Model training method and device, text classification method, electronic device and storage medium

Abstract

The present application discloses a model training method and device, a text classification method, an electronic device and a storage medium. The method comprises: according to annotated text with a label, obtaining a relationship map with the annotated text as the central node, wherein the relationship map comprises neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes comprise text nodes and entity nodes; inputting the text nodes in the relationship map into a text classification model to obtain a predicted pseudo-label and a predicted feature vector for each of the text nodes; according to the predicted pseudo-label, converging each text node with the central node to obtain a converged feature vector; using the converged feature vector as training data, training a graph classification model. The embodiment of the present application realizes an effective expansion of the annotated text, thereby improving the recognition accuracy of the trained model by adding a large amount of training data related to the annotated text.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a model training method and device, a text classification method, an electronic device and a storage medium. Background Art

[0002] With the development of media technology, news is increasingly appearing in people's lives and work, becoming an important part of it. Especially for enterprises, the sharp increase in the amount of news and the increase in the speed of dissemination have made the management of news one of the important tasks of enterprise risk control. For example, the managers of an enterprise need to classify the news that appears and spreads on various media according to the risk categories based on the various tasks carried out by the enterprise, so as to use the corresponding strategies for timely processing. Such artificial humans not only consume huge manpower, but also cannot ensure efficiency, timeliness and accuracy of processing. Therefore, a technical solution that can efficiently classify news texts is needed. Summary of the invention

[0003] The embodiments of the present application provide a model training method and device, a text classification method, an electronic device and a storage medium to solve the defects of low efficiency and low accuracy in classifying news texts in the prior art.

[0004] To achieve the above objectives, the present application provides a model training method, including:

[0005] According to the labeled text, a relationship graph with the labeled text as a central node is obtained, wherein the relationship graph includes neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes include text nodes and entity nodes;

[0006] Inputting the text nodes in the relationship graph into a text classification model to obtain a predicted pseudo label and a predicted feature vector for each of the text nodes;

[0007] According to the predicted pseudo-label, each text node and the central node are aggregated to obtain an aggregated feature vector;

[0008] The aggregated feature vector is used as training data to train a graph classification model.

[0009] The present application also provides a text classification method, including:

[0010] Performing text feature extraction on the target text to obtain a text feature vector of the target text;

[0011] Inputting the text feature vector of the target text into a predetermined graph classification model to calculate the correlation between the target text and each predetermined classification label;

[0012] Determine the classification label of the target text according to the relevance, wherein

[0013] The predetermined graph classification model is trained using the following method:

[0014] According to the annotated text with the classification label, a relationship graph with the annotated text as the central node is obtained, wherein the relationship graph includes neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes include text nodes and entity nodes;

[0015] Inputting the text nodes in the relationship graph into a text classification model to obtain a predicted pseudo label and a predicted feature vector for each of the text nodes;

[0016] According to the predicted pseudo-label, each text node and the central node are aggregated to obtain an aggregated feature vector;

[0017] The graph classification model is trained using the aggregated feature vector as training data.

[0018] The present application also provides a model training device, including:

[0019] An acquisition module, configured to acquire, based on the labeled text, a relationship graph with the labeled text as a central node, wherein the relationship graph includes neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes include text nodes and entity nodes;

[0020] A text classification module, used for inputting the text nodes in the relationship graph into a text classification model to obtain a predicted pseudo label and a predicted feature vector for each of the text nodes;

[0021] A convergence module, used for converging each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector;

[0022] A training module is used to use the aggregated feature vector as training data to train a graph classification model.

[0023] The present application also provides an electronic device, including:

[0024] Memory, used to store programs;

[0025] A processor is used to run the program stored in the memory, and when the program is run, the model training method provided in the embodiment of the present application is executed.

[0026] An embodiment of the present application also provides a computer-readable storage medium on which a computer program executable by a processor is stored, wherein the program, when executed by the processor, implements the model training method provided in the embodiment of the present application.

[0027] The model training method and device, text classification method, electronic device and storage medium provided in the embodiments of the present application obtain a relationship graph with the annotated text as the central node based on the annotated text with labels, input the text nodes in the relationship graph into the text classification model to obtain a predicted pseudo-label and a predicted feature vector for each text node, and then converge each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector, and use the converged feature vector as training data to train the graph classification model. Therefore, the annotated text can be used as the central node in the relationship graph to expand and obtain neighbor nodes including unlabeled text, and then the nodes of the unlabeled text are converged with the central node according to the predicted pseudo-label of each unlabeled text, thereby generating a converged feature vector that is both related to the annotated text and different from the annotated text, so that the converged feature vector can be used together with the label of the annotated text as new training data to train the graph classification model, thereby achieving effective expansion of the annotated text, and thus improving the recognition accuracy of the trained model by adding a large amount of training data related to the annotated text.

[0028] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0030] Figure 1 A schematic diagram of a scenario of a model training solution provided in an embodiment of the present application;

[0031] Figure 2 A flowchart of an embodiment of the model training method provided in this application;

[0032] Figure 3 A flowchart of another embodiment of the model training method provided by the present application;

[0033] Figure 4 A schematic diagram of the structure of an embodiment of the model training device provided in this application;

[0034] Figure 5 A schematic diagram of the structure of an electronic device embodiment provided in this application. DETAILED DESCRIPTION

[0035] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0036] Embodiment 1

[0037] The solution provided in the embodiments of the present application can be applied to any system with text processing capabilities, such as a server system including a chip with text processing functions and related components, etc. Figure 1 A schematic diagram of a scenario of a model training solution provided in an embodiment of the present application, Figure 1 The scenario shown is only one of the scenarios to which the technical solution of the present application can be applied.

