Risk Audit Method, System, Device and Storage Medium Based on Global Pointer

Abstract

The present invention discloses a risk audit method, system, device and storage medium based on a global pointer. The method includes: obtaining a trade background document, performing optical character recognition on the trade background document to obtain trade background information, classifying the trade background information through a pre-created text classification model to obtain the classified trade background information, extracting information from the classified trade background information through a pre-created global pointer model to obtain the extracted information, and performing a risk audit on the extracted information. The present invention can classify the redundant trade background documents, output effective background information, and then be able to perform a deeper risk audit, judge whether the transaction is real, and perform a one-key, fully automatic risk audit of the trade background, improving the accuracy rate.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a risk audit method, system, device and storage medium based on a global pointer. Background Art

[0002] With the rapid development of artificial intelligence, machines are gradually replacing humans to handle some tedious and complex tasks, including the risk audit of the trade background of customers before bank loans. Currently, most of the trade background risk audits adopt the method of keyword matching. The main process includes: uploading a document, performing character recognition on the document to obtain text, performing word segmentation on the text to obtain multiple keywords, and matching each keyword with a preset risk word. If the match is successful, a risk is prompted.

[0003] Since it is a simple text matching method, the traditional audit method has poor effect, low accuracy, and high false audit rate. In addition, manual classification of the background document is required before the risk audit. Moreover, the traditional audit method cannot make further detailed judgments on the information entered by the customer and the uploaded background document. Summary of the Invention

[0004] The main purpose of the present invention is to provide a risk audit method, system, device and storage medium based on a global pointer, aiming to perform the risk audit of the trade background in a one-key and fully automatic manner and improve the accuracy.

[0005] To achieve the above object, an embodiment of the present invention provides a risk audit method based on a global pointer. The method is applied to a risk audit system, and the risk audit system includes: a text classification model and a global pointer model. The method includes the following steps:

[0006] Obtain a trade background document;

[0007] Perform optical character recognition on the trade background document to obtain trade background information;

[0008] Through a pre-created text classification model, perform text classification on the trade background information to obtain the classified trade background information;

[0009] Through a pre-created global pointer model, perform information extraction on the classified trade background information to obtain the extracted information;

[0010] Perform a risk audit on the extracted information.

[0011] Optionally, the step of performing optical character recognition on the trade background document to obtain trade background information includes:

[0012] Preprocess the trade background document to obtain an image file of black and white dot matrix. The preprocessing includes one or more of grayscale conversion, binarization, and skew correction;

[0013] Detect the image file of the black and white dot matrix, and identify trade background information within the text area of the image file.

[0014] Optionally, before the step of obtaining the trade background document, the following steps are further included:

[0015] Create the text classification model, which specifically includes:

[0016] Obtain a batch of financial text information;

[0017] Assign category labels to the financial text information to obtain data with category labels;

[0018] Parse the data with category labels through a pre-trained model to obtain text feature vectors;

[0019] Based on the text feature vectors, combined with the data with category labels, use cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained text classification model.

[0020] Optionally, before the step of parsing the data with category labels through a pre-trained model to obtain text feature vectors, the following steps are included:

[0021] Convert the data with category labels into one-hot vectors;

[0022] Classify based on the one-hot vectors to obtain text categories and label IDs.

[0023] Optionally, before the step of obtaining the trade background document, the following steps are further included:

[0024] Create the global pointer model, which specifically includes:

[0025] Obtain a batch of financial text information;

[0026] Assign entity labels to the financial text information to obtain data with entity labels;

[0027] Parse the data with entity labels through a pre-trained model to obtain text feature vectors;

[0028] Construct two independent attention matrices, multiply and exchange the attention matrices with the text feature vectors to obtain a start pointer matrix and an end pointer matrix;

[0029] Merge the starting pointer matrix and the ending pointer matrix together to obtain a global pointer matrix for different categories;

[0030] Based on the global pointer matrix, combine the data with entity labels, use multi-label cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained global pointer model.

