An intelligent decision-making method and system for bank asset preservation business

Abstract

The present invention relates to the technical field of intelligent management decision-making of bank assets, and specifically discloses an intelligent decision-making method and system for bank asset preservation business, which uses natural language processing technology based on deep learning to perform sentence-granularity semantic encoding on the text content of a litigation application initiated by a customer manager, thereby mining the semantic feature expressions in the litigation application to capture the potential risk information contained in the litigation application, thereby intelligently judging whether to preserve the entire asset in a preliminary review, and being able to realize automated decision-making on the preliminary review of the entire asset preservation system, thereby improving the efficiency of bank asset preservation processing.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent management decision-making for bank assets, and more specifically, to an intelligent decision-making method and system for bank asset preservation business. Background Art

[0002] With the increasing complexity and competitiveness of the financial market, bank asset preservation business has become increasingly important. Bank asset preservation business refers to various business activities that take various preventive, protective and recovery measures for existing or upcoming non-performing assets in accordance with the law to reduce losses and protect bank assets. However, the existing asset preservation business management work usually requires manual participation and mainly relies on manual review and processing. This method is not only inefficient, but also prone to misjudgment due to human factors. With the rapid development of banking business, the gradual strengthening of supervision, and the refinement of management, the existing system has shown a certain lag and is gradually unable to meet the development and management requirements of bank asset preservation business.

[0003] Therefore, in order to strengthen the standardized management of bank asset preservation business, speed up the collection and resolution of non-performing assets, and prevent new risks from arising in the process of asset preservation, we look forward to an intelligent decision-making method and system for bank asset preservation business. Summary of the invention

[0004] In order to solve the above technical problems, this application is proposed. The embodiment of this application provides an intelligent decision-making method and system for bank asset preservation business, which uses natural language processing technology based on deep learning to perform sentence-based semantic encoding on the text content of the litigation application initiated by the account manager, and mines the semantic feature expression in the litigation application to capture the potential risk information contained in the litigation application, so as to intelligently judge whether to preserve the entire asset for preliminary review. In this way, it is possible to achieve automated decision-making on the preliminary review of the asset preservation department and improve the efficiency of bank asset preservation processing.

[0005] Accordingly, according to one aspect of the present application, there is provided a method for intelligent decision-making of bank asset preservation business, which includes:

[0006] Obtain litigation applications initiated by account managers;

[0007] Semantically encoding the text portion of the litigation application to obtain a sequence of sentence-level semantic understanding feature vectors of the litigation application text;

[0008] Simultaneously mapping each litigation application text sentence granular semantic understanding feature vector in the sequence of litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain a sequence of mapped litigation application text sentence granular semantic understanding feature vectors;

[0009] Based on the overall semantic feature distribution importance of each mapped litigation application text sentence granularity semantic understanding feature vector relative to the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors, attention fusion is performed on the sequence of litigation application text sentence granularity semantic understanding feature vectors to obtain a sentence granularity attention litigation application text understanding feature vector;

[0010] Based on the sentence-granular attention litigation application text understanding feature vector, determine whether to retain the entire preliminary examination.

[0011] In the above-mentioned intelligent decision-making method for bank asset preservation business, the text part in the litigation application is semantically encoded to obtain a sequence of sentence-granularity semantic understanding feature vectors of the litigation application text, including: extracting the text part in the litigation application to obtain a litigation application text description; performing sentence-granularity semantic encoding on the litigation application text description to obtain a sequence of sentence-granularity semantic understanding feature vectors of the litigation application text.

[0012] In the above-mentioned intelligent decision-making method for bank asset preservation business, sentence-granularity semantic encoding is performed on the text description of the litigation application to obtain a sequence of sentence-granularity semantic understanding feature vectors of the litigation application text, including: sentence-segmenting the text description of the litigation application and then passing it through a semantic encoder including a sentence encoder to obtain a sequence of sentence-granularity semantic understanding feature vectors of the litigation application text.

[0013] In the above-mentioned intelligent decision-making method for bank asset preservation business, each litigation application text sentence granularity semantic understanding feature vector in the sequence of litigation application text sentence granularity semantic understanding feature vectors is simultaneously mapped to the same attention space to obtain a sequence of litigation application text sentence granularity semantic understanding feature vectors after mapping, including: inputting the sequence of litigation application text sentence granularity semantic understanding feature vectors into an attention space mapper based on a fully connected layer to simultaneously map each litigation application text sentence granularity semantic understanding feature vector to the same attention space to obtain a sequence of litigation application text sentence granularity semantic understanding feature vectors after mapping.

[0014] In the above-mentioned intelligent decision-making method for bank asset preservation business, based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence granularity semantic understanding feature vectors relative to the sequence of the mapped litigation application text sentence granularity semantic understanding feature vectors, the sequence of the litigation application text sentence granularity semantic understanding feature vectors is subjected to attention fusion to obtain the sentence granularity attention litigation application text understanding feature vector, including: calculating the overall attention weight of the feature distribution of each of the mapped litigation application text sentence granularity semantic understanding feature vectors in the sequence of the mapped litigation application text sentence granularity semantic understanding feature vectors relative to the sequence of the mapped litigation application text sentence granularity semantic understanding feature vectors to obtain a sequence of attention weights; based on the sequence of attention weights, fusing the sequence of the litigation application text sentence granularity semantic understanding feature vectors to obtain the sentence granularity attention litigation application text understanding feature vector.

