People's Bank of China credit information processing method and system based on LSTM model

Abstract

The present application discloses a method and system for processing credit information of the People's Bank of China based on the LSTM model. It first passes the credit report of the People's Bank of China of the user to be detected through a context encoder to obtain multiple word context semantic feature vectors, then arranges the multiple word context semantic feature vectors into a one-dimensional feature vector and passes through a phrase granularity understanding model to obtain a first-scale report understanding feature vector, then inputs the multiple word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report understanding feature vector, then cascades the multiple word context semantic feature vectors to obtain a third-scale report understanding feature vector, and finally, passes the classification feature vector obtained by fusing the first to third-scale report understanding feature vectors through a classifier to obtain a classification result for indicating whether the user to be detected is a compensation repurchase user. In this way, it is possible to accurately detect and judge whether the user to be detected is a compensation repurchase user.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection technology, and more specifically, to a PBOC credit information processing method and system based on an LSTM model. Background Art

[0002] PBOC credit report repurchase refers to the practice of financial institutions proactively repaying or settling overdue payments on behalf of clients to maintain the activity of their platforms. This practice results in the removal or weakening of overdue information on clients' PBOC credit reports. This undermines the system for thwarting dishonesty, jeopardizes the development of the social credit system, and can even cause other financial institutions to suffer financial losses due to misguided information. Therefore, identifying PBOC credit report repurchase is crucial and significant.

[0003] Patent application number CN112634028A discloses a method for identifying instances of loan repayment and repurchase in a People's Bank of China (PBOC) credit report. The method comprises the following steps: Step A: Automatically extracting features of loan repayment and repurchase in a PBOC credit report using a recurrent neural network; Step B: Utilizing an attention mechanism to enhance the network's effectiveness in identifying loan repayment and repurchase features in a PBOC credit report; and Step C: Building a model based on the user's loan agreement level in a practical application scenario. When a user has a loan that has loan repayment and repurchase behavior, the user is identified as a loan repayment and repurchase user. This method utilizes deep learning to automatically extract features and uses an algorithm to identify features that can better distinguish loan repayment and repurchase users. However, this method does not fully understand the semantics of PBOC credit reports, namely, it does not consider the contextual semantic features of PBOC credit reports. Furthermore, PBOC credit reports often contain a large amount of irrelevant semantic information, resulting in low accuracy in determining whether a user is a loan repayment and repurchase user.

[0004] Therefore, considering that the identification of repurchase and repayment users in the PBOC credit report requires a thorough understanding of the semantics of the PBOC credit report to more accurately determine whether a user is a repurchase and repayment user, the PBOC credit report contains a large amount of irrelevant semantic information, which makes it difficult to determine whether a user is a repurchase and repayment user.

[0005] Therefore, an optimized PBOC credit information processing solution is desired, which can accurately detect and determine whether the user to be detected is a compensation repurchase user. Summary of the Invention

[0006] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a method and system for processing the credit information of the People's Bank of China based on the LSTM model. It first passes the credit report of the People's Bank of China of the user to be detected through a context encoder to obtain a plurality of word context semantic feature vectors, then arranges the plurality of word context semantic feature vectors into a one-dimensional feature vector and passes through a phrase granularity understanding model to obtain a first-scale report understanding feature vector, then inputs the plurality of word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report understanding feature vector, then cascades the plurality of word context semantic feature vectors to obtain a third-scale report understanding feature vector, and finally, passes the classification feature vector obtained by fusing the first to third-scale report understanding feature vectors through a classifier to obtain a classification result for indicating whether the user to be detected is a compensation repurchase user. In this way, it is possible to accurately detect and judge whether the user to be detected is a compensation repurchase user.

[0007] According to one aspect of the present application, a method for processing credit information of the People's Bank of China based on an LSTM model is provided, which includes:

[0008] Obtain the People's Bank of China credit report of the user to be tested;

[0009] Passing the PBOC credit report through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors;

[0010] Arranging the multiple word context semantic feature vectors into a one-dimensional feature vector and then applying the result to a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector;

[0011] Inputting the plurality of word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report comprehension feature vector;

[0012] cascading the multiple word context semantic feature vectors to obtain a third scale report comprehension feature vector;

[0013] The first to third scale report comprehension feature vectors are integrated to obtain a classification feature vector; and the classification feature vector is passed through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the user to be detected is a compensation repurchase user.

[0014] In the above-mentioned PBOC credit information processing method based on the LSTM model, the PBOC credit report is passed through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors, including:

[0015] Performing word segmentation processing on the PBOC credit report to convert the PBOC credit report into a word sequence consisting of multiple words;

[0016] Use the embedding layer of the context encoder to map each word in the word sequence to a word vector to obtain a sequence of word vectors; and use the context encoder to perform global context semantic encoding on the sequence of word vectors to obtain the multiple word context semantic feature vectors.

[0017] In the above-mentioned PBOC credit information processing method based on the LSTM model, the method of arranging the multiple word context semantic feature vectors into a one-dimensional feature vector and then passing the result through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector includes: using each layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network to perform the following on the input data in the forward pass of the layer:

[0018] Performing convolution processing on the input data to obtain a convolution feature vector;

[0019] Performing pooling processing on the convolution feature vector to obtain a pooled feature vector; and performing nonlinear activation on the pooled feature vector to obtain an activated feature vector;

[0020] The output of the last layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the first-scale report understanding feature vector, and the input of the first layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the one-dimensional feature vector.