[0038] With the development of media technology, news has appeared more and more in people's lives and work, becoming an important part of it. Especially for enterprises, the sharp increase in the amount of news and the increase in the speed of dissemination have made the management of news one of the important tasks of enterprise risk control. For example, the managers of enterprises need to classify the news that appears and spreads on various media according to the risk categories based on the various tasks carried out by the enterprise, so as to use the corresponding strategies for timely processing. Such artificial humans not only consume huge manpower, but also cannot ensure efficiency, timeliness and accuracy of processing.

[0039] To this end, there have been schemes in the prior art that use machine learning models to classify news texts that appear in the media. Specifically, in the prior art, by training the machine learning model for news text classification, the machine learning model can automatically identify the news that appears in the media and calculate the probability that it belongs to each category, thereby determining the category with the highest probability as the category to which the news text belongs. Such a scheme greatly improves the efficiency of news classification because it eliminates the manual processing link. However, the accuracy of such recognition processing by the machine learning model depends on the training data. For example, a large amount of annotated data needs to be prepared for the machine learning model as training data to ensure the accuracy of the trained model in identifying the category of the newly appeared news text. At the current stage, the annotated data usually needs to be manually classified and annotated by specialized personnel for various existing news texts before it can become annotated data. This makes the amount of annotated data at this stage subject to a large restriction, which in turn leads to a low recognition accuracy rate of the classification model generated by training the model using limited annotated data.

[0040] In particular, some solutions have been proposed in the prior art for improving the classification accuracy of models when there is less effective labeled data. For example, regarding the text classification solution of few-shot learning, a solution using meta-learning has been proposed in the prior art. The essence of the meta-learning solution is transfer learning, for example, the model learned from the meta-training set is transferred to a small sample, that is, the meta-test set, to perform classification prediction. The purpose is to further divide the meta-training set and the test set on the basis of the small sample, and at the same time, to obtain better classification results. Therefore, in actual use, the meta-learning solution is difficult to achieve such an ideal goal. Another solution has been proposed in the prior art, namely the semi-supervised learning solution, which can be particularly targeted at situations where the amount of unlabeled data is large and the labeling cost is high. In such an existing solution, the classification result with the highest probability for each unlabeled sample data can be obtained as its pseudo-label by election. At the same time, in order to reduce the impact of noise, in this solution, the corresponding loss can be calculated only when the confidence is greater than the threshold.

[0041] In the case of classification scenarios such as news texts, as mentioned above, there are actually a large number of unlabeled news texts, that is, there are relatively few labeled news texts. Therefore, if the above-mentioned semi-supervised algorithm is used in this case, there are actually relatively few news with designated labels, that is, designated classifications, in the massive amount of news texts. Therefore, if a pre-trained classification model is used to add pseudo-labels to the unlabeled news texts, and then valid data is selected based on a high confidence threshold, not only is the cost high, but the probability of noise being screened out is also increased. In addition, in this case, since the news texts involve different corporate entities, the content of the news descriptions of different corporate entities may vary greatly, and the effectiveness of auxiliary news information with large entity differences may be poor.

[0042] For example, in Figure 1 In the scenario shown in , when training a model for classifying news texts, multiple news texts can be obtained from media such as the Internet, and one or more annotated texts can be obtained by manually annotating these news texts. In an embodiment of the present application, the annotated texts generally include text objects that have been manually classified and classification results obtained by the classification process, such as labels. Therefore, in the prior art, text objects in such annotated texts are generally used to input into the model to be trained, and the annotated classification results or labels are used to compare with the output results of the model to be trained, so that the model can be trained. However, as mentioned above, due to the different styles of different news texts, the labor cost of annotating them is also relatively high, and the high cost of manual annotation also leads to a small number of annotated texts, which in turn makes the effect of training the model poor.

[0043] Therefore, in an embodiment of the present application, a relationship map with the labeled text as the central node can be obtained based on the labeled text. The relationship map may include neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes may include text nodes and entity nodes. Afterwards, the text nodes in the relationship map can be input into a text classification model to obtain a predicted pseudo-label and a predicted feature vector for each text node. According to the predicted pseudo-label, each text node is converged with the central node to obtain a converged feature vector. Finally, the converged feature vector can be used as training data to train the graph classification model.

[0044] Specifically, one or more multi-label classified text data that have been annotated can be obtained in advance, so that this batch of annotated texts can be used as a training set. Figure 1 An example scenario of expanding training data with an annotated text is shown in Figure 1In the scenario shown in , a labeled text N_labeled can be obtained from a labeled data source such as the Internet and used as the central node in the relationship graph. In the embodiment of the present application, the relationship graph can be the enterprise entity corresponding to each news text, and the entity set can be obtained by setting rules or using an entity extraction model to pre-extract the news text. Figure 1 In the example shown in , a training scenario of a corporate news text is shown. The annotated text is a piece of news and has been manually classified / labeled. In addition, the news text can be subjected to entity extraction processing by using a pre-set rule or an entity extraction model as described above to obtain the enterprise corresponding to the news text as the entity subject. After obtaining the enterprise entity corresponding to the news text, other entities related to the entity subject can be further obtained to form an enterprise map.

[0045] For example, an enterprise graph can be constructed based on various relationships involved in an enterprise. For example, other entities related to the entity subject can be obtained based on various relationships such as enterprise-investment-enterprise, enterprise-common executives-enterprise, enterprise-common legal persons-enterprise, etc. For example, based on the relationship of enterprise-investment-enterprise, enterprise entity A obtained by entity extraction processing of news text can be used to obtain enterprise entity B that has an investment relationship with entity A, that is, enterprise entity B has implemented an investment behavior on enterprise entity A, and therefore has an investment relationship with enterprise entity A, so that enterprise entity B can be obtained from enterprise entity A through this relationship, and enterprise entity B can be used as an entity associated with enterprise entity A in the enterprise graph of enterprise entity A. By analogy, enterprise entity C that has a common executive relationship with enterprise entity A, for example, can also be obtained. Therefore, an enterprise graph with multiple enterprise entity nodes is obtained in the above manner. Afterwards, news texts related to these enterprise entities can be further obtained from the text data source based on each enterprise entity node in the enterprise graph.

