Letter of credit review method, device, electronic device and storage medium

By integrating directed graphs and two-way recurrent neural network models, the terms and characteristics of letter of credit data are automatically extracted and reviewed, the problem of cumbersome and inefficient letter of credit review process is solved, efficient and accurate letter of credit review is achieved, and labor costs are reduced and transaction risks.

CN114492408BActive Publication Date: 2024-11-26AGRICULTURAL BANK OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210100993.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-27
Publication Date
2024-11-26
Estimated Expiration
2042-01-27

AI Technical Summary

Technical Problem

The existing letter of credit review process is cumbersome and inefficient, especially for witnesses with native Chinese speaking, it is prone to errors, which increases labor costs and trade transaction risks.

Method used

A model that integrates directed graphs and bidirectional recurrent neural networks is adopted to automatically extract the clause characteristics in letter of credit data and review them based on the pre-generated certificate review rules engine library to reduce manual intervention.

Benefits of technology

It improves the efficiency of letter of credit audit, reduces errors, reduces labor costs, and meets the needs of emergency audits, avoiding the risks of trade transactions in goods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114492408B_ABST
    Figure CN114492408B_ABST
Patent Text Reader

Abstract

The embodiment of the present application discloses a method, device, electronic device and storage medium for reviewing a letter of credit. The method comprises: when receiving the letter of credit data to be reviewed, intercepting the content of each message domain in the letter of credit data; for any message domain, inputting the message domain content into a pre-trained neural network model, obtaining the clause features of the message domain content output by the neural network model, the neural network model being a model integrating a directed graph and a bidirectional recurrent neural network; based on the clause feature specifications of each message domain in the pre-generated review rule matching engine library and the logical relationship between each message domain, reviewing each message domain of the letter of credit data according to the clause features can avoid manual review and effectively improve efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of document review, and in particular to a method, device, electronic device and storage medium for reviewing a letter of credit. Background Art

[0002] With the continuous development of international trade, the types of international letter of credit business of banks have become more and more diverse, and the review process has become more and more complicated. The content of letters of credit is generally written in English and contains a lot of professional language, which often requires reviewers to have a strong English foundation.

[0003] Currently, it is usually only carried out through manual review. Even experienced and well-trained professional review experts can only review about 5 orders a day. In addition, due to the information asymmetry of the content of the letter of credit, there is a lot of repetitive work and heavy workload in the letter of credit review process, which increases the difficulty of the review. Especially for reviewers whose native language is Chinese, it is extremely easy to make various errors.

[0004] This often requires another professional reviewer to conduct the review, which obviously increases the labor cost of reviewing letters of credit. In addition, the review of letters of credit has a relatively high time limit. If it is not completed on time, it may face the transaction risk of goods trade and cause irreparable losses. Therefore, manual review not only requires high capabilities of reviewers, but also the traditional method is prone to errors and low efficiency. If the review is not timely, it will also lead to risks in goods trade transactions. Summary of the invention

[0005] The embodiments of the present application provide a method, device, electronic device and storage medium for reviewing a letter of credit to improve the efficiency of reviewing a letter of credit.

[0006] In a first aspect, an embodiment of the present application provides a method for reviewing a letter of credit, the method comprising:

[0007] When receiving the letter of credit data to be reviewed, intercepting the content of each message field in the letter of credit data;

[0008] For any electronic message domain, input the electronic message domain content into a pre-trained neural network model to obtain the clause features of the electronic message domain content output by the neural network model, wherein the neural network model is a model integrating a directed graph and a bidirectional recurrent neural network;

[0009] Based on the pre-generated clause feature specifications of each message field in the verification rule matching engine library and the logical relationship between each message field, each message field of the letter of credit data is reviewed according to the clause features.

[0010] In a second aspect, the embodiment of the present application further provides a letter of credit review device, the letter of credit review device comprising:

[0011] An interception module, for intercepting the content of each message field in the letter of credit data upon receiving the letter of credit data to be reviewed;

[0012] A clause feature output module, for any telegram domain, inputting the content of the telegram domain into a pre-trained neural network model, and obtaining the clause features of the content of the telegram domain output by the neural network model, wherein the neural network model is a model integrating a directed graph and a bidirectional recurrent neural network;

[0013] The review module is used to review each message field of the letter of credit data according to the terms and conditions based on the pre-generated review rule matching engine library and the logical relationship between the message fields.