[0021] In the above-mentioned PBOC credit information processing method based on the LSTM model, the fusing of the first to third scale report understanding feature vectors to obtain a classification feature vector includes: fusing the first to third scale report understanding feature vectors to obtain a classification feature vector using the following formula;

[0022] Wherein, the formula is:

[0023]

[0024] in, Represents the first to third scale reported understanding feature vector, Represents a cascade function, represents the classification feature vector.

[0025] In the above-mentioned PBOC credit information processing method based on the LSTM model, the classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the user to be detected is a compensation repurchase user, including:

[0026] Performing full connection encoding on the classification feature vector using the fully connected layer of the classifier to obtain an encoded classification feature vector;

[0027] The encoded classification feature vector is input into the Softmax classification function of the classifier to obtain the classification result.

[0028] In the above-mentioned PBOC credit information processing method based on the LSTM model, a training step is also included: training the context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier.

[0029] In the above-mentioned PBOC credit information processing method based on the LSTM model, the training step includes:

[0030] Acquire training data, the training data including a training PBOC credit report of the user to be tested, and a true value of whether the user to be tested is a compensation repurchase user;

[0031] Passing the training PBOC credit report through the context encoder including the word embedding layer to obtain a plurality of training word context semantic feature vectors;

[0032] Arranging the plurality of training word context semantic feature vectors into a training one-dimensional feature vector and then applying the one-dimensional convolutional neural network-based phrase granularity understanding model to obtain a first training scale report understanding feature vector;

[0033] Inputting the plurality of training word context semantic feature vectors into the bidirectional LSTM model to obtain a second training scale report comprehension feature vector;

[0034] cascading the plurality of training word context semantic feature vectors to obtain a third training scale report comprehension feature vector;

[0035] fusing the first to third training scale report comprehension feature vectors to obtain a training classification feature vector;

[0036] Passing the training classification feature vector through the classifier to obtain a classification loss function value;

[0037] The context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier are trained based on the classification loss function value and through the direction propagation of gradient descent, wherein in each round of iteration of the training, the free label optimization factor based on the cross-classifier soft similarity of the training classification feature vector is calculated to iterate the training classification feature vector.

[0038] In the above-mentioned PBOC credit information processing method based on the LSTM model, in each round of training iteration, the free label optimization factor based on cross-classifier soft similarity of the training classification feature vector is calculated using the following formula as the label value of the classifier;

[0039] Wherein, the formula is:

[0040]

[0041] is the classification feature vector, is the weight matrix of the classifier for the training classification feature vector, and represent tensor multiplication and tensor addition respectively, represents the distance between vectors, represents the two-norm of the vector, and and is the weight hyperparameter, represents an exponential operation of a vector, wherein the exponential operation of a vector represents calculating the value of a natural exponential function raised to the power of the eigenvalue of each position in the vector.

[0042] According to another aspect of the present application, a PBOC credit information processing system based on an LSTM model is provided, comprising:

[0043] A data acquisition module, used to obtain the People's Bank of China credit report of the user to be tested;

[0044] A context encoding module, configured to pass the PBOC credit report through a context encoder comprising a word embedding layer to obtain a plurality of word context semantic feature vectors;

[0045] A first scale encoding module, configured to arrange the plurality of word context semantic feature vectors into a one-dimensional feature vector and then pass the result through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first scale report understanding feature vector;

[0046] a second-scale encoding module, configured to input the plurality of word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report comprehension feature vector;

[0047] a concatenation module, configured to concatenate the plurality of word context semantic feature vectors to obtain a third-scale report comprehension feature vector;

[0048] A fusion module is used to fuse the first to third scale report understanding feature vectors to obtain a classification feature vector; and a classification result generation module is used to pass the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the user to be detected is a compensation repurchase user.

[0049] In the above-mentioned PBOC credit information processing system based on the LSTM model, the context encoding module includes:

[0050] A word segmentation processing unit, configured to perform word segmentation processing on the PBOC credit report to convert the PBOC credit report into a word sequence consisting of a plurality of words;

[0051] An embedding coding unit, used to use the embedding layer of the context encoder to map each word in the word sequence to a word vector to obtain a sequence of word vectors; and a context coding unit, used to use the context encoder to perform global context semantic encoding on the sequence of word vectors to obtain the multiple word context semantic feature vectors.

[0052] Compared with the prior art, the PBOC credit information processing method and system based on the LSTM model provided by the present application first passes the PBOC credit report of the user to be detected through a context encoder to obtain multiple word context semantic feature vectors, then arranges the multiple word context semantic feature vectors into one-dimensional feature vectors and passes them through a phrase granularity understanding model to obtain a first-scale report understanding feature vector, then inputs the multiple word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report understanding feature vector, then cascades the multiple word context semantic feature vectors to obtain a third-scale report understanding feature vector, and finally, passes the classification feature vector obtained by fusing the first to third-scale report understanding feature vectors through a classifier to obtain a classification result for indicating whether the user to be detected is a compensation repurchase user. In this way, it is possible to accurately detect and judge whether the user to be detected is a compensation repurchase user. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0054] Figure 1 This is a scenario diagram of the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application.

[0055] Figure 2 This is a flowchart of the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application.

[0056] Figure 3 Schematic diagram of the architecture of the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application.

[0057] Figure 4 This is a flowchart of the sub-steps of step S120 in the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application.

[0058] Figure 5 This is a flowchart of the sub-steps of step S130 in the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application.

[0059] Figure 6 This is a flowchart of the sub-steps of step S170 in the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application.

[0060] Figure 7 This is a flowchart of the training steps further included in the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application.