[0031] Optionally, before the step of parsing the data with entity labels through a pre-trained model to obtain text feature vectors, the following steps are included:

[0032] Record the category, starting position, and ending position of the data with entity labels;

[0033] Use the global pointer label to output the category matrix of the data with entity labels.

[0034] Optionally, the step of performing risk review on the extracted information includes:

[0035] Review the amount in the extracted contract and the amount on the invoice provided by the customer;

[0036] Review the two parties (Party A and Party B) in the extracted contract and the information provided by the customer;

[0037] Review the amount the customer needs to loan and the transaction amount in the extracted contract.

[0038] In addition, to achieve the above object, an embodiment of the present invention further proposes a risk review system, and the risk review system includes:

[0039] A character recognition module, configured to obtain a trade background document, perform optical character recognition on the trade background document to obtain trade background information;

[0040] A text classification model, configured to perform text classification on the trade background information to obtain classified trade background information;

[0041] A global pointer model, configured to perform information extraction on the classified trade background information to obtain the extracted information;

[0042] A risk review module, configured to perform risk review on the extracted financial information.

[0043] In addition, to achieve the above object, an embodiment of the present invention further proposes a terminal device, and the terminal device includes a memory, a processor, and a global pointer risk review program stored on the memory and executable on the processor. When the global pointer risk review program is executed by the processor, the steps of the risk review method based on the global pointer as described above are implemented.

[0044] In addition, to achieve the above object, an embodiment of the present invention further provides a computer-readable storage medium, on which a global pointer risk audit program is stored. When the global pointer risk audit program is executed by a processor, the steps of the risk audit method based on the global pointer as described above are implemented.

[0045] A risk audit method, system, device and storage medium based on a global pointer proposed by an embodiment of the present invention, obtain a trade background document; perform optical character recognition on the trade background document to obtain trade background information; through a pre-created text classification model, perform text classification on the trade background information to obtain classified trade background information; through a pre-created global pointer model, perform information extraction on the classified trade background information to obtain the extracted information; perform risk audit on the extracted information. Thus, through the above solution, the redundant trade background documents are classified, and effective background information is output, so that deeper risk audit can be carried out to judge whether the transaction is real, and the risk audit of the trade background is carried out in a one-key and fully automatic manner, improving the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the risk audit method based on the global pointer of the present invention;

[0047] Figure 2 It is a schematic diagram of the global pointer model construction process involved in the embodiment of the risk audit method based on the global pointer of the present invention;

[0048] Figure 3 It is a schematic diagram of the process of the first embodiment of the risk audit method based on the global pointer of the present invention;

[0049] Figure 4 It is a schematic diagram of the process of the second embodiment of the risk audit method based on the global pointer of the present invention;

[0050] Figure 5 It is a schematic diagram of the functional modules of the first embodiment of the risk audit system based on the global pointer of the present invention.

[0051] The implementation, functional features and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0053] The main solution of the embodiment of the present invention is: obtaining a trade background document; performing optical character recognition on the trade background document to obtain trade background information; classifying the trade background information through a pre-created text classification model to obtain classified trade background information; extracting information from the classified trade background information through a pre-created global pointer model to obtain the extracted information; and performing risk review on the extracted information. Thus, through the above solution, the redundant trade background documents are classified to output effective background information, and further deeper risk review can be performed to determine whether the transaction is real, and the risk review of the trade background is performed in a one-key and fully automatic manner, improving the accuracy rate.

[0054] The embodiment of the present invention considers that in the existing related solutions, most of the risk review systems for trade backgrounds adopt keyword matching. Since it is a simple text matching, the review method has poor effect, low accuracy rate, and high false review rate. In addition, manual classification of background documents is required before risk review. Moreover, the traditional review method cannot make further detailed judgments on the information entered by customers and the uploaded background documents.

[0055] Therefore, the embodiment of the present invention proposes a solution that can realize one-key and fully automatic risk review of the trade background, classify the redundant trade background documents, output effective background information, and further perform deeper risk review to determine whether the transaction is real, improving the accuracy rate.

[0056] Specifically, referring to Figure 1 , Figure 1 is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the risk review method based on the global pointer of the present invention.

[0057] The device of the embodiment of the present invention can be a mobile terminal or a server device.