[0014] In a third aspect, an embodiment of the present application further provides an electronic device, the electronic device comprising:

[0015] one or more processors;

[0016] a storage device for storing one or more programs,

[0017] When the one or more programs are executed by the one or more processors, the one or more processors implement the letter of credit review method provided in any embodiment of the present application.

[0018] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, a letter of credit review method as provided in any embodiment of the present application is implemented.

[0019] The technical solution of the embodiment of the present application obtains the clause features of the content of each telegram field in the letter of credit data by utilizing a model that integrates a directed graph and a bidirectional recurrent neural network, and then reviews each telegram field of the letter of credit data based on the clause feature specifications of each telegram field in the pre-generated review rule matching engine library and the logical relationship between the telegram fields. This can avoid manual review and effectively improve efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flowchart of a method for reviewing a letter of credit provided in Example 1 of the present application;

[0021] Figure 2 A structural diagram of a bidirectional recurrent neural network provided in Example 1 of the present application;

[0022] Figure 3 A schematic diagram of the structure of a directed graph provided in Example 1 of the present application;

[0023] Figure 4 A schematic diagram of the structure of a letter of credit review device provided in Example 2 of the present application;

[0024] Figure 5 A schematic diagram of the structure of an electronic device provided in Example 3 of the present application. DETAILED DESCRIPTION

[0025] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only the parts related to the present application, rather than all structures, are shown in the accompanying drawings.

[0026] Embodiment 1

[0027] Figure 1 This is a flow chart of the method for reviewing a letter of credit provided in the first embodiment of the present application. This embodiment is applicable to the scenario of reviewing a letter of credit. The method can be executed by a letter of credit review device, which can be implemented in hardware and / or software and can generally be integrated in an electronic device such as a computer with data computing capabilities, and specifically includes the following steps:

[0028] Step 101: When receiving the letter of credit data to be reviewed, intercept the message field content of each message field in the letter of credit data.

[0029] It should be noted that the letter of credit data to be reviewed refers to the data corresponding to the received letter of credit to be reviewed, which is generally stored in a message. Its format may not be standardized and there may be some spelling errors. Therefore, the letter of credit data can be firstly processed by intelligent error correction based on the preset noise channel model.

[0030] Taking spelling errors as an example, firstly, it is necessary to find out the misspelled words in the content of the telegram domain. This can be achieved by looking up a dictionary, such as scanning each word in turn. If the word is not in the dictionary (not included in the dictionary), it is considered to be a misspelled word.

[0031] Then, we need to select the "most accurate" word from a group of candidate correct words, and this "most accurate" word is the result we are looking for, that is, the correct word corresponding to the wrong word. The correct word candidate can be determined by the edit distance algorithm. For the specific content of the edit distance algorithm, please refer to the relevant technology, which will not be repeated here.

[0032] After intelligent error correction, the content of each message field is intercepted from the letter of credit data according to the pre-agreed correspondence between the message field and the data structure of the letter of credit data. It should be noted that since the letter of credit data exists in the form of a message, and the message has a fixed data structure, that is, it contains multiple data fields, and the position of each data field in the message is also pre-agreed. Therefore, the data fields occupied by each message field, that is, the correspondence between each message field and the data structure of the letter of credit data, can be pre-agreed, and then based on the correspondence, the message field content of the corresponding message field can be extracted.

[0033] It should be noted that the number and types of electronic fields in a letter of credit are fixed, and the content of each electronic field is also fixed. Therefore, the above-mentioned agreed correspondence can be used in the same type of transaction letters of credit.

[0034] Step 102: For any message domain, input the message domain content into a pre-trained neural network model to obtain the clause features of the message domain content output by the neural network model. The neural network model is a model that integrates a directed graph and a bidirectional recurrent neural network.