[0061] Figure 8 This is a block diagram of the PBOC credit information processing system based on the LSTM model according to an embodiment of the present application. DETAILED DESCRIPTION

[0062] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0063] Scenario Overview

[0064] Taking into account the identification of compensation and repurchase in the People's Bank of China credit report, it is necessary to fully understand the semantics of the People's Bank of China credit report in order to more accurately determine whether the user is a compensation and repurchase user. However, since there is a large amount of irrelevant semantic information in the People's Bank of China credit report, it brings difficulties to the detection and judgment of whether the user is a compensation and repurchase user. Based on this, in the technical solution of the present application, an artificial intelligence semantic understanding algorithm based on deep learning is used to complete the detection of whether the user to be detected is a compensation and repurchase user through coding analysis of the People's Bank of China credit report of the user to be detected and classification judgment of the classifier. That is, the artificial intelligence technology based on deep learning is applied to the semantic understanding of the natural language processing of the People's Bank of China credit information to construct a People's Bank of China credit information processing solution based on the LSTM model. In this way, the semantic understanding of the People's Bank of China credit information can be accurately carried out to accurately perform intelligent detection and judgment on whether the user to be detected is a compensation and repurchase user.

[0065] Specifically, in the technical solution of the present application, first, the People's Bank of China credit report of the user to be detected is obtained. Then, considering that the People's Bank of China credit report of the user to be detected is a sentence composed of multiple words or phrases, in order to be able to understand its semantics to obtain the compensation repurchase user detection of the user to be detected, the People's Bank of China credit report is further encoded through a context encoder including a word embedding layer to extract the global high-dimensional semantic features of each word in the People's Bank of China credit report, thereby obtaining multiple word context semantic feature vectors. That is, specifically, first, the People's Bank of China credit report is segmented to facilitate semantic confusion during the subsequent semantic feature extraction; then, the word embedding layer is used to embed each word after the segmentation process to map it into a word embedding vector, and the context encoding converter is used to perform global context semantic encoding on the sequence of the obtained word embedding vectors to generate the multiple word context semantic feature vectors. Further, the multiple word context semantic feature vectors are cascaded to obtain a third-scale report understanding feature vector with global word semantic features.

[0066] Then, considering that the contextual semantic association features of each word in the PBOC credit report of the user to be tested do not enable a good contextual semantic understanding, this is because the PBOC credit report also contains phrases composed of individual words that play an important role in the semantic understanding of the PBOC credit report of the user to be tested. Therefore, in the technical solution of this application, the multiple word contextual semantic feature vectors are further arranged into a one-dimensional feature vector and then extracted using a phrase granularity understanding model based on a one-dimensional convolutional neural network to extract the locally associated phrase feature information of the semantic understanding features of each word, thereby obtaining a first-scale report understanding feature vector.

[0067] It should be understood that, considering the distribution of contextual semantic understanding features of each word, the phrase features composed of each word not only have local short-distance dependency correlation, but also have long-distance dependency correlation feature distribution information. In other words, the contextual semantic features of each word in the PBOC credit report of the user to be detected have long-distance dependency semantic correlation, which can provide a more accurate semantic understanding of the PBOC credit information. Therefore, in the technical solution of the present application, the multiple word contextual semantic feature vectors are further input into the bidirectional LSTM model to obtain a second-scale report understanding feature vector. It should be understood that the bidirectional long short-term memory neural network model (LSTM) enables the weights of the neural network to self-update by adding input gates, output gates, and forget gates. When the network model parameters are fixed, the weight scales of different channels can be changed dynamically, thereby avoiding the problem of gradient disappearance or gradient expansion. In particular, the bidirectional long short-term memory neural network model is composed of a forward LSTM and a backward LSTM. The forward LSTM can learn the previous text information of the current word, and the backward LSTM can learn the subsequent text information of the current word. Therefore, the second-scale report understanding feature vector obtained by the bidirectional long short-term memory neural network model learns the semantic information of the global context.

[0068] Furthermore, the first through third scaled report understanding feature vectors are integrated to integrate the semantic understanding feature distribution information of multi-scale word associations in the PBOC credit report of the user to be tested. This is then used as a classification feature vector for classification processing in a classifier to obtain a classification result indicating whether the user to be tested is a repurchase repurchase user. This enables semantic understanding of the PBOC credit information to accurately and intelligently determine whether the user to be tested is a repurchase repurchase user.

[0069] In particular, in the technical solution of the present application, since the first to third scale report comprehension feature vectors respectively represent contextual association features of word semantics at different scales, it is preferred to obtain the classification feature vector by directly concatenating the first to third scale report comprehension feature vectors, so that the classification feature vector fully expresses the local-remote word semantic context information within the PBOC credit report. However, this will result in the overall distribution of the classification feature vector having a strong discreteness, making classifier training difficult, especially the convergence of the classifier's label values ​​difficult.

[0070] Therefore, the applicant of the present application uses soft label learning to replace the commonly used hard label learning. Specifically, at each iteration, the free label optimization factor based on the cross-classifier soft similarity of the training classification feature vector is calculated as the label value of the classifier using the following formula:

[0071] Wherein, the formula is:

[0072]

[0073] is the classification feature vector, is the weight matrix of the classifier for the training classification feature vector, and represent tensor multiplication and tensor addition respectively, represents the distance between vectors, represents the two-norm of the vector, and and is the weight hyperparameter, represents an exponential operation of a vector, wherein the exponential operation of a vector represents calculating the value of a natural exponential function raised to the power of the eigenvalue of each position in the vector.