[0015] In the above-mentioned intelligent decision-making method for bank asset preservation business, the attention weight of each mapped litigation application text sentence granularity semantic understanding feature vector in the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors relative to the feature distribution of the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors is calculated to obtain a sequence of attention weights, including: calculating the attention weight of each mapped litigation application text sentence granularity semantic understanding feature vector in the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors relative to the feature distribution of the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors using the following attention weight calculation formula to obtain the sequence of attention weights; wherein, the attention weight calculation formula is:

[0016]

[0017] Among them, h i is the i-th mapped litigation application text sentence granular semantic understanding feature vector in the sequence of the mapped litigation application text sentence granular semantic understanding feature vector, and A is 1×N w The matrix, N w is the dimension of the sentence-level semantic understanding feature vector of the i-th mapped litigation application text, and B is 1×N h The matrix, N h is the dimension of each mapped litigation application text sentence granular semantic understanding feature vector, N is the number of mapped litigation application text sentence granular semantic understanding feature vectors, σ is the Sigmoid function, S i is the i-th attention weight value.

[0018] In the above-mentioned intelligent decision-making method for bank asset preservation business, whether to preserve all assets in the preliminary review is determined based on the sentence-granularity attention litigation application text understanding feature vector, including: passing the sentence-granularity attention litigation application text understanding feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether to preserve all assets in the preliminary review.

[0019] The above-mentioned intelligent decision-making method for bank asset preservation business also includes a training step: training the semantic encoder including the sentence encoder, the attention space mapper based on the fully connected layer and the classifier.

[0020] In the above-mentioned intelligent decision-making method for bank asset preservation business, the training step includes: obtaining training data, the training data including training litigation applications initiated by customer managers and the true value of the initial review of whether to preserve the entire litigation application; extracting the text part of the training litigation application to obtain a text description of the training litigation application; after sentence processing of the training litigation application text description, passing it through the semantic encoder including the sentence encoder to obtain a sequence of sentence-granularity semantic understanding feature vectors of the training litigation application text; inputting the sequence of sentence-granularity semantic understanding feature vectors of the training litigation application text into the attention space mapper based on the fully connected layer to simultaneously map each of the sentence-granularity semantic understanding feature vectors of the training litigation application text to the same attention space to obtain a sequence of sentence-granularity semantic understanding feature vectors of the litigation application text after mapping; calculating the sentence-granularity semantic understanding feature vectors of the litigation application text after mapping The attention weight of the whole feature distribution of the sentence granularity semantic understanding feature vector of the litigation application text relative to the sequence of the sentence granularity semantic understanding feature vector of the litigation application text after the training mapping is obtained to obtain a sequence of training attention weights; based on the sequence of training attention weights, the sequence of the training sentence granularity semantic understanding feature vector of the litigation application text is fused to obtain the training sentence granularity attention litigation application text understanding feature vector; the training sentence granularity attention litigation application text understanding feature vector is passed through the classifier to obtain the classification loss function value; the specific loss function value of the training sentence granularity attention litigation application text understanding feature vector and the sequence of the training sentence granularity semantic understanding feature vector of the litigation application text is calculated; the weighted sum of the classification loss function value and the specific loss function value is used as the loss function value to train the semantic encoder including the sentence encoder, the attention space mapper based on the fully connected layer and the classifier.

[0021] According to another aspect of the present application, there is provided an intelligent decision-making system for bank asset preservation business, which includes:

[0022] The litigation application acquisition module is used to obtain the litigation application initiated by the account manager;

[0023] A litigation application semantic understanding module, used for semantically encoding the text portion of the litigation application to obtain a sequence of semantic understanding feature vectors of the litigation application text sentence granularity;

[0024] A same attention space mapping module is used to simultaneously map each litigation application text sentence granular semantic understanding feature vector in the sequence of litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain a sequence of litigation application text sentence granular semantic understanding feature vectors after mapping;

[0025] An attention fusion module, used for performing attention fusion on the sequence of sentence granularity semantic understanding feature vectors of the litigation application text based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence granularity semantic understanding feature vectors relative to the sequence of the mapped litigation application text sentence granularity semantic understanding feature vectors to obtain a sentence granularity attention litigation application text understanding feature vector;

[0026] A management result generation module is used to determine whether to retain the entire preliminary examination based on the sentence-granular attention litigation application text understanding feature vector.

[0027] Compared with the prior art, the intelligent decision-making method and system for bank asset preservation business provided by the present application adopts natural language processing technology based on deep learning to perform sentence-granularity semantic encoding on the text content of the litigation application initiated by the account manager, thereby mining the semantic feature expressions in the litigation application to capture the potential risk information contained in the litigation application, thereby intelligently judging whether to preserve the entire asset for preliminary review, and being able to realize automated decision-making on the preliminary review of the entire asset preservation system, thereby improving the efficiency of bank asset preservation processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used 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 accompanying drawings, the same reference numerals generally represent the same components or steps.