[0035] In this step, the pre-trained neural network model is a model that integrates a directed graph and a bidirectional recurrent neural network, where the bidirectional recurrent neural network refers to a BRNN bidirectional recurrent neural network. For the structure of BRNN, please refer to Figure 2 , Figure 2 A structural diagram of a bidirectional recurrent neural network provided in Example 1 of the present application, such as Figure 2 As shown in the figure, at each time t, the input will simultaneously provide two RNNs in opposite directions, and these two recurrent networks are jointly connected to the output layer.

[0036] Since the hidden layer of BRNN needs to record the two values ​​before and after Ai and Ai' at the same time, and participate in the forward and reverse calculations respectively, the final output yi is determined by Ai and Ai'. In this embodiment, the two values ​​before and after can be reflected as context. For example, a sentence C in the telegram domain is represented as {x1, x2, ..., x t} sequence, then the hidden layer of BRNN can take into account the context of a word in the sentence, which is very helpful in judging whether the input word conforms to the semantic rules in each telegram domain.

[0037] For the input layer of the neural network model (x0~x i One end), its role is to use distributed word vectors (word2vec) to represent the content of each telegram domain. It is a context-based dense vector representation method that can reduce text data from a high-dimensional sparse matrix to a low-dimensional dense matrix.

[0038] For the hidden layer of the neural network model (s0~s i and s0'~s i ' one end), using the word vector matrix obtained by the input layer, the main feature information valuable for the classification of each telegram domain is automatically obtained through the context content, usually by adding activation functions, dot multiplication, matrix multiplication and other operations. For details, please refer to the description of the BRNN bidirectional recurrent neural network related technology, which will not be repeated here.

[0039] It should be noted that the hidden layer can determine whether the input Xi feature is important through the attention mechanism, and filter out unimportant features using the threshold, thereby automatically extracting feature label information that contributes to the classification of each telegram domain.

[0040] The output layer of the neural network model (y0~y i One end) is used to output the feature words (which can also be called central feature label words) corresponding to the feature label information extracted by the aforementioned hidden layer. The feature words serve as nodes of the directed graph model to facilitate the mining of the logical relationship between each telegram domain.

[0041] In this step, the feature of the clause output by the neural network model is the feature word output by the aforementioned output layer (i.e., the central feature label word).

[0042] Step 103: Based on the pre-generated document review rule matching engine library, the clause feature specifications of each message field and the logical relationship between each message field, each message field of the letter of credit data is reviewed according to the clause features.

[0043] It should be noted that the verification rule matching engine library in this step is pre-generated, and specifically, the verification rule matching engine library can be pre-generated based on letter of credit sample data and a neural network model.

[0044] First, sample letter of credit data can be obtained and input into a bidirectional recurrent neural network. For each piece of sample letter of credit data, the central feature label words of each message domain output by the bidirectional recurrent neural network and the weights of each central feature label word are obtained.

[0045] The processing process is the same as the input layer, hidden layer and output layer mentioned above, wherein, for any letter of credit sample data, the preset distributed word vector dictionary is used to map the message domain content of each message domain in the letter of credit sample data to obtain the feature vector corresponding to each message domain; for any message domain, the feature vector of the message domain is input into the bidirectional recurrent neural network to obtain the central feature label word extracted by the bidirectional recurrent neural network; the weight of each central feature label word in the message domain is calculated.

[0046] The weight of each central feature label word in the telegram domain can be calculated by using the softmax nonlinear activation function in machine learning, and then the calculated result is mapped to the (0, 1) interval to obtain the weight. It should be noted that in this embodiment, the weight obtained after the mapping can be used as a judgment standard for the prompt level.

[0047] In a specific example, when the weight is set to [0.8,1], it means that the word must contain the label word in the domain, otherwise a suggestion of "must be modified" is prompted; similarly, when the weight is set to [0.4,0.8), it means that the word is recommended to contain the label word in the domain, otherwise a suggestion of "suggested modification" is prompted; when the weight is set to [0,0.4), it means that the word can have no label word in the domain and no suggestion is displayed.

[0048] In addition, regarding the logical relationship between the features of the clauses in each message domain, according to the UCP600 international convention, expert experience information, practical rules, etc. (such as Table 1), it can be known that the label words in each message domain must be logically related to each other.

