine the primary label corresponding to the text to be classified according to the primary label probability matrix, and the primary label corresponding to the element with the largest value in the primary label probability matrix is the The first-level label corresponding to the classified text;

The rounding module 504 is configured to perform integer processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix is changed from a real number to an integer to obtain the first-level tag integerization matrix. The value of the element in the level label integerization matrix is 0 or 1;

The fourth determining module 505 is configured to determine the first relationship dependence matrix according to the first-level label integerization matrix and the preset relationship dependence matrix;

The fifth determining module 506 is configured to determine a secondary label probability matrix according to the first relationship dependency matrix and the secondary label prediction probability matrix, and each element in the secondary label probability matrix corresponds to a primary and secondary label ；

The sixth determining module 507 is configured to determine the secondary label corresponding to the text to be classified according to the secondary label probability matrix, and the secondary label corresponding to the element with the largest value in the secondary label probability matrix is the text to be classified The corresponding secondary label.

The embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed on a computer, the computer is caused to execute the model construction method or the classification method provided in this embodiment Process.

An embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory. The processor is configured to execute the computer program stored in the memory by calling the computer program stored in the memory. The model building method or the process in the classification method.

For example, the above-mentioned electronic device may be a mobile terminal such as a tablet computer or a smart phone. Please refer to FIG. 15. FIG. 15 is a schematic diagram of the first structure of an electronic device provided by an embodiment of this application.

The electronic device 600 may include components such as a memory 601 and a processor 602. Those skilled in the art can understand that the structure of the electronic device shown in FIG. 15 does not constitute a limitation on the electronic device, and may include more or fewer components than shown in the figure, or a combination of certain components, or different component arrangements.

The memory 601 can be used to store application programs and data. The application program stored in the memory 601 contains executable code. Application programs can be composed of various functional modules. The processor 602 executes various functional applications and data processing by running application programs stored in the memory 601.

The processor 602 is the control center of the electronic device. It uses various interfaces and lines to connect various parts of the entire electronic device, and executes the electronic device by running or executing the application program stored in the memory 601 and calling the data stored in the memory 601 The various functions and processing data of the electronic equipment can be used to monitor the electronic equipment as a whole.

In this embodiment, the processor 602 in the electronic device loads the executable code corresponding to the process of one or more application programs into the memory 601 according to the following instructions, and the processor 601 runs and stores the executable code in the memory 601 The application in 601, so as to realize the process:

Obtain the text to be classified;

Please refer to FIG. 16, which is a schematic diagram of a second structure of an electronic device provided by an embodiment of this application.

The electronic device 600 may include components such as a memory 601, a processor 602, an input unit 603, an output unit 604, and a display screen 605.

The memory 601 can be used to store application programs and data. The application program stored in the memory 601 contains executable code. Application programs can be composed of various functional modules. The processor 602 executes various functional applications and data processing by running application programs stored in the storage 601.

The processor 602 is the control center of the electronic device. It uses various interfaces and lines to connect the various parts of the entire electronic device, and executes the electronic device by running or executing the application program stored in the memory 601 and calling the data stored in the memory 601 The various functions and processing data of the electronic equipment can be used to monitor the electronic equipment as a whole.

The input unit 603 can be used to receive inputted numbers, character information or user characteristic information (such as fingerprints), and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

The output unit 604 may be used to display information input by the user or information provided to the user and various graphical user interfaces of the electronic device. These graphical user interfaces may be composed of graphics, text, icons, videos, and any combination thereof. The output unit may include a display panel.

The display screen 605 can be used to display information such as text and pictures.

In this embodiment, the processor 602 in the electronic device loads the executable code corresponding to the process of one or more application programs into the memory 601 according to the following instructions, and the processor 602 runs and stores the executable code in the memory 601 The application in 601, so as to realize the process:

In some implementation manners, the processor 602 executes the determination of a target loss value based on the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, and the target matrix is based on When the secondary label prediction matrix and the target relationship dependency matrix are determined, it may be executed: according to the primary label prediction matrix and the primary label reference matrix, the first loss value is determined; according to the secondary label prediction Determine the target matrix based on the matrix and the target relationship dependence matrix; determine the second loss value according to the target matrix and the secondary label reference matrix; determine the target according to the first loss value and the second loss value Loss value.

In some implementation manners, when the processor 602 executes the determination of the target loss value according to the first loss value and the second loss value, it may execute: multiply the first loss value by a first weight value, Obtain a third loss value; multiply the second loss value by a second weight value to obtain a fourth loss value, where the second weight value is less than the first weight value; according to the third loss value and the The fourth loss value determines the target loss value.