[0046] In the embodiment of the present application, these news texts may all be news texts that have not been annotated. In other words, in the embodiment of the present application, since the annotation of news texts requires a large amount of human resources, the number of existing annotated texts is small. Therefore, when expanding news texts based on corporate entities in the embodiment of the present application, news texts that have not been annotated can be mainly obtained as extended text nodes of annotated news texts. Of course, annotated texts that have been annotated can also be obtained as extended text points of annotated news text A. The present application has no restrictions on this.

[0047] In the relationship graph formed by the labeled text N_labeled as the central node and the unlabeled text N_unlabeled extended by the entity relationship, each node can include two categories: news nodes and enterprise nodes. And the edges can include two categories: enterprise-enterprise associations and enterprise-news associations. For example, Figure 1 As shown in , the labeled news text N_labeled as the center is a news node, and the corporate entity A associated with it is the corporate node of N_labeled, and the edge between them represents the corporate-news relationship. Corporate entity B, which has an investment relationship with corporate entity A, can be associated with three unlabeled news texts, and these three unlabeled news texts can all be news texts related to corporate entity B. And corporate entity C, which has a common executive relationship with corporate entity A, can be associated with two unlabeled news texts, and these two unlabeled news texts can be news texts related to corporate entity C. In particular, other unlabeled news texts related to corporate entity A can be further obtained as news nodes based on corporate entity A.

[0048] After obtaining the extended unlabeled news text based on the labeled text N_labeled as above, the labeled text N_labeled can be used as input to train the initial multi-label classification model using the text learning model. In an embodiment of the present application, the classification model can be a relatively simple TextCNN model, or a slightly more complex model such as the HAN model, TextRNN, etc., or a more complex model such as BERT or BERT+LSTM. Therefore, such a classification model can be used to train the labeled text N_labeled as training data until convergence. Therefore, a text classification model based on the labeled text N_labeled is obtained, so that the classification model can be used to classify the unlabeled news text nodes in the relationship graph obtained as above, so that the text classification prediction result can be obtained for each unlabeled news node as the pseudo-label P_label of the node, and the text prediction feature vector P_feature can also be obtained.

[0049] Therefore, in the embodiment of the present application, the features of the nodes whose pseudo labels of each unlabeled news node obtained by the text classification model are the same as the labeled labels of the central node N_labeled can be aggregated to the central node. For example, the feature vectors P_feature of the unlabeled news nodes with the same pseudo labels as the labeled labels of the central node N_labeled can be averaged to obtain the average vector Neigh_feature, and then the averaged feature vector Neigh_feature can be concatenated or averaged with the T_feature of the central node to obtain a new central node Next T_feature. Therefore, a new central node related to the central node but with different features can be obtained, so that the new features and the labeled labels can be used together to form new training data to be input into the graph classification model.

[0050] In addition, in the embodiment of the present application, the more similar the neighbor node news is to the central news node N_labeled, the greater the role it should play in the prediction task of the central node, and the node with the greater difference in the central node representation should play as little role as possible in the convergence of the central node. Therefore, on the basis of using the trained text classification model to calculate the pseudo-label P_label of each unlabeled news node in the relationship graph, in the embodiment of the present application, the similarity between the pseudo-label P_label calculated by the above text classification model for each unlabeled news node and the labeled label of the central news node N_labeled can be further calculated. And the similarity between the pseudo-label P_label obtained by the above text classification model for each unlabeled news node and the label of the central node is used as the weight of convergence, and the vectors of the unlabeled nodes with the same pseudo-label as the labeled label of the central node are merged by weighted averaging to obtain Neigh_feature, and then it is spliced ​​or averaged with the current T_feature of the central node to obtain the new feature vector Next T_feature of the central node. Thus, the graph classification model can be trained similarly based on the new feature vector Next T_feature of the central node and the labeled label to form training data.

[0051] In addition, after the graph classification model is trained using the new feature vector as described above, in the embodiment of the present application, the trained model can be used to further perform classification calculations on each unlabeled news node in the relationship graph constructed in the above manner. That is, the text nodes in the relationship graph are input into the trained graph classification model to obtain the predicted pseudo-label and predicted feature vector for each text node. Thus, the predicted pseudo-label and predicted feature vector obtained in this way can be used to update the relationship graph, and then the above calculations are performed again until the graph classification model reaches convergence.

[0052] Therefore, the model training scheme provided in the embodiment of the present application obtains a relationship graph with the annotated text as the central node based on the annotated text with labels, inputs the text nodes in the relationship graph into the text classification model to obtain a predicted pseudo-label and a predicted feature vector for each text node, and then converges each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector, and uses the converged feature vector as training data to train the graph classification model. Therefore, the annotated text can be used as the central node in the relationship graph to expand and obtain neighbor nodes including unlabeled text, and then the nodes of the unlabeled text are converged with the central node according to the predicted pseudo-label of each unlabeled text, thereby generating a converged feature vector that is both related to the annotated text and different from the annotated text, so that the converged feature vector can be used together with the label of the annotated text as new training data to train the graph classification model, thereby achieving effective expansion of the annotated text, and thus improving the recognition accuracy of the trained model by adding a large amount of training data related to the annotated text.

[0053] The above embodiments are illustrations of the technical principles and exemplary application frameworks of the embodiments of the present application. The specific technical solutions of the embodiments of the present application are further described in detail below through multiple embodiments.

[0054] Embodiment 2

[0055] Figure 2 This is a flowchart of an embodiment of the model training method provided by the present application. The execution subject of this method can be various terminals or server devices with text classification capabilities, or it can be a device or chip integrated on these devices. Figure 2 As shown, the model training method includes the following steps:

[0056] S201, based on the annotated text with labels, obtain a relationship graph with the annotated text as the central node.