[0074] Here, the free label optimization factor based on cross-classifier soft similarity simulates pseudo-categories based on the classifier weight matrix through cross-classifier soft similarity of the feature vector and the weight matrix by performing bidirectional clustering before calculating the classification probability of the feature vector with the hard label value. This avoids the classification quantization loss caused by hard label learning through soft similarity learning, and achieves free label optimization that focuses more on the intrinsic weight structure of the classifier, thereby optimizing the training of the classifier label value and improving the training speed of the classifier. In this way, the semantic understanding of the credit information of the People's Bank of China can be more accurately performed, so as to more accurately perform intelligent detection and judgment on whether the user to be detected is a compensation repurchase user.

[0075] Based on this, the present application provides a method for processing credit information of the People's Bank of China based on the LSTM model, which includes: obtaining the credit report of the People's Bank of China of the user to be detected; passing the credit report of the People's Bank of China through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors; arranging the multiple word context semantic feature vectors into one-dimensional feature vectors and then passing them through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector; inputting the multiple word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report understanding feature vector; cascading the multiple word context semantic feature vectors to obtain a third-scale report understanding feature vector; fusing the first to third-scale report understanding feature vectors to obtain a classification feature vector; and passing the classification feature vector through a classifier to obtain a classification result, which is used to indicate whether the user to be detected is a compensation repurchase user.

[0076] Figure 1This is an application scenario diagram of the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application. Figure 1 As shown, in this application scenario, the credit report of the user to be detected is obtained (for example, Figure 1 Then, the PBOC credit report is input to a server that has deployed a PBOC credit information processing algorithm based on an LSTM model (e.g., Figure 1 In the S) shown in , the server is able to generate a classification result indicating whether the user to be detected is a compensation repurchase user based on the PBOC credit information processing algorithm based on the LSTM model.

[0077] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0078] Exemplary Methods

[0079] Figure 2 Flowchart of the PBOC credit information processing method based on the LSTM model according to an embodiment of the present application. Figure 2 As shown, according to the LSTM model-based credit information processing method of the People's Bank of China in an embodiment of the present application, the steps include: S110, obtaining the People's Bank of China credit report of the user to be detected; S120, passing the People's Bank of China credit report through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors; S130, arranging the multiple word context semantic feature vectors into one-dimensional feature vectors and then passing them through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector; S140, inputting the multiple word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report understanding feature vector; S150, cascading the multiple word context semantic feature vectors to obtain a third-scale report understanding feature vector; S160, fusing the first to third-scale report understanding feature vectors to obtain a classification feature vector; and, S170, passing the classification feature vector through a classifier to obtain a classification result, which is used to indicate whether the user to be detected is a compensation repurchase user.

[0080] Figure 3 Schematic diagram of the structure of the PBOC credit information processing method based on the LSTM model according to the embodiment of the present application. Figure 3As shown, in this network architecture, first, the People's Bank of China credit report of the user to be detected is obtained; then, the People's Bank of China credit report is passed through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors; then, the multiple word context semantic feature vectors are arranged into one-dimensional feature vectors and then passed through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector; then, the multiple word context semantic feature vectors are input into a bidirectional LSTM model to obtain a second-scale report understanding feature vector; then, the multiple word context semantic feature vectors are cascaded to obtain a third-scale report understanding feature vector; then, the first to third-scale report understanding feature vectors are fused to obtain a classification feature vector; finally, the classification feature vector is passed through a classifier to obtain a classification result, which is used to indicate whether the user to be detected is a compensation repurchase user.

[0081] More specifically, in step S110, the credit report of the People's Bank of China of the user to be detected is obtained. The identification of the compensation repurchase in the credit report of the People's Bank of China requires a full semantic understanding of the credit report of the People's Bank of China, so as to more accurately determine whether the user is a compensation repurchase user. However, since there is a large amount of irrelevant semantic information in the credit report of the People's Bank of China, it brings difficulties to the detection and judgment of whether the user is a compensation repurchase user. Based on this, in the technical solution of the present application, an artificial intelligence semantic understanding algorithm based on deep learning is adopted to complete the detection of whether the user to be detected is a compensation repurchase user through the coding analysis of the credit report of the People's Bank of China of the user to be detected and the classification judgment of the classifier. That is, the artificial intelligence technology based on deep learning is applied to the semantic understanding of the natural language processing of the credit information of the People's Bank of China to construct a credit information processing solution of the People's Bank of China based on the LSTM model.

[0082] More specifically, in step S120, the PBOC credit report is passed through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors. The PBOC credit report of the user to be detected is a sentence composed of multiple words or phrases. Therefore, in order to be able to perform semantic understanding on it to obtain the compensation repurchase user detection of the user to be detected, the PBOC credit report is further encoded through a context encoder including a word embedding layer to extract global high-dimensional semantic features of each word in the PBOC credit report, thereby obtaining multiple word context semantic feature vectors.

[0083] Specifically, first, the PBOC credit report is segmented to facilitate semantic confusion during subsequent semantic feature extraction; then, each word after the segmentation process is embedded using a word embedding layer to map it into a word embedding vector, and the context encoding converter is used to perform global context semantic encoding on the sequence of the obtained word embedding vectors to generate the multiple word context semantic feature vectors.

[0084] Accordingly, in a specific example, Figure 4 As shown, the process of passing the PBOC credit report through a context encoder comprising a word embedding layer to obtain a plurality of word context semantic feature vectors includes: S121, performing word segmentation processing on the PBOC credit report to convert the PBOC credit report into a word sequence composed of a plurality of words; S122, using the embedding layer of the context encoder to map each word in the word sequence to a word vector to obtain a sequence of word vectors; and, S123, using the context encoder to perform global context semantic encoding on the sequence of word vectors to obtain the plurality of word context semantic feature vectors.