In some embodiments, the processor 602 executes the prediction matrix based on the secondary label and the target relationship dependency matrix, and when determining the target matrix, it may execute: adding the target relationship dependency matrix to a preset value, Obtain a first matrix; multiply the first matrix by the second-level label prediction matrix to obtain a target matrix.

In some implementation manners, each element in the primary label prediction matrix is a real number, and the processor 602 executes the determination of the target relationship dependence matrix according to the primary label prediction matrix and the preset relationship dependence matrix. , It may be executed: performing integerization processing on the first-level label prediction matrix, so that each element in the first-level label prediction matrix is changed from a real number to an integer to obtain a first-level label integer matrix, the first-level label integer matrix The value of each element in is 0 or 1; the first-level label integer matrix is cross-multiplied by the preset relationship dependence matrix to obtain the target relationship dependence matrix.

In some embodiments, the processor 602 executes the acquisition of data to be trained, and the data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, and a word segmentation matrix corresponding to a first-level label corresponding to each text. When the primary label reference matrix formed by codes and the secondary label reference matrix formed by the codes corresponding to the secondary labels corresponding to each text, you can execute: obtain multiple texts, the primary labels corresponding to each text, and the corresponding text Second-level tags; perform word segmentation processing on each text to obtain the word segmentation corresponding to each text; determine the code corresponding to the word segmentation corresponding to each text; determine the word segmentation matrix according to the code corresponding to the word segmentation corresponding to each text; correspond to each text One-hot encoding process is performed on the first-level tags of each text to obtain the codes corresponding to the first-level tags corresponding to each text; the first-level tag reference matrix is determined according to the codes corresponding to the first-level tags corresponding to each text; The tags are subjected to one-hot encoding processing to obtain the codes corresponding to the second-level tags of each text; determine the second-level tag reference matrix according to the codes corresponding to the second-level tags of each text; according to the word segmentation matrix and the first-level tag reference The matrix and the secondary label reference matrix determine the data to be trained.

In some implementation manners, when the processor 602 executes the code corresponding to the word segmentation corresponding to each text, it may execute: construct a dictionary according to the word segmentation corresponding to each text, and the dictionary includes a plurality of word segmentation and their corresponding codes. The encoding; according to the word segmentation corresponding to each text and the dictionary, the encoding corresponding to the word segmentation corresponding to each text is determined.

In some embodiments, before the processor 602 executes the acquisition of the data to be trained, it may execute: acquire multiple primary tags; acquire multiple secondary tags; determine the hierarchical relationship between each primary tag and each secondary tag; The hierarchical relationship establishes a preset relationship dependency matrix.

Obtain the text to be classified;

In the foregoing embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in an embodiment, please refer to the detailed description of the model construction method/classification method above, which will not be repeated here.

The model construction method/classification method device provided by the embodiment of the application belongs to the same concept as the model construction method/classification method in the above embodiment, and the model construction method can be run on the model construction method/classification method device For any method provided in the embodiment of the classification method, for the specific implementation process, please refer to the embodiment of the model construction method/classification method, which will not be repeated here.

It should be noted that for the model construction method/classification method described in the embodiment of this application, a person of ordinary skill in the art can understand that all or part of the process of realizing the model construction method/classification method described in the embodiment of this application can be achieved through a computer The computer program can be stored in a computer readable storage medium, such as stored in a memory, and executed by at least one processor. The execution process can include constructing the model as described above. The flow of the method/classification method of the embodiment. Wherein, the storage medium may be a magnetic disk, an optical disc, a read only memory (ROM, Read Only Memory), a random access memory (RAM, Random Access Memory), etc.

For the model construction method/classification method device of the embodiment of the present application, each functional module may be integrated in a processing chip, or each module may exist alone physically, or two or more modules may be integrated in one Module. The above-mentioned integrated modules can be implemented in the form of hardware or software function modules. If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer readable storage medium, such as a read-only memory, a magnetic disk or an optical disk, etc. .

The model construction method, classification method, device, storage medium, and electronic equipment provided in the embodiments of the application are described in detail above. Specific examples are used in this article to illustrate the principles and implementation of the application. The above embodiments The description is only used to help understand the method and core idea of this application; at the same time, for those skilled in the art, according to the idea of this application, there will be changes in the specific implementation and scope of application. In summary , The content of this manual should not be construed as a limitation on this application.