[0058] As Figure 1 shown, the device may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard). Optionally, the user interface 1003 may further include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory or a stable memory (non-volatile memory), such as a disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0059] Those skilled in the art can understand that Figure 1 the device structure shown in

[0060] does not constitute a limitation on the device, and may include more or fewer components than those shown, or combine certain components, or have different component arrangements. Figure 1 As shown in

[0061] In Figure 1 the system shown, the network interface 1004 is mainly used to connect to a network server and communicate with the network server for data; the user interface 1003 is mainly used to interact with users and receive instructions input by users; and the processor 1001 can be used to call the global pointer risk review program stored in the memory 1005 and perform the following operations:

[0062] Obtain a trade background document;

[0063] Perform optical character recognition on the trade background document to obtain trade background information;

[0064] Perform text classification on the trade background information through a pre-created text classification model to obtain classified trade background information;

[0065] Perform information extraction on the classified trade background information through a pre-created global pointer model to obtain extracted information;

[0066] Perform risk review on the extracted information.

[0067] Furthermore, when the global pointer risk review program is run by the processor, the following operations are also implemented:

[0068] Preprocess the trade background document to obtain an image file of black and white dot matrix, and the preprocessing includes one or more of grayscale conversion, binarization, and skew correction;

[0069] Detect the black and white dot matrix image file and recognize trade background information within the text area of the image file.

[0070] Furthermore, when the global pointer risk review program is run by the processor, the following operations are also implemented:

[0071] Create the text classification model, specifically including:

[0072] Obtain a batch of financial text information;

[0073] Label the financial text information with category labels to obtain data with category labels;

[0074] Parse the data with category labels through a pre-trained model to obtain text feature vectors;

[0075] Based on the text feature vectors, combined with the data with category labels, use cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained text classification model.

[0076] Further, when the global pointer risk audit program is run by the processor, the following operations are also implemented:

[0077] Convert the data with category labels into one-hot vectors;

[0078] Classify based on the one-hot vectors to obtain the text category and label ID.

[0079] Further, when the global pointer risk audit program is run by the processor, the following operations are also implemented:

[0080] Create the global pointer model, specifically including:

[0081] Obtain a batch of financial text information;

[0082] Label the financial text information with entity labels to obtain data with entity labels;

[0083] Parse the data with entity labels through a pre-trained model to obtain text feature vectors;

[0084] Construct two independent attention matrices, multiply and exchange the attention matrices with the text feature vectors to obtain a start pointer matrix and an end pointer matrix;

[0085] Merge the start pointer matrix and the end pointer matrix together to obtain global pointer matrices of different categories;

[0086] Based on the global pointer matrices, combined with the data with entity labels, use multi-label cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained global pointer model.

[0087] Further, when the global pointer risk audit program is run by the processor, the following operations are also implemented:

[0088] Record the category, start position, and end position of the data with entity labels;

[0089] Use the global pointer labels to output the category matrix of the data with entity labels.

[0090] Further, when the global pointer risk audit program is run by the processor, the following operations are also implemented:

[0091] Audit the amount in the extracted contract and the amount on the invoice provided by the customer;

[0092] Audit both parties (Party A and Party B) in the extracted contract and the information provided by the customer;

[0093] Audit the amount the customer needs to borrow and the transaction amount in the extracted contract.

[0094] To better understand the above technical solution, the exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0095] Based on the above terminal device and system architecture but not limited to the above architecture, an embodiment of the visualization lung ventilation monitoring method of the present invention is proposed.