[0057] In step S201, one or more labeled and categorized text data may be obtained, so that the labeled text data is used as a training set. Figure 1 In the scenario shown in FIG. 2 , in step S201 , a labeled text N_labeled may be obtained from a labeled data source such as the Internet, and the labeled text may have a classification label determined for it through the labeling process. Therefore, the labeled text may serve as a central node in the relationship graph.

[0058] In an embodiment of the present application, the relationship graph may be the business entity corresponding to each text, and a graph consisting of an entity set may be obtained by setting rules or using an entity extraction model to pre-extract the annotated text.

[0059] For example, in Figure 1 In the example shown in , a training scenario of a corporate news text is shown. The annotated text is a piece of news and has been manually classified / labeled. In addition, the news text can be subjected to entity extraction processing by using a pre-set rule or an entity extraction model as described above to obtain the enterprise corresponding to the news text as the entity subject. After obtaining the enterprise entity corresponding to the news text, other entities related to the entity subject can be further obtained to form an enterprise map.

[0060] For example, an enterprise graph can be constructed based on various relationships involved in an enterprise. For example, other entities related to the entity subject can be obtained based on various relationships such as enterprise-investment-enterprise, enterprise-common executives-enterprise, enterprise-common legal persons-enterprise, etc. For example, based on the relationship of enterprise-investment-enterprise, enterprise entity A obtained by entity extraction processing of news text can be used to obtain enterprise entity B that has an investment relationship with entity A, that is, enterprise entity B has implemented an investment behavior on enterprise entity A, and therefore has an investment relationship with enterprise entity A, so that enterprise entity B can be obtained from enterprise entity A through this relationship, and enterprise entity B can be used as an entity associated with enterprise entity A in the enterprise graph of enterprise entity A. By analogy, enterprise entity C that has a common executive relationship with enterprise entity A, for example, can also be obtained. Therefore, an enterprise graph with multiple enterprise entity nodes is obtained in the above manner. Afterwards, news texts related to these enterprise entities can be further obtained from the text data source based on each enterprise entity node in the enterprise graph.

[0061] In the embodiment of the present application, these news texts may all be news texts that have not been annotated. In other words, in the embodiment of the present application, since the annotation of news texts requires a large amount of human resources, the number of existing annotated texts is small. Therefore, when expanding news texts based on corporate entities in the embodiment of the present application, news texts that have not been annotated can be mainly obtained as extended text nodes of annotated news texts. Of course, annotated texts that have been annotated can also be obtained as extended text points of annotated news text A. The present application has no restrictions on this.

[0062] In the relational graph formed by the labeled text N_labeled as the central node and the unlabeled text extended by the entity relationship, each node can include two categories: news nodes and enterprise nodes. And the edges can include two categories: enterprise-enterprise associations and enterprise-news associations. For example, Figure 1 As shown in , the labeled news text N_labeled as the center is a news node, and the corporate entity A associated with it is the corporate node of N_labeled, and the edge between them represents the corporate-news relationship. Corporate entity B, which has an investment relationship with corporate entity A, can be associated with three unlabeled news texts, and these three unlabeled news texts can all be news texts related to corporate entity B. And corporate entity C, which has a common executive relationship with corporate entity A, can be associated with two unlabeled news texts, and these two unlabeled news texts can be news texts related to corporate entity C. In particular, other unlabeled news texts related to corporate entity A can be further obtained as news nodes based on corporate entity A.

[0063] S202, input the text nodes in the relationship graph into the text classification model to obtain a predicted pseudo label and a predicted feature vector for each text node.

[0064] After obtaining the relationship graph including the extended unlabeled news text based on the labeled text N_labeled in step S201, in step S202, the labeled text N_labeled can be first used as input to train an initial multi-label classification model using a text learning model.

[0065] In an embodiment of the present application, the classification model used in step S202 can be a relatively simple TextCNN model, or a slightly more complex model such as a HAN model, TextRNN, etc., or a more complex model such as BERT or BERT+LSTM. Therefore, such a classification model can be used to train the labeled text N_labeled as training data until convergence. Therefore, a text classification model based on the labeled text N_labeled is obtained, so that in step S202, the classification model can be used to classify the unlabeled news text nodes in the relationship graph obtained in step S201, so that the text classification prediction result can be obtained for each unlabeled news node as the pseudo-label P_label of the node, and the text prediction feature vector P_feature can also be obtained.

[0066] S203: according to the predicted pseudo-label, aggregate each text node and the central node to obtain an aggregated feature vector.

[0067] In step S203, the feature vector of each unlabeled text node in the relationship graph obtained in step S201 obtained by the text classification model can be aggregated with the features of the labeled nodes, such as by concatenation or averaging, to obtain an aggregated feature vector.

[0068] For example, in step S203, the features of the nodes with the same pseudo-labels of the unlabeled text obtained in step S202 and the labeled labels of the labeled text N_labeled can be converged to the central node by connection or averaging. For example, in step S203, the pseudo-labels obtained by the text classification model for each text node of the unlabeled text obtained in step S202 can be first compared with the labels of the labeled text in step S201, so that the feature vectors P_feature of the unlabeled text with the same pseudo-labels as the labeled text N_labeled can be averaged to obtain the average vector Neigh_feature, and then the averaged feature vector Neigh_feature can be spliced ​​or averaged with the T_feature of the central node to obtain a new central node Next T_feature.

[0069] S204, using the aggregated feature vector as training data to train the graph classification model.

[0070] Therefore, in step S204, the new central node obtained in step S203 that is related to the annotated text but has different features can be used, and the new features and the annotated labels can be used together to form new training data to be input into the graph classification model.