Claims

A model construction method, which includes:

Obtain the data to be trained. The data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, a first-level label reference matrix composed of codes corresponding to a first-level label corresponding to each text, and each text The reference matrix of the secondary label formed by the codes corresponding to the corresponding secondary label;

Acquiring a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label;

Inputting the to-be-trained data and the preset relationship dependency matrix into the to-be-trained model to obtain a primary label prediction matrix and a secondary label prediction matrix;

Determine a target relationship dependence matrix according to the first-level label prediction matrix and the preset relationship dependence matrix;

Determine a target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, and the target matrix is based on the secondary label prediction matrix and the target relationship Dependent matrix determination;

Whenever a batch of training data is trained, a target loss value is obtained. For each target loss value obtained, the target loss value is returned to the model to be trained to adjust the parameters of the model to be trained until the model to be trained converges. Confirm that the model training is over and get the trained model.
The model construction method according to claim 1, wherein the target loss value is determined according to the primary label prediction matrix, the primary label reference matrix, the target matrix and the secondary label reference matrix, and The target matrix is determined according to the secondary label prediction matrix and the target relationship dependency matrix, and includes:

Determine a first loss value according to the primary label prediction matrix and the primary label reference matrix;

Determine a target matrix according to the secondary label prediction matrix and the target relationship dependency matrix;

Determine a second loss value according to the target matrix and the secondary label reference matrix;

According to the first loss value and the second loss value, a target loss value is determined.
The model construction method according to claim 2, wherein the determining the target loss value according to the first loss value and the second loss value comprises:

Multiply the first loss value by the first weight value to obtain a third loss value;

Multiplying the second loss value by a second weight value to obtain a fourth loss value, where the second weight value is smaller than the first weight value;

According to the third loss value and the fourth loss value, a target loss value is determined.
The model construction method according to claim 2, wherein the determining the target matrix according to the secondary label prediction matrix and the target relationship dependency matrix comprises:

Adding the target relationship dependency matrix to a preset value to obtain a first matrix;

Multiply the first matrix by the second-level label prediction matrix to obtain a target matrix.
The model construction method according to claim 1, wherein each element in the first-level label prediction matrix is a real number, and the target relationship is determined according to the first-level label prediction matrix and the preset relationship dependency matrix The dependency matrix includes:

Perform integerization processing on the first-level label prediction matrix, so that each element in the first-level label prediction matrix is changed from a real number to an integer to obtain a first-level label integer matrix, and each element in the first-level label integer matrix The value is 0 or 1;

Multiplying the first-level label integer matrix by the preset relationship dependence matrix to obtain a target relationship dependence matrix.
The model construction method according to claim 1, wherein the acquired data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, and a first level corresponding to each text The primary label reference matrix formed by the codes corresponding to the labels and the secondary label reference matrix formed by the codes corresponding to the secondary labels corresponding to each text include:

Obtain multiple pieces of text, the first-level label corresponding to each text, and the second-level label corresponding to each text;

Perform word segmentation processing on each text to obtain the word segmentation corresponding to each text;

Determine the code corresponding to the word segmentation corresponding to each text;

Determine the word segmentation matrix according to the code corresponding to the word segmentation corresponding to each text;

Performing one-hot encoding processing on the first-level label corresponding to each text to obtain the code corresponding to the first-level label corresponding to each text;

Determine the first-level label reference matrix according to the code corresponding to the first-level label corresponding to each text;

Performing one-hot encoding processing on the secondary label corresponding to each text to obtain the code corresponding to the secondary label corresponding to each text;

Determine the reference matrix of the secondary label according to the code corresponding to the secondary label corresponding to each text;

Determine the data to be trained according to the word segmentation matrix, the primary label reference matrix, and the secondary label reference matrix.
The model construction method according to claim 6, wherein said determining the code corresponding to the word segmentation corresponding to each text comprises:

Construct a dictionary according to the word segmentation corresponding to each text, the dictionary including a plurality of word segmentation and their corresponding codes;

According to the word segmentation corresponding to each text and the dictionary, the code corresponding to the word segmentation corresponding to each text is determined.
The model construction method according to claim 1, wherein before said obtaining the data to be trained, it further comprises:

Get multiple first-level labels;

Obtain multiple secondary labels;

Determine the hierarchical relationship between each primary label and each secondary label;

According to the hierarchical relationship, a preset relationship dependency matrix is established.
A classification method, which includes:

Obtain the text to be classified;