[0096] Refer to Figure 2 , Figure 2 is a schematic diagram of the global pointer model construction process involved in the embodiment of the risk audit method based on the global pointer of the present invention. As Figure 2As shown in the figure, the adopted global pointer model structure is a NEZHA model connected to a global pointer layer, and the NEZHA model structure is similar to the BERT model. The detailed structure is as follows: The text feature vector [batch_size, seq_len, hidden_size] obtained after parsing by the pre-trained model NEZHA model represents the semantic information of each input sequence in each dimension, where batch_size refers to the number of samples in batch processing, seq_len refers to the length of the text sequence, and hidden_size refers to the size of the hidden layer. At this time, two independent bi-affine attention matrices are constructed and multiplied with the intermediate feature vector of the text to realize the association between text features, and a start pointer matrix and an end pointer matrix [batch_size, seq_len, num_cate, head] are obtained, where batch_size refers to the number of samples in batch processing, seq_len refers to the length of the text sequence, num_cate refers to the number of entity types to be recognized, and head refers to the size of the head of the attention matrix. Finally, the associated results, that is, the two matrices, are combined together as the output of the global pointer, and the dimension is [batch_size, num_cate, seq_len, seq_len], where batch_size refers to the number of samples in batch processing, num_cate refers to the number of entity types to be recognized, and seq_len refers to the length of the text sequence. In this way, the start matrix and the end matrix of the label category are no longer unassociated, and each label category has a matrix with a dimension of (sequence length, sequence length) to represent the start position and the end position of the label category.

[0097] Refer to Figure 3 , Figure 3 is a schematic flowchart of the first embodiment of the risk audit method based on the global pointer of the present invention. As Figure 3 shown, the first embodiment of the present invention proposes a risk audit method based on the global pointer. The method is applied to a risk audit system, and the risk audit system includes: a text classification model, a global pointer model. The solution of this embodiment involves a one-key and fully automatic risk audit of the trade background to improve the accuracy.

[0098] Specifically, the risk audit method based on the global pointer in this embodiment includes:

[0099] Step S101, obtain a trade background document.

[0100] Among them, the trade background documents of the customer include: picture documents such as contracts, invoices, and letters of explanation. After obtaining the customer's trade background documents, they are uploaded to the risk audit system.

[0101] Step S102, perform optical character recognition on the trade background document to obtain trade background information.

[0102] More specifically, first preprocess the trade background document uploaded to the risk review system. The preprocessing includes: grayscale conversion, binarization, skew correction, etc., to convert the trade background document into a black-and-white dot matrix image file. Detect the area with text in the black-and-white dot matrix image file, identify the text in the area, and output the obtained trade background information.

[0103] Specifically, each pixel of the grayscale image only needs one byte to store the grayscale value, and the grayscale value range is 0 - 255. At the same time, the weighted average method is used to obtain the grayscale value of each pixel point. During the binarization process, the grayscale value of the pixel points on the image is set to 0 or 255, presenting an obvious visual effect of only black and white for the entire image. In addition, the trade background image scanned by the scanner may be skewed. Through optical character recognition, the skew angle of the image can be automatically detected and corrected.

[0104] Step S103: Through a pre-created text classification model, classify the trade background information to obtain the classified trade background information.

[0105] Specifically, the text recognized in the input document is predicted through a pre-created text classification model, and the category to which the trade background information belongs can be predicted, thereby classifying the trade background information into different categories.

[0106] The text classification model is a deep learning model. The structure is that the output of the CLS of a pre-training model similar to BERT is taken, and then a fully connected layer is connected. The pre-training model is the NEZHA model.

[0107] Step S104: Through a pre-created global pointer model, extract information from the classified trade background information to obtain the extracted information.

[0108] Specifically, through a pre-created global pointer model, information extraction is performed on the document that needs to extract information. For example, for a document belonging to a trade contract, important financial information such as the name of Party A, the name of Party B, the total contract amount, the contract signing date, the contract termination date, and the settlement method are extracted. According to the input trade background information for prediction, the corresponding entity segments in the financial text information can be predicted and extracted.

[0109] The global pointer model is a deep learning model. The structure is that a pre-training model similar to BERT is connected with a global pointer layer. The pre-training model is the NEZHA model.

[0110] Step S105: Conduct a risk review on the extracted information.

[0111] Specifically, according to different business scenarios, the extracted financial text information is used for risk review. For example, to determine whether the trade background of a trade contract transaction is true: whether the amount in the extracted contract is consistent with the amount on the invoice provided by the customer; whether the two parties (Party A and Party B) in the extracted contract are consistent with the information provided by the customer; whether the amount the customer needs to loan is less than or equal to the transaction amount in the extracted contract. In addition, other risk review points that can be paired through keywords, such as whether the goods traded are within the specified range, etc.