[0029] Figure 1 It is a flow chart of the intelligent decision-making method for bank asset preservation business according to the embodiment of the present application.

[0030] Figure 2 Schematic diagram of the architecture of the intelligent decision-making method for bank asset preservation business according to an embodiment of the present application.

[0031] Figure 3The present invention is a flowchart of a sequence of semantic encoding of the text portion of the litigation application in the intelligent decision-making method for bank asset preservation business according to an embodiment of the present application to obtain a sentence-granular semantic understanding feature vector of the litigation application text.

[0032] Figure 4 A flowchart for performing attention fusion on the sequence of sentence-granularity semantic understanding feature vectors of the mapped litigation application text based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence-granularity semantic understanding feature vectors relative to the sequence of mapped litigation application text sentence-granularity semantic understanding feature vectors in the intelligent decision-making method for bank asset preservation business according to an embodiment of the present application to obtain a sentence-granularity attention litigation application text understanding feature vector.

[0033] Figure 5 It is a flowchart of the training steps in the intelligent decision-making method for bank asset preservation business according to the embodiment of the present application.

[0034] Figure 6 It is a block diagram of the intelligent decision-making system for bank asset preservation business according to the embodiment of the present application. DETAILED DESCRIPTION

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

[0036] The approval process for litigation execution plans is usually as follows: first, the account manager initiates a litigation application; then, the branch president reviews it. If the branch is directly affiliated to the head office, it will go directly to the insurance department for initial review. If the branch is not directly affiliated to the head office, it will go to the branch risk department for review; then, the insurance department will review it again. If it is not within the authority, it will need to go through the secretary position, risk committee director review, and president approval stages. If it is within the authority, it can directly proceed to the approval registration / issuance step.

[0037] As mentioned in the above background technology, the existing asset preservation business management work usually requires manual participation. With the rapid development of banking business, the gradual strengthening of supervision, and the improvement of refined management, it has shown a certain lag and is gradually unable to meet the development and management requirements of the bank's asset preservation business.

[0038] In response to the above technical problems, the technical concept of this application is to use natural language processing technology based on deep learning to perform sentence-based semantic encoding on the text content of the litigation application initiated by the account manager, thereby mining the semantic feature expressions in the litigation application to capture the potential risk information contained in the litigation application, and thus intelligently determine whether to preserve all assets in the preliminary review, which can realize automated decision-making on the preliminary review of the asset preservation system and improve the efficiency of the bank's asset preservation processing.

[0039] Figure 1 It is a flow chart of the intelligent decision-making method for bank asset preservation business according to the embodiment of the present application. Figure 2 FIG. 1 is a schematic diagram of the architecture of the intelligent decision-making method for bank asset preservation business according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the embodiment of the present application, the intelligent decision-making method for bank asset preservation business includes the steps of: S110, obtaining a litigation application initiated by a customer manager; S120, semantically encoding the text part of the litigation application to obtain a sequence of sentence-granularity semantic understanding feature vectors of the litigation application text; S130, simultaneously mapping each litigation application text sentence-granularity semantic understanding feature vector in the sequence of sentence-granularity semantic understanding feature vectors of the litigation application text to the same attention space to obtain a sequence of mapped litigation application text sentence-granularity semantic understanding feature vectors; S140, based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence-granularity semantic understanding feature vectors relative to the sequence of mapped litigation application text sentence-granularity semantic understanding feature vectors, performing attention fusion on the sequence of the litigation application text sentence-granularity semantic understanding feature vectors to obtain a sentence-granularity attention litigation application text understanding feature vector; S150, based on the sentence-granularity attention litigation application text understanding feature vector, determining whether to preserve the entire text for preliminary review.

[0040] In the above-mentioned intelligent decision-making method for bank asset preservation business, the step S110 obtains the litigation application initiated by the account manager. It should be understood that in banking business, the account manager is usually responsible for communicating and interacting with customers and understanding their needs and problems. When customers are involved in asset preservation issues, such as debt collection, handling of collateral, etc., the account manager usually files a lawsuit application with the court on behalf of the bank. Usually, the litigation application initiated by the account manager details the bank's requirements and legal basis, as well as all evidence related to the debt, such as the loan agreement, debt amount, repayment period, repayment plan, and the borrower's repayment record. This information provides details about the background of the case, the dispute points, and the required preservation measures. Based on this information, natural language processing technology can be used to semantically encode and analyze the litigation application to achieve the goal of automated decision-making in the initial review of the asset preservation department.

[0041] In the above-mentioned intelligent decision-making method for bank asset preservation business, the step S120 performs semantic encoding on the text part of the litigation application to obtain a sequence of semantic understanding feature vectors of the litigation application text sentence granularity. Specifically, Figure 3 The flowchart is a sequence of semantic coding of the text part in the litigation application to obtain the semantic understanding feature vector of the litigation application text sentence granularity in the intelligent decision-making method for bank asset preservation business according to the embodiment of the present application. Figure 3 As shown, the step S120 includes: S121, extracting the text portion of the litigation application to obtain a text description of the litigation application; S122, performing sentence-granular semantic encoding on the text description of the litigation application to obtain a sequence of sentence-granular semantic understanding feature vectors of the litigation application text.