[0085] More specifically, in step S130, the multiple word context semantic feature vectors are arranged into a one-dimensional feature vector and then extracted through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector. Considering that the context semantic association features of each word in the PBOC credit report of the user to be detected are not able to perform context semantic understanding very well, this is because the PBOC credit report also contains phrases composed of each word, which play an important role in the semantic understanding of the PBOC credit report of the user to be detected. Therefore, in the technical solution of the present application, the multiple word context semantic feature vectors are further arranged into a one-dimensional feature vector and then extracted through a phrase granularity understanding model based on a one-dimensional convolutional neural network to extract the local associated phrase feature information of the semantic understanding features of each word, thereby obtaining a first-scale report understanding feature vector.

[0086] Accordingly, in a specific example, Figure 5 As shown, the method of arranging the multiple word context semantic feature vectors into one-dimensional feature vectors and then passing them through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector includes: using each layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network to perform the following on the input data in the forward pass of the layer: S131, convolution processing on the input data to obtain a convolution feature vector; S132, pooling processing on the convolution feature vector to obtain a pooled feature vector; and, S133, nonlinear activation on the pooled feature vector to obtain an activation feature vector; wherein the output of the last layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the first-scale report understanding feature vector, and the input of the first layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the one-dimensional feature vector.

[0087] In an embodiment of the present application, when a one-dimensional convolution kernel performs one-dimensional convolution encoding, the length of the one-dimensional convolution kernel corresponds to multiple words. Therefore, the essence of one-dimensional convolution encoding is to capture the contextual semantic association features between multiple local words of the People's Bank of China credit report. Therefore, the convolutional neural network model containing a one-dimensional convolution kernel is defined as a phrase granularity understanding model, that is, the above-mentioned phrase granularity understanding model is a convolutional neural network model containing a one-dimensional convolution kernel or a one-dimensional convolutional neural network model, whose input data is the one-dimensional feature vector, and the output data is the first-scale report understanding feature vector. It can be understood by those skilled in the art that the principle and network structure of the convolutional neural network model containing a one-dimensional convolution kernel belong to the well-known technology in the field and will not be repeated here.

[0088] More specifically, in step S140, the multiple word context semantic feature vectors are input into a bidirectional LSTM model to obtain a second-scale report comprehension feature vector. It should be understood that, considering the distribution of the context semantic comprehension features of each word, the phrase features composed of each word possess not only local short-range dependency correlations but also long-range dependency correlation feature distribution information. In other words, the context semantic features of each word in the PBOC credit report of the user to be tested possess long-range dependency semantic correlations, which enable more accurate semantic understanding of the PBOC credit information. Therefore, in the technical solution of the present application, the multiple word context semantic feature vectors are further input into a bidirectional LSTM model to obtain a second-scale report comprehension feature vector. It should be understood that the bidirectional long short-term memory (LSTM) neural network model, by adding input gates, output gates, and forget gates, enables the neural network weights to self-update. While the network model parameters are fixed, the weight scales of different channels can be dynamically changed, thereby avoiding the problems of vanishing gradients or gradient inflation. In particular, the bidirectional long short-term memory neural network model is composed of a forward LSTM and a backward LSTM. The forward LSTM can learn the previous text information of the current word, and the backward LSTM can learn the subsequent text information of the current word. Therefore, the second-scale report understanding feature vector obtained by the bidirectional long short-term memory neural network model learns the semantic information of the global context.

[0089] More specifically, in step S150 , the multiple word context semantic feature vectors are concatenated to obtain a third-scale report comprehension feature vector.

[0090] More specifically, in step S160, the first to third scaled report comprehension feature vectors are fused to obtain a classification feature vector. By fusing the first to third scaled report comprehension feature vectors, the semantic comprehension feature distribution information of multi-scale word associations in the PBOC credit report of the user to be tested is integrated, and this is used as a classification feature vector for classification processing in a classifier to obtain a classification result indicating whether the user to be tested is a repurchase user.

[0091] Accordingly, in a specific example, fusing the first to third scale report understanding feature vectors to obtain a classification feature vector includes: fusing the first to third scale report understanding feature vectors to obtain a classification feature vector using the following formula; wherein the formula is:

[0092]

[0093] in, Represents the first to third scale reported understanding feature vector, Represents a cascade function, represents the classification feature vector.

[0094] More specifically, in step S170, the classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the user to be detected is a compensation repurchase user.

[0095] Accordingly, in a specific example, Figure 6 As shown, the classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the user to be detected is a compensation repurchase user, including: S171, using the fully connected layer of the classifier to fully connect encode the classification feature vector to obtain an encoded classification feature vector; S172, inputting the encoded classification feature vector into the Softmax classification function of the classifier to obtain the classification result.

[0096] Accordingly, in a specific example, the PBOC credit information processing method based on the LSTM model further includes a training step: training the context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier.