Input the text to be classified into the trained model to obtain a first-level label probability matrix and a second-level label prediction probability matrix. Each element in the first-level label probability matrix corresponds to a first-level label, and the first-level label Each element in the probability matrix is a real number, each element in the secondary label prediction probability matrix corresponds to a primary and secondary label, and each element in the secondary label prediction probability matrix is a real number;

Determine the primary label corresponding to the text to be classified according to the primary label probability matrix, and the primary label corresponding to the element with the largest value in the primary label probability matrix is the primary label corresponding to the text to be classified;

Perform integerization processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix changes from a real number to an integer to obtain a first-level tag integerization matrix. The value of the element is 0 or 1;

Determine the first relationship dependence matrix according to the first-level label integerization matrix and the preset relationship dependence matrix;

Determining a secondary label probability matrix according to the first relationship dependency matrix and the secondary label prediction probability matrix, where each element in the secondary label probability matrix corresponds to a primary and secondary label;

According to the secondary label probability matrix, the secondary label corresponding to the text to be classified is determined, and the secondary label corresponding to the element with the largest value in the secondary label probability matrix is the secondary label corresponding to the text to be classified.
A model building device, which includes:

The first acquisition module is used to acquire data to be trained. The data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, and a word segmentation matrix composed of codes corresponding to first-level tags corresponding to each text. A level-label reference matrix and a second-level label reference matrix composed of codes corresponding to the second-level labels corresponding to each text;

The second acquiring module is configured to acquire a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label;

The first training module is configured to input the data to be trained and the preset relationship dependency matrix into the model to be trained to obtain a primary label prediction matrix and a secondary label prediction matrix;

A first determining module, configured to determine a target relationship dependence matrix according to the first-level label prediction matrix and the preset relationship dependence matrix;

The second determining module is configured to determine a target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, and the target matrix is based on the secondary label The prediction matrix and the target relationship dependency matrix are determined;

The second training module is used to obtain a target loss value after the training of a batch of training data is completed, and for each target loss value obtained, the target loss value is transmitted back to the model to be trained to adjust the parameters of the model to be trained , Until the model to be trained converges, confirm the end of the model training, and get the trained model.
A classification device, which includes:

The third obtaining module is used to obtain the text to be classified;

The prediction module is used to input the text to be classified into the trained model to obtain a first-level label probability matrix and a second-level label prediction probability matrix, each element in the first-level label probability matrix corresponds to a first-level label, Each element in the primary label probability matrix is a real number, each element in the secondary label prediction probability matrix corresponds to a primary and secondary label, and each element in the secondary label prediction probability matrix is a real number;

The third determining module is configured to determine the primary label corresponding to the text to be classified according to the primary label probability matrix, and the primary label corresponding to the element with the largest value in the primary label probability matrix is the to be classified The first level label corresponding to the text;

The rounding module is used to perform integerization processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix is changed from a real number to an integer to obtain the first-level tag integerization matrix. The value of the element in the label integerization matrix is 0 or 1;

A fourth determining module, configured to determine a first relationship dependence matrix according to the first-level label integerization matrix and a preset relationship dependence matrix;

A fifth determining module, configured to determine a secondary label probability matrix according to the first relationship dependency matrix and the secondary label prediction probability matrix, and each element in the secondary label probability matrix corresponds to a primary and secondary label;

The sixth determining module is configured to determine the secondary label corresponding to the text to be classified according to the secondary label probability matrix, and the secondary label corresponding to the element with the largest value in the secondary label probability matrix is the to be classified The secondary label corresponding to the text.
A storage medium, wherein a computer program is stored in the storage medium, and when the computer program is run on a computer, the computer is caused to execute the model construction method or claim of any one of claims 1 to 8 9 the classification method.
An electronic device, wherein the electronic device includes a processor and a memory, and a computer program is stored in the memory, and the processor is configured to execute:

Obtain the data to be trained. The data to be trained includes a word segmentation matrix composed of codes corresponding to word segmentation corresponding to each text in a plurality of texts, a first-level label reference matrix composed of codes corresponding to a first-level label corresponding to each text, and each text The reference matrix of the secondary label formed by the codes corresponding to the corresponding secondary label;

Acquiring a preset relationship dependency matrix, where the preset relationship dependency matrix is used to represent the hierarchical relationship between the primary label and the secondary label;

Inputting the to-be-trained data and the preset relationship dependency matrix into the to-be-trained model to obtain a primary label prediction matrix and a secondary label prediction matrix;

Determine a target relationship dependence matrix according to the first-level label prediction matrix and the preset relationship dependence matrix;