[0042] Specifically, the step S121 extracts the text portion of the litigation application to obtain a text description of the litigation application. It should be understood that the litigation application is usually a text containing a large amount of text description. In order to extract the key information of the litigation application, the text portion of the litigation application is further extracted to remove irrelevant and redundant format content, so that subsequent processing is more efficient and accurate. In addition, by extracting the text portion of the litigation application, natural language processing technology can be further used to perform semantic analysis on complex text content, so as to perform the subsequent automated decision-making task of preserving the entire preliminary review.

[0043] Specifically, the step S122 performs sentence-granular semantic encoding on the text description of the litigation application to obtain a sequence of sentence-granular semantic understanding feature vectors of the litigation application text. In a specific example of the present application, the encoding method for performing sentence-granular semantic encoding on the text description of the litigation application to obtain a sequence of sentence-granular semantic understanding feature vectors of the litigation application text is to perform sentence processing on the text description of the litigation application and then pass it through a semantic encoder containing a sentence encoder to obtain a sequence of sentence-granular semantic understanding feature vectors of the litigation application text. In the technical solution of the present application, the long litigation application text is divided into multiple independent sentence representations through sentence processing, so that the semantic encoding of each sentence is relatively independent, which helps the model to better handle long texts and complex semantic structures, better capture the logical relationship and contextual information between sentences, and reduce the model's dependence on the entire text sequence. Then, the semantic encoder containing the sentence encoder is further used to semantically encode each sentence separately to capture the semantic features of each sentence and to perform a more detailed semantic understanding of each sentence. It should be understood that the sentence encoder is a neural network model that can convert sentence-level text into a vector representation of a fixed dimension with semantic information. Sentence encoders can be implemented using models such as recurrent neural networks (such as long short-term memory networks LSTM or gated recurrent units GRU) or convolutional neural networks (CNN). These models can map sentences into a low-dimensional semantic space by learning the semantic and contextual information in sentences, thereby capturing the semantic information of litigation applications and providing more accurate and comprehensive input for subsequent automated decision-making.

[0044] In the above-mentioned intelligent decision-making method for bank asset preservation business, the step S130 maps each of the litigation application text sentence granular semantic understanding feature vectors in the sequence of the litigation application text sentence granular semantic understanding feature vectors to the same attention space at the same time to obtain a sequence of the litigation application text sentence granular semantic understanding feature vectors after mapping. In a specific example of the present application, the implementation method of simultaneously mapping each of the litigation application text sentence granular semantic understanding feature vectors in the sequence of the litigation application text sentence granular semantic understanding feature vectors to the same attention space is to input the sequence of the litigation application text sentence granular semantic understanding feature vectors into an attention space mapper based on a fully connected layer to simultaneously map each of the litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain a sequence of the litigation application text sentence granular semantic understanding feature vectors after mapping. A person of ordinary skill in the art should know that a fully connected layer is a common neural network layer in which each neuron is connected to all neurons in the previous layer and is generally used to realize feature interaction and fusion. In the technical solution of the present application, a fully connected layer is used to simultaneously map each of the litigation application text sentence granular semantic understanding feature vectors to the same attention space. It should be understood that the attention mechanism is widely used in natural language processing tasks, which can help the model pay more attention to important parts when processing text. By using the attention mechanism, the semantic features of the sentences can be weighted according to the importance and relevance of different sentences, so as to better capture the semantic information of the litigation application text. Here, through the mapping processing of the attention space mapper based on the fully connected layer, the granular semantic understanding feature vectors of each of the litigation application text sentences can be mapped to the same attention space, so that the attention weights between different sentence features have consistent calculation rules and representation methods. That is, the same attention calculation method can be used to measure the importance and relevance of sentences, more accurately reflect the importance of sentences, and then realize the unified calculation, feature fusion and interaction of sentence features, providing a basis for subsequent attention weight calculation. In addition, the fully connected layer can map the sequence of the input litigation application text sentence granular semantic understanding feature vectors to the appropriate dimension by adjusting the dimension of the weight matrix to ensure that the mapped feature vector matches the subsequent processing steps, facilitating the fusion and interaction of subsequent features.