[0049] Table 1

[0050]

[0051] Therefore, this embodiment forms a directed graph structure based on the central feature label words of each message field. For details, please refer to Figure 3 , Figure 3 A schematic diagram of a directed graph structure provided in Example 1 of the present application. Figure 3 As shown in the figure, the central feature label words of each message domain can be used as the node BRNNi of the directed graph, and the directed edge labels between the nodes can be represented as eij, and the weight of the edge label is represented as wij. The larger the weight value of the edge label, the closer the association between the two domains.

[0052] In this embodiment, an edge label weight threshold can be set to determine whether there is a correlation between two message domains. For example, the edge label weight wij can be standardized so that it is in the interval [0,1]. When the edge label weight output value is [0,0.4), it means that there is no correlation between the two message domains, and the correlation is automatically discarded.

[0053] The remaining directed edge labels represent the logical relationship between the message domains. In order to integrate the logical relationship into the verification rule matching engine library, the logical relationship can be standardized in the form of triples or tuples. For example, the triple (message domain 1, message domain 2, logical reasoning relationship) can be shown in Table 2, and the tuple (message domain 1, message domain 2, ..., message domain n, logical reasoning relationship) can be shown in Table 3.

[0054] Table 2

[0055]

[0056] Table 3

[0057]

[0058] In addition, in this step, the clause features include the central feature label words contained in the telegram domain. The audit process can be specifically performed by comparing the coverage of the central feature label words in the telegram domain according to the clause feature specifications of each telegram domain in the audit rule matching engine library; determining the content association of the central feature label words in the telegram domain according to the logical relationship between each telegram domain in the audit rule matching engine library; and determining the audit results of each telegram domain of the letter of credit data based on the coverage and content association.

[0059] It should be noted that the coverage mainly indicates whether all the feature words required to be included in the telegram domain exist, and the content association refers to whether the associated feature words between the telegram domains are correctly associated.

[0060] In addition, in order to ensure that the review rule matching engine library is updated at all times, after each review of the letter of credit data, the existence weights corresponding to the terms features of each message field output by the neural network model based on the letter of credit data to be reviewed and the association weights of the terms features of each message field are obtained; the review rule matching engine library is updated based on the existence weights and association weights.

[0061] In this embodiment, by utilizing a model that integrates a directed graph and a bidirectional recurrent neural network, the clause features of the content of each message field in the letter of credit data are obtained, and then based on the pre-generated review rule matching engine library, the clause feature specifications of each message field and the logical relationship between the message fields are reviewed. According to the clause features, each message field of the letter of credit data is reviewed, which can avoid manual review and effectively improve efficiency.

[0062] Embodiment 2

[0063] Figure 4 This is a schematic diagram of the structure of a letter of credit review device provided in the second embodiment of the present application. The letter of credit review device provided in the embodiment of the present application can execute the letter of credit review method provided in any embodiment of the present application, and has the corresponding functional modules and beneficial effects of the execution method. The device can be implemented in software and / or hardware, such as Figure 4 As shown, the letter of credit review device specifically includes: an interception module 401, a clause feature output module 402, and a review module 403.

[0064] The interception module is used to intercept the content of each message field in the letter of credit data when receiving the letter of credit data to be reviewed;

[0065] A clause feature output module is used to input the content of any telegram domain into a pre-trained neural network model to obtain the clause features of the content of the telegram domain output by the neural network model, where the neural network model is a model integrating a directed graph and a bidirectional recurrent neural network;

[0066] The review module is used to review each message field of the letter of credit data according to the terms and conditions based on the pre-generated review rule matching engine library and the logical relationship between the message fields.

[0067] In this embodiment, by utilizing a model that integrates a directed graph and a bidirectional recurrent neural network, the clause features of the content of each message field in the letter of credit data are obtained, and then based on the pre-generated review rule matching engine library, the clause feature specifications of each message field and the logical relationship between the message fields are reviewed. According to the clause features, each message field of the letter of credit data is reviewed, which can avoid manual review and effectively improve efficiency.

[0068] Furthermore, the device also includes:

[0069] The engine library generation module is used to pre-generate the letter of credit rule matching engine library based on letter of credit sample data and neural network model.