[0071] Therefore, the model training method provided in the embodiment of the present application obtains a relationship graph with the annotated text as the central node based on the annotated text with labels, inputs the text nodes in the relationship graph into the text classification model to obtain a predicted pseudo-label and a predicted feature vector for each text node, and then converges each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector, and uses the converged feature vector as training data to train the graph classification model. Therefore, the annotated text can be used as the central node in the relationship graph to expand and obtain neighbor nodes including unlabeled text, and then the nodes of the unlabeled text are converged with the central node according to the predicted pseudo-label of each unlabeled text, thereby generating a converged feature vector that is both related to the annotated text and different from the annotated text, so that the converged feature vector can be used together with the label of the annotated text as new training data to train the graph classification model, thereby achieving effective expansion of the annotated text, and thus improving the recognition accuracy of the trained model by adding a large amount of training data related to the annotated text.

[0072] Embodiment 3

[0073] Figure 3 This is a flowchart of another embodiment of the model training method provided by the present application. The execution subject of this method can be various terminals or server devices with text classification capabilities, or it can be a device or chip integrated on these devices. Figure 3 As shown, the model training method provided in the embodiment of the present application may include the following steps:

[0074] S301, based on the annotated text with labels, obtain a relationship graph with the annotated text as the central node.

[0075] In step S301, one or more labeled and categorized text data may be obtained, so that the labeled text data is used as a training set. Figure 1 In the scenario shown in FIG. 3 , in step S301 , a labeled text N_labeled may be obtained from a labeled data source such as the Internet, and the labeled text may have a classification label determined for it through the labeling process. Therefore, the labeled text may serve as a central node in the relationship graph.

[0076] In an embodiment of the present application, the relationship graph may be the business entity corresponding to each text, and a graph consisting of an entity set may be obtained by setting rules or using an entity extraction model to pre-extract the annotated text.

[0077] For example, in Figure 1 In the example shown in , a training scenario of a corporate news text is shown. The annotated text is a piece of news and has been manually classified / labeled. In addition, the news text can be subjected to entity extraction processing by using a pre-set rule or an entity extraction model as described above to obtain the enterprise corresponding to the news text as the entity subject. After obtaining the enterprise entity corresponding to the news text, other entities related to the entity subject can be further obtained to form an enterprise map.

[0078] For example, an enterprise graph can be constructed based on various relationships involved in an enterprise. For example, other entities related to the entity subject can be obtained based on various relationships such as enterprise-investment-enterprise, enterprise-common executives-enterprise, enterprise-common legal persons-enterprise, etc. For example, based on the relationship of enterprise-investment-enterprise, enterprise entity A obtained by entity extraction processing of news text can be used to obtain enterprise entity B that has an investment relationship with entity A, that is, enterprise entity B has implemented an investment behavior on enterprise entity A, and therefore has an investment relationship with enterprise entity A, so that enterprise entity B can be obtained from enterprise entity A through this relationship, and enterprise entity B can be used as an entity associated with enterprise entity A in the enterprise graph of enterprise entity A. By analogy, enterprise entity C that has a common executive relationship with enterprise entity A, for example, can also be obtained. Therefore, an enterprise graph with multiple enterprise entity nodes is obtained in the above manner. Afterwards, news texts related to these enterprise entities can be further obtained from the text data source based on each enterprise entity node in the enterprise graph.

[0079] In the embodiment of the present application, these news texts may all be news texts that have not been annotated. In other words, in the embodiment of the present application, since the annotation of news texts requires a large amount of human resources, the number of existing annotated texts is small. Therefore, when expanding news texts based on corporate entities in the embodiment of the present application, news texts that have not been annotated can be mainly obtained as extended text nodes of annotated news texts. Of course, annotated texts that have been annotated can also be obtained as extended text points of annotated news text A. The present application has no restrictions on this.

[0080] In the relational graph formed by the labeled text N_labeled as the central node and the unlabeled text extended by the entity relationship, each node can include two categories: news nodes and enterprise nodes. And the edges can include two categories: enterprise-enterprise associations and enterprise-news associations. For example, Figure 1 As shown in , the labeled news text N_labeled as the center is a news node, and the corporate entity A associated with it is the corporate node of N_labeled, and the edge between them represents the corporate-news relationship. Corporate entity B, which has an investment relationship with corporate entity A, can be associated with three unlabeled news texts, and these three unlabeled news texts can all be news texts related to corporate entity B. And corporate entity C, which has a common executive relationship with corporate entity A, can be associated with two unlabeled news texts, and these two unlabeled news texts can be news texts related to corporate entity C. In particular, other unlabeled news texts related to corporate entity A can be further obtained as news nodes based on corporate entity A.

[0081] S302: Use the annotated text as training data to train a text classification model.

[0082] After obtaining the relationship graph including the extended unlabeled news text based on the labeled text N_labeled in step S301, in step S302, the labeled text N_labeled can be first used as input to train an initial multi-label classification model using a text learning model.

[0083] In the embodiment of the present application, the classification model used in step S302 can be a relatively simple TextCNN model, or a slightly more complex model such as a HAN model, TextRNN, etc., or a more complex model such as BERT or BERT+LSTM. Therefore, such a classification model can be used to train the labeled text N_labeled as training data until convergence. Therefore, in step S302, a text classification model based on the labeled text N_labeled can be obtained.

[0084] S303, input the text nodes in the relationship graph into the text classification model to obtain a predicted pseudo label and a predicted feature vector for each text node.

[0085] In step S303, the classification model obtained in step S302 can be used to classify the unlabeled news text nodes in the relationship graph obtained in step S301, so that a text classification prediction result can be obtained for each unlabeled news node as a pseudo label P_label of the node, and a text prediction feature vector P_feature can also be obtained.

[0086] S304, obtaining a text node whose predicted pseudo label is the same as the label of the central node as the first text node.

[0087] S305: Calculate the average value of the predicted feature vectors of the first text node as a first average vector.

[0088] S306: Concatenate or average the first average vector and the feature vector of the central node to obtain a converged feature vector.