[0045] In the above-mentioned intelligent decision-making method for bank asset preservation business, the step S140 performs attention fusion on the sequence of the sentence granularity semantic understanding feature vectors of the litigation application text based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence granularity semantic understanding feature vectors relative to the sequence of the mapped litigation application text sentence granularity semantic understanding feature vectors to obtain the sentence granularity attention litigation application text understanding feature vector. Specifically, Figure 4A flowchart of performing attention fusion on the sequence of the sentence granularity semantic understanding feature vectors of the litigation application text based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence granularity semantic understanding feature vectors relative to the sequence of the mapped litigation application text sentence granularity semantic understanding feature vectors in the intelligent decision-making method for bank asset preservation business according to the embodiment of the present application to obtain the sentence granularity attention litigation application text understanding feature vector. Figure 4 As shown, the step S140 includes: S141, calculating the overall attention weight of the feature distribution of each mapped litigation application text sentence granularity semantic understanding feature vector in the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors relative to the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors to obtain a sequence of attention weights; S142, based on the sequence of attention weights, fusing the sequence of litigation application text sentence granularity semantic understanding feature vectors to obtain the sentence granularity attention litigation application text understanding feature vector.

[0046] Specifically, the step S141 calculates the overall attention weight of the feature distribution of each mapped litigation application text sentence granular semantic understanding feature vector in the sequence of the mapped litigation application text sentence granular semantic understanding feature vector relative to the sequence of the mapped litigation application text sentence granular semantic understanding feature vector to obtain a sequence of attention weights. In the technical solution of the present application, the importance and relevance of each sentence in the entire text are determined by calculating the attention weight of each vector in the sequence of the mapped litigation application text sentence granular semantic understanding feature vector relative to the overall feature distribution. Specifically, the attention weight represents the importance of the semantic features of each sentence in the entire text, and the attention weight of each sentence represents its importance relative to the overall feature distribution. Therefore, by dynamically allocating attention weights based on the relevance of each mapped litigation application text sentence granular semantic understanding feature vector relative to the overall semantic feature distribution of the application litigation text, the importance distribution of each sentence can be obtained, and then weighted operations can be performed according to these attention weights in subsequent processing, so that the network model pays more attention to important sentences, thereby improving the understanding and processing capabilities of the litigation application text.

[0047] In a specific example of the present application, the step S141 includes: calculating the overall attention weight of each mapped litigation application text sentence granularity semantic understanding feature vector in the sequence of the mapped litigation application text sentence granularity semantic understanding feature vector relative to the sequence of the mapped litigation application text sentence granularity semantic understanding feature vector using the following attention weight calculation formula to obtain the sequence of attention weights; wherein the attention weight calculation formula is:

[0048]

[0049] Among them, h i is the i-th mapped litigation application text sentence granular semantic understanding feature vector in the sequence of the mapped litigation application text sentence granular semantic understanding feature vector, and A is 1×N w The matrix, N w is the dimension of the sentence-level semantic understanding feature vector of the i-th mapped litigation application text, and B is 1×N h The matrix, N h is the dimension of each mapped litigation application text sentence granular semantic understanding feature vector, N is the number of mapped litigation application text sentence granular semantic understanding feature vectors, σ is the Sigmoid function, S i is the i-th attention weight value.

[0050] Specifically, the step S142, based on the sequence of the attention weights, fuses the sequence of the litigation application text sentence granularity semantic understanding feature vectors to obtain the sentence granularity attention litigation application text understanding feature vector. It should be understood that through the calculation of the above attention weights, the importance distribution of each sentence in the entire text is obtained, and further using the attention weights as weight coefficients, the weighted summation of the litigation application text sentence granularity semantic understanding feature vectors in the sequence of the litigation application text sentence granularity semantic understanding feature vectors can make the semantic features of important sentences get higher weights in the fusion process, thereby strengthening the attention to important sentences, capturing important semantic information in the litigation application text, and having a deeper understanding of the litigation application text, so as to better support the management of bank asset preservation business.

[0051] In the above-mentioned intelligent decision-making method for bank asset preservation business, the step S150 determines whether to perform a preliminary review of the entire asset preservation based on the sentence-granular attention litigation application text understanding feature vector. In a specific example of the present application, the implementation method of determining whether to perform a preliminary review of the entire asset preservation based on the sentence-granular attention litigation application text understanding feature vector is to pass the sentence-granular attention litigation application text understanding feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether to perform a preliminary review of the entire asset preservation. It should be understood that a classifier is a machine learning model that can classify or predict based on input features. In the technical solution of the present application, the classifier is trained into a binary classifier, in which one category indicates that a preliminary review of the entire asset preservation is required, and the other category indicates that a preliminary review of the entire asset preservation is not required. By using a large amount of sample data with known classification results to train the classifier, the classifier learns the mapping relationship between different litigation applications and classification results, and performs classification prediction on unknown samples. Furthermore, the sentence-granular attention litigation application text understanding feature vector is input into the trained classifier, and the feature pattern and mapping law learned by the classifier can be used to determine whether the litigation application needs a preliminary review of the entire asset preservation. If the classification results indicate that the litigation application requires preliminary review of the entire preservation system, the corresponding processes and operations can be triggered to achieve automated decision-making on the preliminary review of the entire asset preservation system, thereby improving processing efficiency and accuracy.

[0052] It should be understood that before using the above neural network model, it is necessary to train the semantic encoder including the sentence encoder, the attention space mapper based on the fully connected layer, and the classifier. That is to say, in the intelligent decision-making method for bank asset preservation business of the present application, it also includes a training step of training the semantic encoder including the sentence encoder, the attention space mapper based on the fully connected layer, and the classifier.