[0112] In this embodiment, through the above solution, a trade background document is obtained; optical character recognition is performed on the trade background document to obtain trade background information; through a pre-created text classification model, the trade background information is classified to obtain classified trade background information; through a pre-created global pointer model, information extraction is performed on the classified trade background information to obtain the extracted information; risk review is performed on the extracted information. Thus, through the above solution, it is possible to automatically classify the redundant background documents, focus on important materials in the financial field, such as contracts, invoices, etc., and be able to structure the documents, outputting effective information, including: information of both parties (Party A and Party B), contract amount and time information, settlement method, etc., and then perform a deeper level of risk review to determine whether the transaction is true, improving the accuracy rate.

[0113] Refer to Figure 4 , Figure 4 which is a schematic flowchart of the second embodiment of the risk review method based on the global pointer of the present invention.

[0114] In this embodiment, based on the above Figure 3 shown embodiment, before the above step S101, the following steps are further included:

[0115] Step S1001, create the text classification model;

[0116] Step S1002, create the global pointer model.

[0117] Specifically, the training process of the text classification model can be as follows:

[0118] Such as Figure 4As shown in the figure, first obtain a batch of financial text information, label the batch of financial text information with category labels to obtain data with category labels. Immediately preprocess the data with category labels, convert the data with category labels into one-hot vectors, and classify according to the corresponding one-hot vectors to obtain the corresponding text categories and label IDs. For example, there are a total of 5 categories. The first category vector is [1 0 0 0 0], the second category vector is [0 1 0 0 0], the third category vector is [0 0 1 0 0], the fourth category vector is [0 0 0 1 0], and the fifth category vector is [0 0 0 0 1]. For example, the ID of the contract category is 0, the ID of the notice letter is 1, the ID of the invoice is 2, the ID of the credit record category is 3, and the ID of the other category is 4. After being parsed by the pre-trained model, a text feature vector is obtained. Based on the text feature vector, a large amount of financial information is used for model training and fitting, and the cross-entropy is used to construct the loss function of the model. The model is trained with this loss function to obtain the trained text classification model. And during the training process, a part of the validation set is used to verify the effect of the model.

[0119] After the model training is completed, save the overall structure and weights of the existing model, load the saved text classification model, and make predictions according to the input text information. The model can predict the category to which the text information belongs, so as to classify the redundant financial text information, which is convenient for subsequent information extraction. For example, after passing through the pre-created text classification model, the output vector is [0.8 0.1 0.1 0 0], and it can be concluded that the probability of this text belonging to [1 0 0 0 0] is the largest, and it is predicted that this text belongs to the first category and is classified into the first category.

[0120] In this embodiment, the text classification model structure adopted is to take the output of CLS of the NEZHA model, that is, the vector representation of the entire sequence, and then connect a traditional fully connected layer with the output dimension being the number of categories to construct a classic text classification model. The NEZHA model structure is similar to the BERT model and adopts the relative position encoding method, so that the text length in text classification is no longer limited, solving the limitation of the text length being too long in traditional text classification methods in review, and greatly reducing the misreview rate.

[0121] Specifically, the model training process of the global pointer model can be as follows:

[0122] First, obtain a batch of financial text information, label the batch of financial text information with entity tags, and record the tag category and word position information. Immediately preprocess the data with entity tags, record the category, start position, and end position of the entity, use the global pointer tag to obtain the corresponding entity category matrix, with all default values being 0, and then fill in the values according to the tags. For example, if the word "Five thousand yuan" has position information 4, 5, 6 in the text, the category is contract amount, and the category ID is 2, then the element [2, 4, 6] of this matrix is 1. The text feature vector obtained after parsing by the pre-trained model represents the semantic information of each input sequence in each dimension. At the same time, construct two independent attention matrices, perform a product transformation with the intermediate feature vector of the text to realize the association between text features, obtain a start pointer matrix and an end pointer matrix, and finally merge the associated results, that is, the two matrices, together as the output of the global pointer. In this way, the start matrix and end matrix of the tag category are no longer unassociated, and each tag category has a matrix to represent the start position of the tag category. Based on the global pointer matrix, combine the data with entity tags to train and fit the model, use multi-label cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained global pointer model. And during the training process, use part of the validation set to verify the effect of the model.