[0070] Furthermore, the engine library generation module includes:

[0071] A bidirectional recurrent neural network processing unit is used to obtain letter of credit sample data, and input the letter of credit sample data into the bidirectional recurrent neural network, and for each letter of credit sample data, obtain the central feature label words of each message domain output by the bidirectional recurrent neural network and the weight of each central feature label word;

[0072] A directed graph reasoning unit, used to form a directed graph structure based on the central feature label words of each message field, and to perform logical reasoning based on the directed graph structure to determine the logical relationship between the clause features in each message field;

[0073] The engine library generation unit is used to generate the verification rule matching engine library according to the weight of each central feature tag word and the logical relationship between the clause features in each telegram domain.

[0074] Furthermore, the bidirectional recurrent neural network processing unit includes:

[0075] A feature vector determination unit is used to map the content of each message field in the sample data of the letter of credit using a preset distributed word vector dictionary for any sample data of the letter of credit, so as to obtain a feature vector corresponding to each message field;

[0076] The central feature label word extraction unit is used for inputting the feature vector of the message domain into the bidirectional recurrent neural network for any message domain, and obtaining the central feature label word extracted by the bidirectional recurrent neural network;

[0077] The weight calculation unit is used to calculate the weight of each central feature tag word in the electronic message domain.

[0078] Furthermore, the interception module includes:

[0079] A correction unit, used for performing intelligent correction processing on the letter of credit data based on a preset noise channel model;

[0080] The interception unit is used to intercept the message field content of each message field from the letter of credit data according to the pre-agreed correspondence between the message field and the data structure of the letter of credit data.

[0081] Furthermore, the clause features include the central feature tag words contained in the telegram domain;

[0082] The audit modules include:

[0083] A coverage determination unit is used to compare the coverage of the central feature tag words of the message domain according to the clause feature specifications of each message domain in the verification rule matching engine library;

[0084] A content association determination unit, used to determine the content association of the central feature tag word of the message field according to the logical relationship between the message fields in the verification rule matching engine library;

[0085] The audit result determination unit is used to determine the audit result of each electronic message field of the letter of credit data based on the coverage and content association.

[0086] Furthermore, the device also includes:

[0087] Update the data acquisition module, which is used to obtain the existence weights corresponding to the clause features of each message field output by the neural network model based on the letter of credit data to be reviewed, and the association weights of the clause features of each message field;

[0088] The engine library update module is used to update the verification rule matching engine library based on the existence weight and the association weight.

[0089] Embodiment 3

[0090] Figure 5A schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present application is shown in FIG. Figure 3 As shown, the electronic device includes a processor 510, a memory 520, an input device 530 and an output device 540; the number of the processor 510 in the electronic device can be one or more. Figure 5 A processor 510 is taken as an example; the processor 510, the memory 520, the input device 530 and the output device 540 in the electronic device can be connected via a bus or other means. Figure 3 The example of connecting through bus is taken in the following.

[0091] The memory 520, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the letter of credit review method in the embodiment of the present invention (for example, the interception module 401, the clause feature output module 402, and the review module 403 in the letter of credit review device). The processor 510 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory 520, that is, realizes the above-mentioned letter of credit review method:

[0092] When receiving the letter of credit data to be reviewed, intercept the content of each message field in the letter of credit data;

[0093] For any message domain, the message domain content is input into a pre-trained neural network model to obtain the clause features of the message domain content output by the neural network model, where the neural network model is a model that integrates a directed graph and a bidirectional recurrent neural network;

[0094] Based on the pre-generated review rule matching engine library, the clause feature specifications of each message field and the logical relationship between each message field, each message field of the letter of credit data is reviewed according to the clause features.

[0095] Furthermore, the method also includes:

[0096] A credit review rule matching engine library is pre-generated based on letter of credit sample data and neural network models.