[0053] Figure 5 Flow chart of the training steps in the intelligent decision-making method for bank asset preservation business according to the embodiment of the present application. Figure 5As shown, the training step includes: S210, obtaining training data, the training data including training litigation applications initiated by customer managers, and whether to keep the true value of all preliminary reviews; S220, extracting the text part of the training litigation application to obtain a text description of the training litigation application; S230, after sentence processing of the training litigation application text description, passing it through the semantic encoder including the sentence encoder to obtain a sequence of sentence-granularity semantic understanding feature vectors of the training litigation application text; S240, inputting the sequence of sentence-granularity semantic understanding feature vectors of the training litigation application text into the attention space mapper based on the fully connected layer to simultaneously map each of the sentence-granularity semantic understanding feature vectors of the training litigation application text to the same attention space to obtain a sequence of sentence-granularity semantic understanding feature vectors of the training mapped litigation application text; S250, calculating each training mapped litigation application text in the sequence of sentence-granularity semantic understanding feature vectors of the training mapped litigation application text The sentence-granularity semantic understanding feature vector is relative to the feature distribution of the sequence of sentence-granularity semantic understanding feature vectors of the litigation application text after the training mapping to obtain a sequence of training attention weights; S260, based on the sequence of training attention weights, the sequence of sentence-granularity semantic understanding feature vectors of the training litigation application text is fused to obtain the training sentence-granularity attention litigation application text understanding feature vector; S270, the training sentence-granularity attention litigation application text understanding feature vector is passed through the classifier to obtain a classification loss function value; S280, the specific loss function value of the training sentence-granularity attention litigation application text understanding feature vector and the sequence of the training litigation application text sentence-granularity semantic understanding feature vector is calculated; S290, the weighted sum of the classification loss function value and the specific loss function value is used as the loss function value to train the semantic encoder including the sentence encoder, the attention space mapper based on the fully connected layer and the classifier.

[0054] In the technical solution of the present application, the sequence of sentence-granular semantic understanding feature vectors of the training mapped litigation application text expresses the sentence-based encoded text semantic features of the training litigation application text description, and the overall attention weight of the feature distribution of each training mapped litigation application text sentence-granular semantic understanding feature vector in the sequence of the training mapped litigation application text sentence-granular semantic understanding feature vector is calculated relative to the sequence of the training mapped litigation application text sentence-granular semantic understanding feature vector, and based on the sequence of the training attention weights, the sequence of the training litigation application text sentence-granular semantic understanding feature vectors is fused to obtain the training sentence-granular attention litigation application text understanding feature vector, and then the sentence-based encoded text semantic features can be focused on based on the importance of the overall semantic features of the training litigation application text description. However, this will also cause the training sentence-granular attention litigation application text understanding feature vector to have a different feature group density representation relative to the sequence of the training litigation application text sentence-granular semantic understanding feature vector, so that when the model is trained as a whole, there will be an iterative imbalance between the text semantic feature encoding of the semantic encoder and the fusion of the training attention weights, affecting the overall training efficiency of the model.

[0055] Therefore, in the technical solution of the present application, in order to improve the consistency of the feature group density representation of the training sentence granularity attention litigation application text understanding feature vector relative to the sequence of the training litigation application text sentence granularity semantic understanding feature vector, a specific loss function for the sequence of the training sentence granularity attention litigation application text understanding feature vector and the training litigation application text sentence granularity semantic understanding feature vector is further introduced.

[0056] Specifically, the step S280 includes: calculating a specific loss function value of a sequence of the training sentence granular attention litigation application text understanding feature vector and the training litigation application text sentence granular semantic understanding feature vector using the following specific loss function, wherein the specific loss function is expressed as:

[0057]

[0058] V1 is the training sentence granular attention litigation application text understanding feature vector, V2 is the second feature vector after the sequence cascade of the training litigation application text sentence granular semantic understanding feature vector, v 1i and v 2i are the eigenvalues ​​of the training sentence granular attention litigation application text understanding feature vector V1 and the second feature vector V2, L is the length of the feature vector, and represents the square of the vector's two-norm, is the specific loss function value.

[0059] Here, the specific loss function performs group counting attention based on feature group density, and uses group count as a recursive mapping of the output feature group density to perform adaptive attention to different density representation patterns between the sequence of the sentence-granular attention litigation application text understanding feature vector and the sequence of the litigation application text sentence-granular semantic understanding feature vector. By using it as a specific loss function to train the model, the model can avoid over-estimation and under-estimation for different density patterns under the feature distribution of the sequence of the sentence-granular attention litigation application text understanding feature vector and the sequence of the litigation application text sentence-granular semantic understanding feature vector, and learn the correspondence between the feature value distribution and the group density distribution, thereby achieving the optimization of the feature group density representation consistency between the sequence of the sentence-granular attention litigation application text understanding feature vector and the sequence of the litigation application text sentence-granular semantic understanding feature vector with different feature densities, so as to improve the overall training efficiency of the model.