[0089] In steps S304-S306, the features of the nodes with the same pseudo labels of the unlabeled text obtained in step S302 and the labeled labels of the labeled text N_labeled can be aggregated to the central node by connection or averaging. For example, in step S304, the pseudo labels of each text node of the unlabeled text obtained in step S302 obtained by the text classification model can be first compared with the labels of the labeled text in step S301, so that the text nodes with the same labels can be used as the first text nodes.

[0090] In step S305 , the feature vectors P_feature of the unlabeled texts with the same pseudo-label as the labeled text N_labeled determined in step S304 may be averaged to obtain an average vector Neigh_feature, which is used as the first average vector.

[0091] In step S306, the average feature vector Neigh_feature obtained in step S305 may be concatenated or averaged with the T_feature of the central node to obtain a converged feature vector Next T_feature.

[0092] S307, calculating the similarity between the predicted pseudo-label of each text node and the label of the central node as a weight value.

[0093] S308: Calculate a weighted average value of the predicted feature vectors of the text nodes according to the weight values ​​as a second average vector.

[0094] S309: Concatenate or average the second average vector and the feature vector of the central node to obtain a converged feature vector.

[0095] In the embodiment of the present application, the more similar the neighboring node news is to the central news node N_labeled, the greater the role it should play in the prediction task of the central node, and the more different the node is from the central node, the smaller the role it should play in the convergence of the central node. Therefore, in addition to steps S304-S306, in steps S307-S309, based on the pseudo-label P_label of each unlabeled news node in the relationship graph obtained in step S301 calculated in step S303 using the text classification model trained in step S302, the similarity between the pseudo-label P_label calculated for each unlabeled news node in step S303 by the text classification model trained in step S302 and the labeled label of the central news node N_labeled can be further calculated in step S307.

[0096] In step S308, the similarity between the pseudo-label P_label of each unlabeled news node obtained in step S307 through the above text classification model and the label of the central node can be used as the aggregation weight, and the vectors of the unlabeled nodes with the same pseudo-label as the labeled label of the central node can be merged by weighted averaging to obtain Neigh_feature.

[0097] In step S309, the second average vector obtained in step S308 is concatenated or averaged with the current T_feature of the central node to obtain a new feature vector Next T_feature of the central node.

[0098] S310, using the aggregated feature vector as training data to train the graph classification model.

[0099] Therefore, in step S204, the new central node obtained in step S203 that is related to the annotated text but has different features can be used, and the new features and the annotated labels can be used together to form new training data to be input into the graph classification model.

[0100] S311, input the text nodes in the relationship graph into the graph classification model to obtain the predicted pseudo-label and predicted feature vector for each text node.

[0101] In step S311, the graph classification model trained in step S310 can be used to further perform classification calculations on each unlabeled news node in the relationship graph constructed in step S301. That is, the text nodes in the relationship graph obtained in step S301 are input into the graph classification model trained in step S310 to obtain a predicted pseudo-label and predicted feature vector for each text node. Thus, the predicted pseudo-label and predicted feature vector obtained in this way can be used to update the relationship graph.

[0102] S312: Update the text classification model using the prediction feature layer including the prediction feature vectors of each text node.

[0103] In step S312, the text classification model in step S303 can be updated based on the predicted feature vectors obtained for each node in step S311 using the graph classification model after the previous round of training. In particular, in an embodiment of the present application, the predicted feature vectors generated in step S311 can be used to form a predicted feature layer, so the predicted feature layer can include the predicted feature vectors generated by the graph classification model trained through steps S304-S306 and steps S307-S309. Therefore, such a predicted feature layer can be used in step S312 as the result of the previous round to substitute into the text classification model, so that the calculations of steps S304-S306 and steps S307-S309 can be performed again in an iterative manner until the graph classification model reaches convergence.

[0104] Therefore, the model training method provided in the embodiment of the present application obtains a relationship graph with the annotated text as the central node based on the annotated text with labels, inputs the text nodes in the relationship graph into the text classification model to obtain a predicted pseudo-label and a predicted feature vector for each text node, and then converges each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector, and uses the converged feature vector as training data to train the graph classification model. Therefore, the annotated text can be used as the central node in the relationship graph to expand and obtain neighbor nodes including unlabeled text, and then the nodes of the unlabeled text are converged with the central node according to the predicted pseudo-label of each unlabeled text, thereby generating a converged feature vector that is both related to the annotated text and different from the annotated text, so that the converged feature vector can be used together with the label of the annotated text as new training data to train the graph classification model, thereby achieving effective expansion of the annotated text, and thus improving the recognition accuracy of the trained model by adding a large amount of training data related to the annotated text.

[0105] Embodiment 4

[0106] Figure 4 A schematic diagram of the structure of an embodiment of the model training device provided in this application can be used to perform Figure 2 or Figure 3 The model training method shown in . Figure 4 As shown, the model training device may include: an acquisition module 41, a text classification module 42, an aggregation module 43 and a training module 44.

[0107] The acquisition module 41 can be used to acquire a relationship graph with the annotated text as a central node based on the annotated text with the label.

[0108] The acquisition module 41 can acquire one or more labeled and classified text data that have been annotated, so as to use this batch of annotated text as a training set. Figure 1 In the scenario shown in , the acquisition module 41 can acquire a labeled text N_labeled from a labeled data source such as the Internet, and the labeled text can have a classification label determined for it through the labeling process. Therefore, the labeled text can be used as a central node in the relationship map.

[0109] In an embodiment of the present application, the relationship graph may be the business entity corresponding to each text, and a graph consisting of an entity set may be obtained by setting rules or using an entity extraction model to pre-extract the annotated text.

[0110] In the embodiment of the present application, the news texts may all be news texts that have not been annotated. In other words, in the embodiment of the present application, since the annotation of news texts requires a large amount of human resources, the number of existing annotated texts is small. Therefore, when expanding news texts based on corporate entities in the embodiment of the present application, news texts that have not been annotated can be mainly obtained as extended text nodes of annotated news texts. Of course, annotated texts that have been annotated can also be obtained as extended text points of annotated news text A. The present application has no restrictions on this.