[0123] After the model training is completed, save the overall structure and weights of the existing model, load the saved global pointer model, and make predictions according to the input text information. The model can predict and extract the corresponding entity segments in the text information. For example, after passing through the pre-created global pointer model, the start row and end position of the corresponding entity category matrix in the output sentence are 1, and the others that are not entities are values below 0, and the corresponding text information is extracted.

[0124] In this embodiment, the global pointer model structure adopted is a NEZHA model connected to a global pointer layer. The NEZHA model structure is similar to the BERT model and adopts the relative position encoding method, so that the text length in text classification is no longer restricted, greatly reducing the misjudgment rate. In addition, the global pointer technology used in the global pointer model solves the problem of entity nesting in information extraction in the financial field. In the traditional review method, a text belongs to only one entity, and problems such as information extraction errors and information extraction incompleteness will occur during the information extraction process. However, the global pointer technology associates the start matrix and the end matrix, and a text can belong to multiple entities, and can accurately extract the financial text information that needs to be known, greatly improving the accuracy.

[0125] In addition, the embodiment of the present invention also proposes a risk review system, referring to Figure 5 , Figure 5This is a schematic diagram of the functional modules of the first embodiment of the global pointer-based risk audit system of the present invention. As Figure 5 shown, the risk audit system includes:

[0126] A character recognition module, configured to obtain a trade background document, perform optical character recognition on the trade background document, and obtain trade background information;

[0127] A text classification model, configured to perform text classification on the trade background information to obtain classified trade background information;

[0128] A global pointer model, configured to perform information extraction on the classified trade background information to obtain extracted information;

[0129] A risk audit module, configured to perform risk audit on the extracted financial information.

[0130] For the principle and implementation process of implementing global pointer-based risk audit in this embodiment, please refer to the above-mentioned embodiments, and details will not be repeated here.

[0131] In addition, an embodiment of the present invention further provides a terminal device, where the terminal device includes a memory, a processor, and a global pointer risk audit program stored on the memory and executable on the processor. When the global pointer risk audit program is executed by the processor, it implements the steps of the risk audit method based on the global pointer as described in the above embodiments.

[0132] Since all the technical solutions of the foregoing all embodiments are adopted when the global pointer risk audit program is executed by the processor, it has at least all the beneficial effects brought by all the technical solutions of the foregoing all embodiments, and details will not be repeated here.

[0133] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which a global pointer risk audit program is stored. When the global pointer risk audit program is executed by the processor, it implements the steps of the risk audit method based on the global pointer as described in the above embodiments.

[0134] Since all the technical solutions of the foregoing all embodiments are adopted when the global pointer risk audit program is executed by the processor, it has at least all the beneficial effects brought by all the technical solutions of the foregoing all embodiments, and details will not be repeated here.

[0135] Compared with the prior art, the risk audit method, system, device and storage medium based on the global pointer proposed in the embodiments of the present invention obtain a trade background document; perform optical character recognition on the trade background document to obtain trade background information; classify the trade background information through a pre-created text classification model to obtain the classified trade background information; extract information from the classified trade background information through a pre-created global pointer model to obtain the extracted information; and perform risk audit on the extracted information. This solution realizes the classification of redundant trade background documents, outputs effective background information, and can further perform deeper risk audits to determine whether the transaction is real, and perform one-key, fully automatic risk audits of the trade background, improving the accuracy rate.

[0136] It should be noted that in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the phrase "including a..." does not exclude the existence of additional identical elements in the process, method, article or system including that element.

[0137] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0138] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium as described above (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, controlled terminal, or network device, etc.) to execute the methods of each embodiment of the present invention.