[0060] In summary, the intelligent decision-making method for the bank asset preservation business according to the embodiment of the present application is explained, which uses natural language processing technology based on deep learning to perform sentence-based semantic encoding on the text content of the litigation application initiated by the account manager, thereby mining the semantic feature expressions in the litigation application to capture the potential risk information contained in the litigation application, thereby intelligently judging whether to preserve the entire asset for preliminary review. In this way, it is possible to achieve automated decision-making on the preliminary review of the asset preservation department and improve the efficiency of the bank's asset preservation processing.

[0061] Figure 6 FIG. 1 is a block diagram of an intelligent decision-making system for bank asset preservation services according to an embodiment of the present application. Figure 6 As shown, according to the embodiment of the present application, the intelligent decision-making system 100 for bank asset preservation business includes: a litigation application acquisition module 110, which is used to acquire the litigation application initiated by the customer manager; a litigation application semantic understanding module 120, which is used to semantically encode the text part in the litigation application to obtain a sequence of sentence-granularity semantic understanding feature vectors of the litigation application text; a same attention space mapping module 130, which is used to simultaneously map each litigation application text sentence-granularity semantic understanding feature vector in the sequence of the litigation application text sentence-granularity semantic understanding feature vector to the same attention space to obtain a sequence of mapped litigation application text sentence-granularity semantic understanding feature vectors; an attention fusion module 140, which is used to perform attention fusion on the sequence of the litigation application text sentence-granularity semantic understanding feature vectors based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence-granularity semantic understanding feature vectors relative to the sequence of the mapped litigation application text sentence-granularity semantic understanding feature vectors to obtain a sentence-granularity attention litigation application text understanding feature vector; a management result generation module 150, which is used to determine whether to preserve the entire preliminary review based on the sentence-granularity attention litigation application text understanding feature vector.

[0062] Here, those skilled in the art can understand that the specific operations of each module in the above-mentioned bank asset preservation business intelligent decision-making system have been referred to above. Figures 1 to 5 The invention has been introduced in detail in the description of the intelligent decision-making method for bank asset preservation business, and therefore, its repeated description will be omitted.

[0063] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed devices, systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the module division is only a logical function division, and there may be other division methods in actual implementation. The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0064] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0065] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference to a figure in a claim should not be considered as limiting the claim to which it relates.

[0066] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. An intelligent decision-making method for bank asset preservation business, characterized in that: include: Obtain litigation applications initiated by account managers; Semantically encoding the text portion of the litigation application to obtain a sequence of sentence-level semantic understanding feature vectors of the litigation application text; Simultaneously mapping each litigation application text sentence granular semantic understanding feature vector in the sequence of litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain a sequence of mapped litigation application text sentence granular semantic understanding feature vectors; Based on the overall semantic feature distribution importance of each mapped litigation application text sentence granularity semantic understanding feature vector relative to the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors, attention fusion is performed on the sequence of litigation application text sentence granularity semantic understanding feature vectors to obtain a sentence granularity attention litigation application text understanding feature vector; Determine whether to retain the entire preliminary hearing based on the sentence-granular attention litigation application text understanding feature vector; The sentence-granular semantic encoding of the litigation application text description is performed to obtain a sequence of sentence-granular semantic understanding feature vectors of the litigation application text, including: After the litigation application text description is processed into sentences, a semantic encoder including a sentence encoder is used to obtain a sequence of sentence-granular semantic understanding feature vectors of the litigation application text; The method of simultaneously mapping each litigation application text sentence granular semantic understanding feature vector in the sequence of litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain a sequence of litigation application text sentence granular semantic understanding feature vectors after mapping includes: Inputting the sequence of the litigation application text sentence granular semantic understanding feature vectors into an attention space mapper based on a fully connected layer to simultaneously map each of the litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain the sequence of the mapped litigation application text sentence granular semantic understanding feature vectors; Wherein, based on the sentence-granular attention litigation application text understanding feature vector, determining whether to keep the entire preliminary review includes: Passing the sentence-granular attention litigation application text understanding feature vector through a classifier to obtain a classification result, wherein the classification result is used to indicate whether to retain the entire preliminary review; The method further comprises a training step of: training the semantic encoder including the sentence encoder, the fully connected layer-based attention space mapper, and the classifier; The training step comprises: Obtaining training data, the training data including training litigation applications initiated by the account manager and whether to preserve all true values ​​of the initial review; Extracting a text portion of the training litigation application to obtain a text description of the training litigation application; After the training litigation application text description is processed into sentences, the training litigation application text description is passed through the semantic encoder including the sentence encoder to obtain a sequence of sentence-granular semantic understanding feature vectors of the training litigation application text; Inputting the sequence of the training litigation application text sentence granular semantic understanding feature vectors into the fully connected layer-based attention space mapper to simultaneously map each of the training litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain a sequence of the training mapped litigation application text sentence granular semantic understanding feature vectors; Calculate the overall attention weight of each training mapped litigation application text sentence granular semantic understanding feature vector in the sequence of training mapped litigation application text sentence granular semantic understanding feature vectors relative to the sequence of training mapped litigation application text sentence granular semantic understanding feature vectors to obtain a sequence of training attention weights; Based on the sequence of training attention weights, fusing the sequence of training litigation application text sentence granularity semantic understanding feature vectors to obtain a training sentence granularity attention litigation application text understanding feature vector; Passing the training sentence granular attention litigation application text understanding feature vector through the classifier to obtain a classification loss function value; Calculate a specific loss function value of a sequence of the training sentence granular attention litigation application text understanding feature vector and the training litigation application text sentence granular semantic understanding feature vector; Using a weighted sum of the classification loss function value and the specific loss function value as a loss function value, the semantic encoder including the sentence encoder, the fully connected layer-based attention space mapper, and the classifier are trained; The specific loss function value of the sequence of the training sentence granular attention litigation application text understanding feature vector and the training litigation application text sentence granular semantic understanding feature vector is calculated using the following specific loss function, wherein the specific loss function is expressed as: V1 is the training sentence granular attention litigation application text understanding feature vector, V2 is the second feature vector after the sequence cascade of the training litigation application text sentence granular semantic understanding feature vector, v 1i and v 2i are the eigenvalues ​​of the training sentence granular attention litigation application text understanding feature vector V1 and the second feature vector V2, L is the length of the feature vector, and represents the square of the vector's two-norm, is the specific loss function value.