[0111] In the relational graph formed by the labeled text N_labeled as the central node and the unlabeled text extended by the entity relationship, each node can include two categories: news nodes and enterprise nodes. And the edges can include two categories: enterprise-enterprise associations and enterprise-news associations. For example, Figure 1 As shown in , the labeled news text N_labeled as the center is a news node, and the corporate entity A associated with it is the corporate node of N_labeled, and the edge between them represents the corporate-news relationship. Corporate entity B, which has an investment relationship with corporate entity A, can be associated with three unlabeled news texts, and these three unlabeled news texts can all be news texts related to corporate entity B. And corporate entity C, which has a common executive relationship with corporate entity A, can be associated with two unlabeled news texts, and these two unlabeled news texts can be news texts related to corporate entity C. In particular, other unlabeled news texts related to corporate entity A can be further obtained as news nodes based on corporate entity A.

[0112] The text classification module 42 may be used to input the text nodes in the relationship graph into a text classification model to obtain a predicted pseudo label and a predicted feature vector for each text node.

[0113] After the acquisition module 41 obtains the relationship graph including the extended unlabeled news text based on the labeled text N_labeled, the text classification module 42 can first use the labeled text N_labeled as input to train an initial multi-label classification model using a text learning model.

[0114] In an embodiment of the present application, the classification model used by the text classification module 42 can be a relatively simple TextCNN model, or a slightly more complex model such as a HAN model, TextRNN, etc., or a more complex model such as BERT or BERT+LSTM. Therefore, such a classification model can be used to train the labeled text N_labeled as training data until convergence. Therefore, a text classification model based on the labeled text N_labeled is obtained, so that the text classification module 42 can use the classification model to classify the unlabeled news text nodes in the relationship graph obtained by the acquisition module 41, so that the text classification prediction result can be obtained for each unlabeled news node as the pseudo-label P_label of the node, and the text prediction feature vector P_feature can also be obtained.

[0115] The aggregation module 43 may be used to aggregate each text node and the central node according to the predicted pseudo-label to obtain an aggregated feature vector.

[0116] The aggregation module 43 may aggregate the feature vectors of each unlabeled text node in the relationship graph obtained by the acquisition module 41 through the text classification model with the features of the labeled nodes, for example by concatenation or averaging, so as to obtain an aggregated feature vector.

[0117] For example, the aggregation module 43 can aggregate the features of the nodes with the same pseudo-label of the unlabeled text obtained in step S202 and the labeled label of the labeled text N_labeled to the central node by connecting or averaging. For example, the aggregation module 43 can first compare the pseudo-label obtained by the text classification model of each text node of the unlabeled text obtained by the text classification module 42 with the label of the labeled text obtained in the acquisition module 41, so that the feature vector P_feature of the unlabeled text with the same pseudo-label as the labeled text N_labeled can be averaged to obtain the average vector Neigh_feature, and then the averaged feature vector Neigh_feature is spliced ​​or averaged with the T_feature of the central node to obtain a new central node Next T_feature.

[0118] In addition, in the embodiment of the present application, the more similar the neighbor node news is to the central news node N_labeled, the greater the role it should play in the prediction task of the central node, and the node with the greater difference in the central node representation should play as little role as possible in the convergence of the central node. Therefore, on the basis of using the trained text classification model to calculate the pseudo-label P_label of each unlabeled news node in the relationship graph, in the embodiment of the present application, the similarity between the pseudo-label P_label calculated by the above text classification model for each unlabeled news node and the labeled label of the central news node N_labeled can be further calculated. And the similarity between the pseudo-label P_label obtained by the above text classification model for each unlabeled news node and the label of the central node is used as the weight of convergence, and the vectors of the unlabeled nodes with the same pseudo-label as the labeled label of the central node are merged by weighted averaging to obtain Neigh_feature, and then it is spliced ​​or averaged with the current T_feature of the central node to obtain the new feature vector Next T_feature of the central node. Thus, the graph classification model can be trained similarly based on the new feature vector Next T_feature of the central node and the labeled label to form training data.

[0119] The training module 44 may be used to train a graph classification model using the aggregated feature vectors as training data.

[0120] Therefore, the training module 44 can use the new central node obtained by the aggregation module 43 that is related to the annotated text but has different features, and can use the new features together with the annotated labels to form new training data to input into the graph classification model.

[0121] In addition, after the graph classification model is trained using the new feature vector as described above, in the embodiment of the present application, the trained model can be used to further perform classification calculations on each unlabeled news node in the relationship graph constructed in the above manner. That is, the text nodes in the relationship graph are input into the trained graph classification model to obtain the predicted pseudo-label and predicted feature vector for each text node. Thus, the predicted pseudo-label and predicted feature vector obtained in this way can be used to update the relationship graph, and then the above calculations are performed again until the graph classification model reaches convergence.

[0122] Therefore, the model training device provided in the embodiment of the present application obtains a relationship graph with the annotated text as the central node based on the annotated text with labels, inputs the text nodes in the relationship graph into the text classification model to obtain a predicted pseudo-label and a predicted feature vector for each text node, and then converges each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector, and uses the converged feature vector as training data to train the graph classification model. Therefore, the annotated text can be used as the central node in the relationship graph to expand and obtain neighboring nodes including unlabeled text, and then the nodes of the unlabeled text are converged with the central node according to the predicted pseudo-label of each unlabeled text, thereby generating a converged feature vector that is both related to the annotated text and different from the annotated text, so that the converged feature vector can be used together with the label of the annotated text as new training data to train the graph classification model, thereby achieving effective expansion of the annotated text, and thus improving the recognition accuracy of the trained model by adding a large amount of training data related to the annotated text.