[0097] Accordingly, in a specific example, Figure 7As shown, the training step includes: S210, obtaining training data, the training data including the training PBOC credit report of the user to be tested, and the true value of whether the user to be tested is a compensation repurchase user; S220, passing the training PBOC credit report through the context encoder including the word embedding layer to obtain multiple training word context semantic feature vectors; S230, arranging the multiple training word context semantic feature vectors into training one-dimensional feature vectors and then passing them through the phrase granularity understanding model based on the one-dimensional convolutional neural network to obtain a first training scale report understanding feature vector; S240, inputting the multiple training word context semantic feature vectors into the bidirectional LSTM model to obtain a second training scale report understanding feature vector; S250, passing the multiple training word context semantic feature vectors into the bidirectional LSTM model to obtain a second training scale report understanding feature vector; The context semantic feature vectors of the training words are cascaded to obtain a third training scale report understanding feature vector; S260, the first to third training scale report understanding feature vectors are fused to obtain a training classification feature vector; S270, the training classification feature vector is passed through the classifier to obtain a classification loss function value; S280, based on the classification loss function value and through the direction propagation of gradient descent, the context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier are trained, wherein, in each round of iteration of the training, the free label optimization factor of the training classification feature vector based on the soft similarity across classifiers is calculated to iterate the training classification feature vector.

[0098] In particular, in the technical solution of the present application, since the first to third scale report understanding feature vectors respectively represent the contextual association features of word semantics at different scales, it is preferred to obtain the classification feature vector by directly cascading the first to third scale report understanding feature vectors, so that the classification feature vector fully expresses the short-range-long-range word semantic context information in the People's Bank of China credit report. However, this will cause the overall distribution of the classification feature vector to have a strong discreteness, making the training of the classifier difficult, especially the convergence of the classifier's label value difficult. Therefore, the applicant of the present application uses soft label learning instead of the commonly used hard label learning. Specifically, at each iteration, the free label optimization factor of the classification feature vector based on cross-classifier soft similarity is calculated.

[0099] Accordingly, in a specific example, in each iteration of the training, the free label optimization factor based on cross-classifier soft similarity of the training classification feature vector is calculated as the label value of the classifier using the following formula;

[0100] Wherein, the formula is:

[0101]

[0102] is the classification feature vector, is the weight matrix of the classifier for the training classification feature vector, and represent tensor multiplication and tensor addition respectively, represents the distance between vectors, represents the two-norm of the vector, and and is the weight hyperparameter, represents an exponential operation of a vector, wherein the exponential operation of a vector represents calculating the value of a natural exponential function raised to the power of the eigenvalue of each position in the vector.

[0103] Here, the free label optimization factor based on cross-classifier soft similarity simulates pseudo-categories based on the classifier weight matrix through cross-classifier soft similarity of the feature vector and the weight matrix by performing bidirectional clustering before calculating the classification probability of the feature vector with the hard label value. This avoids the classification quantization loss caused by hard label learning through soft similarity learning, and achieves free label optimization that focuses more on the intrinsic weight structure of the classifier, thereby optimizing the training of the classifier label value and improving the training speed of the classifier. In this way, the semantic understanding of the credit information of the People's Bank of China can be more accurately performed, so as to more accurately perform intelligent detection and judgment on whether the user to be detected is a compensation repurchase user.

[0104] In summary, the PBOC credit information processing method based on the LSTM model according to the embodiment of the present application first passes the PBOC credit report of the user to be detected through a context encoder to obtain a plurality of word context semantic feature vectors, then arranges the plurality of word context semantic feature vectors into a one-dimensional feature vector and passes them through a phrase granularity understanding model to obtain a first-scale report understanding feature vector, then inputs the plurality of word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report understanding feature vector, then cascades the plurality of word context semantic feature vectors to obtain a third-scale report understanding feature vector, and finally, passes the classification feature vector obtained by fusing the first to third-scale report understanding feature vectors through a classifier to obtain a classification result indicating whether the user to be detected is a compensation repurchase user. In this way, it is possible to accurately detect and judge whether the user to be detected is a compensation repurchase user.

[0105] Exemplary Systems

[0106] Figure 8 FIG is a block diagram of a PBOC credit information processing system 100 based on an LSTM model according to an embodiment of the present application. Figure 8As shown, according to an embodiment of the present application, the PBOC credit information processing system 100 based on the LSTM model includes: a data acquisition module 110 for acquiring the PBOC credit report of the user to be detected; a context encoding module 120 for passing the PBOC credit report through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors; a first scale encoding module 130 for arranging the multiple word context semantic feature vectors into one-dimensional feature vectors and then passing them through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first scale report understanding feature vector; a second scale encoding module 140 for inputting the multiple word context semantic feature vectors into a bidirectional LSTM model to obtain a second scale report understanding feature vector; a cascade module 150 for cascading the multiple word context semantic feature vectors to obtain a third scale report understanding feature vector; a fusion module 160 for fusing the first to third scale report understanding feature vectors to obtain a classification feature vector; and a classification result generation module 170 for passing the classification feature vector through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the user to be detected is a compensation repurchase user.

[0107] In one example, in the above-mentioned PBOC credit information processing system 100 based on the LSTM model, the context encoding module 120 includes: a word segmentation processing unit, which is used to perform word segmentation processing on the PBOC credit report to convert the PBOC credit report into a word sequence composed of multiple words; an embedding encoding unit, which is used to use the embedding layer of the context encoder to map each word in the word sequence to a word vector to obtain a sequence of word vectors; and a context encoding unit, which is used to use the context encoder to perform global context semantic encoding on the sequence of word vectors to obtain the multiple word context semantic feature vectors.

[0108] In one example, in the above-mentioned PBOC credit information processing system 100 based on the LSTM model, the first-scale encoding module 130 is used to: use each layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network to respectively perform the following on the input data in the forward pass of the layer: convolution processing on the input data to obtain a convolution feature vector; pooling processing on the convolution feature vector to obtain a pooled feature vector; and, nonlinear activation on the pooled feature vector to obtain an activation feature vector; wherein the output of the last layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the first-scale report understanding feature vector, and the input of the first layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the one-dimensional feature vector.