2. The intelligent decision-making method for bank asset preservation business according to claim 1, characterized in that: Semantically encoding the text portion of the litigation application to obtain a sequence of sentence-level semantic understanding feature vectors of the litigation application text, including: Extracting a text portion of the litigation application to obtain a text description of the litigation application; The text description of the litigation application is semantically encoded at sentence granularity to obtain a sequence of sentence granularity semantic understanding feature vectors of the litigation application text.

3. The intelligent decision-making method for bank asset preservation business according to claim 2 is characterized in that: Based on the overall semantic feature distribution importance of each mapped litigation application text sentence granularity semantic understanding feature vector relative to the sequence of mapped litigation application text sentence granularity semantic understanding feature vectors, the sequence of litigation application text sentence granularity semantic understanding feature vectors is subjected to attention fusion to obtain a sentence granularity attention litigation application text understanding feature vector, including: Calculating the overall attention weight of each mapped litigation application text sentence granular semantic understanding feature vector in the sequence of mapped litigation application text sentence granular semantic understanding feature vectors relative to the sequence of mapped litigation application text sentence granular semantic understanding feature vectors to obtain a sequence of attention weights; Based on the sequence of attention weights, the sequence of sentence-granularity semantic understanding feature vectors of the litigation application text is fused to obtain the sentence-granularity attention litigation application text understanding feature vector.

4. The intelligent decision-making method for bank asset preservation business according to claim 3 is characterized in that: Calculating the overall attention weight of each mapped litigation application text sentence granular semantic understanding feature vector in the sequence of mapped litigation application text sentence granular semantic understanding feature vectors relative to the sequence of mapped litigation application text sentence granular semantic understanding feature vectors to obtain a sequence of attention weights, including: The attention weight calculation formula is used to calculate the overall attention weight of each mapped litigation application text sentence granular semantic understanding feature vector in the sequence of the mapped litigation application text sentence granular semantic understanding feature vector relative to the sequence of the mapped litigation application text sentence granular semantic understanding feature vector to obtain the sequence of attention weights; wherein the attention weight calculation formula is: Among them, h i is the i-th mapped litigation application text sentence granular semantic understanding feature vector in the sequence of the mapped litigation application text sentence granular semantic understanding feature vector, and A is 1×N w The matrix, N w is the dimension of the sentence-level semantic understanding feature vector of the i-th mapped litigation application text, and B is 1×N h The matrix, N h is the dimension of each mapped litigation application text sentence granular semantic understanding feature vector, N is the number of mapped litigation application text sentence granular semantic understanding feature vectors, σ is the Sigmoid function, S i is the i-th attention weight value.

5. A bank asset preservation business intelligent decision-making system, used to execute the bank asset preservation business intelligent decision-making method as claimed in claim 4, characterized in that: include: The litigation application acquisition module is used to obtain the litigation application initiated by the account manager; A litigation application semantic understanding module, used for semantically encoding the text portion of the litigation application to obtain a sequence of semantic understanding feature vectors of the litigation application text sentence granularity; A same attention space mapping module is used to simultaneously map each litigation application text sentence granular semantic understanding feature vector in the sequence of litigation application text sentence granular semantic understanding feature vectors to the same attention space to obtain a sequence of litigation application text sentence granular semantic understanding feature vectors after mapping; An attention fusion module, used for performing attention fusion on the sequence of sentence granularity semantic understanding feature vectors of the litigation application text based on the overall semantic feature distribution importance of each of the mapped litigation application text sentence granularity semantic understanding feature vectors relative to the sequence of the mapped litigation application text sentence granularity semantic understanding feature vectors to obtain a sentence granularity attention litigation application text understanding feature vector; A management result generation module is used to determine whether to retain the entire preliminary examination based on the sentence-granular attention litigation application text understanding feature vector.

Images