[0123] Embodiment 5

[0124] The above describes the internal functions and structure of the model training device, which can be implemented as an electronic device. Figure 5 This is a schematic diagram of the structure of an electronic device embodiment provided by this application. Figure 5 As shown, the electronic device includes a memory 51 and a processor 52 .

[0125] The memory 51 is used to store programs. In addition to the above programs, the memory 51 can also be configured to store various other data to support operations on the electronic device. Examples of such data include instructions for any application or method operating on the electronic device, contact data, phone book data, messages, pictures, videos, etc.

[0126] The memory 51 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0127] The processor 52 is not limited to a processor (CPU), but may also be a processing chip such as a graphics processing unit (GPU), a field programmable gate array (FPGA), an embedded neural network processor (NPU) or an artificial intelligence (AI) chip. The processor 52 is coupled to the memory 51 and executes the program stored in the memory 51 to execute the model training method of the above-mentioned embodiment 2 or 3.

[0128] Further, if Figure 5 As shown, the electronic device may also include: a communication component 53, a power component 54, an audio component 55, a display 56 and other components. Figure 5 Only some components are shown schematically, which does not mean that the electronic device only includes Figure 5 Components shown.

[0129] The communication component 53 is configured to facilitate wired or wireless communication between the electronic device and other devices. The electronic device can access a wireless network based on a communication standard, such as a mobile communication network such as WiFi, 2G, 3G, 4G / LTE, 5G, or a combination thereof. In an exemplary embodiment, the communication component 53 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 53 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0130] The power supply component 54 provides power to various components of the electronic device. The power supply component 54 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device.

[0131] The audio component 55 is configured to output and / or input audio signals. For example, the audio component 55 includes a microphone (MIC), and when the electronic device is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 51 or sent via the communication component 53. In some embodiments, the audio component 55 also includes a speaker for outputting audio signals.

[0132] The display 56 includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0133] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A model training method, comprising: According to the labeled text, a relationship graph with the labeled text as a central node is obtained, wherein the relationship graph includes neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes include text nodes and entity nodes; Inputting the text nodes in the relationship graph into a text classification model to obtain a predicted pseudo label and a predicted feature vector for each of the text nodes; According to the predicted pseudo-label, each text node is converged with the central node to obtain a converged feature vector; According to the predicted pseudo-label, each text node is converged with the central node to obtain a converged feature vector, including: obtaining the text node whose predicted pseudo-label is the same as the label of the central node as the first text node; calculating the average value of the predicted feature vector of the first text node as the first average vector; splicing or averaging the first average vector with the feature vector of the central node to obtain the converged feature vector; The aggregated feature vector is used as training data to train a graph classification model.

2. The model training method according to claim 1, wherein: The step of aggregating each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector includes: Calculate the similarity between the predicted pseudo-label of each text node and the label of the central node as a weight value; According to the weight value, calculating the weighted average of the predicted feature vectors of the text node as a second average vector; The second average vector is concatenated or averaged with the feature vector of the central node to obtain the converged feature vector.

3. The model training method according to claim 1, wherein: The method further comprises: The annotated text is used as training data to train the text classification model.

4. The model training method according to claim 1, wherein: The method further comprises: The text nodes in the relationship graph are input into the graph classification model to obtain a predicted pseudo label and a predicted feature vector for each of the text nodes.

5. The model training method according to claim 4, wherein: When the graph classification model is used for prediction, the method further includes: Using the output of the intermediate feature layer of the graph classification model as the predicted feature vector of the text node; Calculate the similarity between the predicted feature vector of each text node and the feature vector of the central node as a weight value; According to the weight value, calculating the weighted average of the predicted feature vectors of the text node as a third average vector; The third average vector and the feature vector of the central node are concatenated or averaged to obtain the converged feature vector.

6. A text classification method, comprising: Performing text feature extraction on the target text to obtain a text feature vector of the target text; Inputting the text feature vector of the target text into a predetermined graph classification model to calculate the correlation between the target text and each predetermined classification label; Determine the classification label of the target text according to the relevance, wherein The predetermined graph classification model is trained using the following method: According to the annotated text with the classification label, a relationship graph with the annotated text as the central node is obtained, wherein the relationship graph includes neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes include text nodes and entity nodes; Inputting the text nodes in the relationship graph into a text classification model to obtain a predicted pseudo label and a predicted feature vector for each of the text nodes; According to the predicted pseudo-label, each text node is converged with the central node to obtain a converged feature vector; According to the predicted pseudo-label, each text node is converged with the central node to obtain a converged feature vector, including: obtaining the text node whose predicted pseudo-label is the same as the label of the central node as the first text node; calculating the average value of the predicted feature vector of the first text node as the first average vector; splicing or averaging the first average vector with the feature vector of the central node to obtain the converged feature vector; The graph classification model is trained using the aggregated feature vector as training data.

7. A model training device, comprising: An acquisition module, configured to acquire, based on the labeled text, a relationship graph with the labeled text as a central node, wherein the relationship graph includes neighbor nodes directly or indirectly associated with the central node, and the neighbor nodes include text nodes and entity nodes; A text classification module, used for inputting the text nodes in the relationship graph into a text classification model to obtain a predicted pseudo label and a predicted feature vector for each of the text nodes; A convergence module is used to converge each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector; the convergence of each text node with the central node according to the predicted pseudo-label to obtain a converged feature vector includes: obtaining the text node whose predicted pseudo-label is the same as the label of the central node as the first text node; calculating the average value of the predicted feature vector of the first text node as the first average vector; concatenating or averaging the first average vector with the feature vector of the central node to obtain the converged feature vector; A training module is used to use the aggregated feature vector as training data to train a graph classification model.

8. An electronic device, comprising: Memory, used to store programs; A processor, configured to run the program stored in the memory to execute the model training method as described in any one of claims 1 to 5.

9. A computer-readable storage medium having stored thereon a computer program executable by a processor, wherein: When the program is executed by a processor, the model training method as described in any one of claims 1 to 5 is implemented.

Images