[0109] In one example, in the PBOC credit information processing system 100 based on the LSTM model, the fusion module 160 is further configured to fuse the first to third scale report comprehension feature vectors using the following formula to obtain a classification feature vector; wherein the formula is:

[0110]

[0111] in, Represents the first to third scale reported understanding feature vector, Represents a cascade function, represents the classification feature vector.

[0112] In one example, in the above-mentioned PBOC credit information processing system 100 based on the LSTM model, the classification result generation module 170 is further used to: use the fully connected layer of the classifier to fully connect the classification feature vector to obtain an encoded classification feature vector; input the encoded classification feature vector into the Softmax classification function of the classifier to obtain the classification result.

[0113] In one example, the PBOC credit information processing system 100 based on the LSTM model further includes a training module for training the context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier.

[0114] In one example, in the above-mentioned PBOC credit information processing system 100 based on the LSTM model, the training module includes: a training data acquisition module for acquiring training data, wherein the training data includes a training PBOC credit report of a user to be detected, and a true value of whether the user to be detected is a compensation repurchase user; a training context encoding module for passing the training PBOC credit report through the context encoder including the word embedding layer to obtain a plurality of training word context semantic feature vectors; a training first scale encoding module for arranging the plurality of training word context semantic feature vectors into training one-dimensional feature vectors and then passing them through the phrase granularity understanding model based on the one-dimensional convolutional neural network to obtain a first training scale report understanding feature vector; a training second scale encoding module for inputting the plurality of training word context semantic feature vectors into the bidirectional LSTM model to obtain a second training scale report understanding feature vector. A scale report understanding feature vector; a training cascade module for cascading the multiple training word context semantic feature vectors to obtain a third training scale report understanding feature vector; a training fusion module for fusing the first to third training scale report understanding feature vectors to obtain a training classification feature vector; a classification loss function value calculation module for passing the training classification feature vector through the classifier to obtain a classification loss function value; a training module for training the context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier based on the classification loss function value and propagating in the direction of gradient descent, wherein in each round of the training iteration, the free label optimization factor of the training classification feature vector based on the soft similarity across classifiers is calculated to iterate the training classification feature vector.

[0115] In one example, in the PBOC credit information processing system 100 based on the LSTM model, in each iteration of the training, the free label optimization factor based on cross-classifier soft similarity of the training classification feature vector is calculated as the label value of the classifier using the following formula;

[0116] Wherein, the formula is:

[0117]

[0118] is the classification feature vector, is the weight matrix of the classifier for the training classification feature vector, and represent tensor multiplication and tensor addition respectively, represents the distance between vectors, represents the two-norm of the vector, and and is the weight hyperparameter, represents an exponential operation of a vector, wherein the exponential operation of a vector represents calculating the value of a natural exponential function raised to the power of the eigenvalue of each position in the vector.

[0119] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the LSTM model-based credit information processing system 100 have been described in detail above. Figures 1 to 7 The description of the PBOC credit information processing method based on the LSTM model has been introduced in detail, and therefore, its repeated description will be omitted.

[0120] As described above, the PBOC credit information processing system 100 based on the LSTM model according to the embodiment of the present application can be implemented in various wireless terminals, such as a server of the PBOC credit information processing algorithm based on the LSTM model. In one example, the PBOC credit information processing system 100 based on the LSTM model according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the PBOC credit information processing system 100 based on the LSTM model can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the PBOC credit information processing system 100 based on the LSTM model can also be one of the many hardware modules of the wireless terminal.

[0121] Alternatively, in another example, the PBOC credit information processing system 100 based on the LSTM model and the wireless terminal may also be separate devices, and the PBOC credit information processing system 100 based on the LSTM model may be connected to the wireless terminal through a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0122] The basic principles of the present application have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in this application are merely illustrative and not restrictive, and it should not be assumed that these advantages, strengths, and effects are required of each embodiment of this application. In addition, the specific details disclosed above are merely illustrative and facilitating understanding, and are not restrictive. The above details do not limit this application to necessarily being implemented using the above specific details.

[0123] The block diagrams of the devices, devices, equipment, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As will be appreciated by those skilled in the art, these devices, devices, equipment, and systems can be connected, arranged, or configured in any manner. Words such as "include," "comprise," "have," and the like are open-ended words, meaning "including but not limited to," and can be used interchangeably therewith. The words "or" and "and" used herein refer to the words "and / or" and can be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to," and can be used interchangeably therewith.

[0124] It should also be noted that in the apparatus, device, and method of the present application, each component or each step can be decomposed and / or recombined, and such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0125] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0126] The above description has been provided for the purpose of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although a number of example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for processing credit information of the People's Bank of China based on the LSTM model, characterized in that: include: Obtain the People's Bank of China credit report of the user to be tested; Passing the PBOC credit report through a context encoder including a word embedding layer to obtain multiple word context semantic feature vectors; Arranging the multiple word context semantic feature vectors into a one-dimensional feature vector and then applying the result to a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first-scale report understanding feature vector; Inputting the plurality of word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report comprehension feature vector; cascading the plurality of word context semantic feature vectors to obtain a third scale report comprehension feature vector; fusing the first to third scale report comprehension feature vectors to obtain a classification feature vector; as well as The classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the user to be detected is a compensation repurchase user.