Determine a target loss value according to the primary label prediction matrix, the primary label reference matrix, the target matrix, and the secondary label reference matrix, and the target matrix is based on the secondary label prediction matrix and the target relationship Dependent matrix determination;

Whenever a batch of training data is trained, a target loss value is obtained. For each target loss value obtained, the target loss value is returned to the model to be trained to adjust the parameters of the model to be trained until the model to be trained converges. Confirm that the model training is over and get the trained model.
The electronic device according to claim 13, wherein the processor is configured to execute:

Determine a first loss value according to the primary label prediction matrix and the primary label reference matrix;

Determine a target matrix according to the secondary label prediction matrix and the target relationship dependency matrix;

Determine a second loss value according to the target matrix and the secondary label reference matrix;

According to the first loss value and the second loss value, a target loss value is determined.
The electronic device according to claim 14, wherein the processor is configured to execute:

Multiply the first loss value by the first weight value to obtain a third loss value;

Multiplying the second loss value by a second weight value to obtain a fourth loss value, where the second weight value is smaller than the first weight value;

According to the third loss value and the fourth loss value, a target loss value is determined.
The electronic device according to claim 14, wherein the processor is configured to execute:

Adding the target relationship dependency matrix to a preset value to obtain a first matrix;

Multiply the first matrix by the second-level label prediction matrix to obtain a target matrix.
The electronic device according to claim 13, wherein each element in the first-level label prediction matrix is a real number, and the processor is configured to execute:

Perform integerization processing on the first-level label prediction matrix, so that each element in the first-level label prediction matrix is changed from a real number to an integer to obtain a first-level label integer matrix, and each element in the first-level label integer matrix The value is 0 or 1;

Multiplying the first-level label integer matrix by the preset relationship dependence matrix to obtain a target relationship dependence matrix.
The electronic device according to claim 13, wherein the processor is configured to execute:

Obtain multiple pieces of text, the first-level label corresponding to each text, and the second-level label corresponding to each text;

Perform word segmentation processing on each text to obtain the word segmentation corresponding to each text;

Determine the code corresponding to the word segmentation corresponding to each text;

Determine the word segmentation matrix according to the code corresponding to the word segmentation corresponding to each text;

Performing one-hot encoding processing on the first-level label corresponding to each text to obtain the code corresponding to the first-level label corresponding to each text;

Determine the first-level label reference matrix according to the code corresponding to the first-level label corresponding to each text;

Performing one-hot encoding processing on the secondary label corresponding to each text to obtain the code corresponding to the secondary label corresponding to each text;

Determine the reference matrix of the secondary label according to the code corresponding to the secondary label corresponding to each text;

Determine the data to be trained according to the word segmentation matrix, the primary label reference matrix, and the secondary label reference matrix.
The electronic device according to claim 18, wherein the processor is configured to execute:

Construct a dictionary according to the word segmentation corresponding to each text, the dictionary including a plurality of word segmentation and their corresponding codes;

According to the word segmentation corresponding to each text and the dictionary, the code corresponding to the word segmentation corresponding to each text is determined.
An electronic device, wherein the electronic device includes a processor and a memory, and a computer program is stored in the memory, and the processor is configured to execute:

Obtain the text to be classified;

Input the text to be classified into the trained model to obtain a first-level label probability matrix and a second-level label prediction probability matrix. Each element in the first-level label probability matrix corresponds to a first-level label, and the first-level label Each element in the probability matrix is a real number, each element in the secondary label prediction probability matrix corresponds to a primary and secondary label, and each element in the secondary label prediction probability matrix is a real number;

Determine the primary label corresponding to the text to be classified according to the primary label probability matrix, and the primary label corresponding to the element with the largest value in the primary label probability matrix is the primary label corresponding to the text to be classified;

Perform integerization processing on the first-level tag probability matrix, so that each element in the first-level tag probability matrix changes from a real number to an integer to obtain a first-level tag integerization matrix. The value of the element is 0 or 1;

Determine the first relationship dependence matrix according to the first-level label integerization matrix and the preset relationship dependence matrix;

Determining a secondary label probability matrix according to the first relationship dependency matrix and the secondary label prediction probability matrix, where each element in the secondary label probability matrix corresponds to a primary and secondary label;

According to the secondary label probability matrix, the secondary label corresponding to the text to be classified is determined, and the secondary label corresponding to the element with the largest value in the secondary label probability matrix is the secondary label corresponding to the text to be classified.