[0139] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A risk audit method based on a global pointer, characterized in that, The method is applied to a risk audit system, which includes: a text classification model and a global pointer model. The method includes the following steps: Obtain a trade background document; Perform optical character recognition on the trade background document to obtain trade background information; Through a pre-created text classification model, classify the trade background information to obtain the classified trade background information; Through a pre-created global pointer model, extract information from the classified trade background information to obtain the extracted information; Conduct a risk audit on the extracted information; Before the step of obtaining the trade background document, it further includes: Create the global pointer model, specifically including: Obtain a batch of financial text information; Label the financial text information with entity tags to obtain data with entity tags; Preprocess the data with entity tags, and record the category, start position, and end position of the entity; Through a pre-trained model, parse the data with entity tags to obtain text feature vectors; Construct two independent attention matrices, perform a product transformation on the attention matrices and the text feature vectors to obtain a start pointer matrix and an end pointer matrix; Merge the start pointer matrix and the end pointer matrix together to obtain global pointer matrices of different categories; Based on the global pointer matrices, combined with the data with entity tags, use multi-label cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained global pointer model.

2. The risk audit method based on a global pointer according to claim 1, wherein The step of performing optical character recognition on the trade background document to obtain trade background information includes: Preprocess the trade background document to obtain an image file of black and white dot matrices. The preprocessing includes one or more of grayscale conversion, binarization, and skew correction; Detect the image file of black and white dot matrices, and identify trade background information within the text area of the image file.

3. The risk audit method based on a global pointer according to claim 1, characterized in that, Before the step of obtaining the trade background document, it further includes: Create the text classification model, specifically including: Obtain a batch of financial text information; Label the financial text information with category tags to obtain data with category tags; Through a pre-trained model, parse the data with category tags to obtain text feature vectors; Based on the text feature vectors, combined with the data with category tags, use cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained text classification model.

4. The risk audit method based on a global pointer according to claim 3, wherein Before the step of parsing the data with category tags through a pre-trained model to obtain text feature vectors, it includes: Convert the data with category tags into one-hot vectors; Classify based on the one-hot vectors to obtain text categories and label IDs.

5. The risk audit method based on a global pointer according to claim 1, wherein Before the step of parsing the data with entity tags through a pre-trained model to obtain text feature vectors, it includes: Record the category, start position, and end position of the data with entity tags; Use global pointer tags to output the category matrix of the data with entity tags.

6. The risk audit method based on a global pointer according to any one of claims 1-5, characterized in that The step of conducting a risk audit on the extracted information includes: Verify the amounts in the extracted contracts and the amounts on the invoices provided by the customers; Verify the two parties (Party A and Party B) in the extracted contracts and the information provided by the customers; Verify the amount the customer needs to borrow and the transaction amount in the extracted contract.

7. A risk audit system, characterized in that, The risk verification system includes: A character recognition module, which is used to obtain a trade background document, perform optical character recognition on the trade background document, and obtain trade background information; A text classification model, which is used to classify the trade background information to obtain the classified trade background information; A global pointer model, which is used to extract information from the classified trade background information to obtain the extracted information; A risk verification module, which is used to verify the risks of the extracted financial information; The creation process of the global pointer model includes: Obtain a batch of financial text information; Mark entity labels on the financial text information to obtain data with entity labels; Preprocess the data with entity labels, and record the category, start position, and end position of the entity; Parse the data with entity labels through a pre-trained model to obtain text feature vectors; Construct two independent attention matrices, perform a product transformation on the attention matrices and the text feature vectors to obtain a start pointer matrix and an end pointer matrix; Merge the start pointer matrix and the end pointer matrix together to obtain global pointer matrices of different categories; Based on the global pointer matrices, combine the data with entity labels, use multi-label cross-entropy to construct the loss function of the model, and train the model with this loss function to obtain the trained global pointer model.

8. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a global pointer risk verification program stored on the memory and executable on the processor. When the global pointer risk verification program is executed by the processor, it implements the steps of the risk verification method based on the global pointer as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, A global pointer risk verification program is stored on the computer-readable storage medium. When the global pointer risk verification program is executed by the processor, it implements the steps of the risk verification method based on the global pointer as described in any one of claims 1-6.

Images