2. The LSTM model-based credit information processing method of the People's Bank of China according to claim 1 is characterized in that: The step of passing the PBOC credit report through a context encoder including a word embedding layer to obtain a plurality of word context semantic feature vectors includes: Performing word segmentation processing on the PBOC credit report to convert the PBOC credit report into a word sequence consisting of multiple words; Mapping each word in the word sequence to a word vector using the embedding layer of the context encoder to obtain a sequence of word vectors; and The context encoder is used to perform global context semantic encoding on the sequence of word vectors to obtain the multiple word context semantic feature vectors.

3. The LSTM model-based credit information processing method of the People's Bank of China according to claim 2 is characterized in that: The step of arranging the plurality of word context semantic feature vectors into a one-dimensional feature vector and then applying the one-dimensional convolutional neural network-based phrase granularity understanding model to obtain a first-scale report understanding feature vector comprises: using each layer of the one-dimensional convolutional neural network-based phrase granularity understanding model to perform the following on the input data in the forward pass of the layer: Performing convolution processing on the input data to obtain a convolution feature vector; Performing pooling processing on the convolution feature vector to obtain a pooled feature vector; and Performing nonlinear activation on the pooled feature vector to obtain an activated feature vector; The output of the last layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the first-scale report understanding feature vector, and the input of the first layer of the phrase granularity understanding model based on the one-dimensional convolutional neural network is the one-dimensional feature vector.

4. The LSTM model-based credit information processing method of the People's Bank of China according to claim 3 is characterized in that: The fusing the first to third scale report comprehension feature vectors to obtain a classification feature vector comprises: fusing the first to third scale report comprehension feature vectors to obtain a classification feature vector using the following formula; Wherein, the formula is: in, Represents the first to third scale reported understanding feature vector, Represents a cascade function, represents the classification feature vector.

5. The PBOC credit information processing method based on the LSTM model according to claim 4 is characterized in that: The classification feature vector is passed through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the user to be detected is a compensation repurchase user, including: Performing full connection encoding on the classification feature vector using the fully connected layer of the classifier to obtain an encoded classification feature vector; The encoded classification feature vector is input into the Softmax classification function of the classifier to obtain the classification result.

6. The PBOC credit information processing method based on the LSTM model according to claim 5 is characterized in that: It also includes a training step: training the context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier.

7. The PBOC credit information processing method based on the LSTM model according to claim 6 is characterized in that: The training step comprises: Acquire training data, the training data including a training PBOC credit report of the user to be tested, and a true value of whether the user to be tested is a compensation repurchase user; Passing the training PBOC credit report through the context encoder including the word embedding layer to obtain a plurality of training word context semantic feature vectors; Arranging the plurality of training word context semantic feature vectors into a training one-dimensional feature vector and then applying the one-dimensional convolutional neural network-based phrase granularity understanding model to obtain a first training scale report understanding feature vector; Inputting the plurality of training word context semantic feature vectors into the bidirectional LSTM model to obtain a second training scale report comprehension feature vector; cascading the plurality of training word context semantic feature vectors to obtain a third training scale report comprehension feature vector; fusing the first to third training scale report comprehension feature vectors to obtain a training classification feature vector; Passing the training classification feature vector through the classifier to obtain a classification loss function value; The context encoder including the word embedding layer, the phrase granularity understanding model based on the one-dimensional convolutional neural network, the bidirectional LSTM model and the classifier are trained based on the classification loss function value and through the direction propagation of gradient descent, wherein in each round of iteration of the training, the free label optimization factor based on the cross-classifier soft similarity of the training classification feature vector is calculated to iterate the training classification feature vector.

8. The PBOC credit information processing method based on the LSTM model according to claim 7 is characterized in that: In each iteration of the training, the free label optimization factor based on cross-classifier soft similarity of the training classification feature vector is calculated as the label value of the classifier using the following formula; Wherein, the formula is: is the classification feature vector, is the weight matrix of the classifier for the training classification feature vector, and represent tensor multiplication and tensor addition respectively, represents the distance between vectors, represents the two-norm of the vector, and and is the weight hyperparameter, represents an exponential operation of a vector, wherein the exponential operation of a vector represents calculating the value of a natural exponential function raised to the power of the eigenvalue of each position in the vector.

9. A PBOC credit information processing system based on LSTM model, characterized in that: include: A data acquisition module, used to obtain the People's Bank of China credit report of the user to be tested; A context encoding module, configured to pass the PBOC credit report through a context encoder comprising a word embedding layer to obtain a plurality of word context semantic feature vectors; A first scale encoding module, configured to arrange the plurality of word context semantic feature vectors into a one-dimensional feature vector and then pass the result through a phrase granularity understanding model based on a one-dimensional convolutional neural network to obtain a first scale report understanding feature vector; a second-scale encoding module, configured to input the plurality of word context semantic feature vectors into a bidirectional LSTM model to obtain a second-scale report comprehension feature vector; a concatenation module, configured to concatenate the plurality of word context semantic feature vectors to obtain a third-scale report comprehension feature vector; A fusion module, configured to fuse the first to third scale report comprehension feature vectors to obtain a classification feature vector; as well as The classification result generation module is used to pass the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the user to be detected is a compensation repurchase user.

10. The PBOC credit information processing system based on the LSTM model according to claim 9, characterized in that: The context encoding module includes: A word segmentation processing unit, configured to perform word segmentation processing on the PBOC credit report to convert the PBOC credit report into a word sequence consisting of a plurality of words; an embedding encoding unit, configured to map each word in the word sequence to a word vector using the embedding layer of the context encoder to obtain a sequence of word vectors; and A context encoding unit is used to use the context encoder to perform global context semantic encoding on the sequence of word vectors to obtain the multiple word context semantic feature vectors.

Images