[0097] Furthermore, based on the letter of credit sample data and the neural network model, a letter of credit rule matching engine library is pre-generated, including:

[0098] Obtain letter of credit sample data, and input the letter of credit sample data into a bidirectional recurrent neural network. For each letter of credit sample data, obtain the central feature label words of each message domain output by the bidirectional recurrent neural network and the weight of each central feature label word;

[0099] Based on the central feature label words of each message field, a directed graph structure is formed, and logical reasoning is performed based on the directed graph structure to determine the logical relationship between the clause features in each message field;

[0100] The verification rule matching engine library is generated according to the weight of each central feature tag word and the logical relationship between the clause features in each telegram domain.

[0101] Furthermore, the letter of credit sample data is input into the bidirectional recurrent neural network. For each letter of credit sample data, the central feature label words of each message domain output by the bidirectional recurrent neural network and the weights of each central feature label word are obtained, including:

[0102] For any letter of credit sample data, the preset distributed word vector dictionary is used to map the message domain content of each message domain in the letter of credit sample data to obtain the feature vector corresponding to each message domain;

[0103] For any message domain, the feature vector of the message domain is input into the bidirectional recurrent neural network to obtain the central feature label word extracted by the bidirectional recurrent neural network;

[0104] Calculate the weight of each central feature tag word in the message domain.

[0105] Further, when the letter of credit data is received, the content of each message field in the letter of credit data is intercepted, including:

[0106] Based on the preset noise channel model, intelligent correction processing is performed on the letter of credit data;

[0107] According to the pre-agreed correspondence between the message domain and the data structure of the letter of credit data, the message domain content of each message domain is intercepted from the letter of credit data.

[0108] Furthermore, the clause features include the central feature tag words contained in the telegram domain;

[0109] Based on the pre-generated verification rule matching engine library, the clause feature specifications of each message field and the logical relationship between each message field, each message field of the letter of credit data is reviewed according to the clause features, including:

[0110] According to the verification rule matching engine library, the clause feature specifications of each message domain are compared with the coverage of the central feature tag words of the message domain;

[0111] According to the logical relationship between each message domain in the verification rule matching engine library, the content association of the central feature tag word of the message domain is determined;

[0112] The audit results of each electronic message field of the letter of credit data are determined based on the coverage and content relevance.

[0113] Furthermore, the method also includes:

[0114] Obtaining the existence weights corresponding to the clause features of each message field output by the neural network model based on the letter of credit data to be reviewed and the association weights of the clause features of each message field;

[0115] The verification rule matching engine library is updated based on the existence weight and association weight.

[0116] The memory 520 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory 520 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 520 may further include a memory remotely arranged relative to the processor 510, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0117] Embodiment 4

[0118] Embodiment 4 of the present application further provides a storage medium including computer executable instructions, which are used to execute a letter of credit review method when executed by a computer processor, the method comprising:

[0119] When receiving the letter of credit data to be reviewed, intercept the content of each message field in the letter of credit data;

[0120] For any message domain, the message domain content is input into a pre-trained neural network model to obtain the clause features of the message domain content output by the neural network model, where the neural network model is a model that integrates a directed graph and a bidirectional recurrent neural network;

[0121] Based on the pre-generated review rule matching engine library, the clause feature specifications of each message field and the logical relationship between each message field, each message field of the letter of credit data is reviewed according to the clause features.

[0122] Furthermore, the method also includes:

[0123] A credit review rule matching engine library is pre-generated based on letter of credit sample data and neural network models.

[0124] Furthermore, based on the letter of credit sample data and the neural network model, a letter of credit rule matching engine library is pre-generated, including:

[0125] Obtain letter of credit sample data, and input the letter of credit sample data into a bidirectional recurrent neural network. For each letter of credit sample data, obtain the central feature label words of each message domain output by the bidirectional recurrent neural network and the weight of each central feature label word;

[0126] Based on the central feature label words of each message field, a directed graph structure is formed, and logical reasoning is performed based on the directed graph structure to determine the logical relationship between the clause features in each message field;

[0127] The verification rule matching engine library is generated according to the weight of each central feature tag word and the logical relationship between the clause features in each telegram domain.

[0128] Furthermore, the letter of credit sample data is input into the bidirectional recurrent neural network. For each letter of credit sample data, the central feature label words of each message domain output by the bidirectional recurrent neural network and the weights of each central feature label word are obtained, including:

[0129] For any letter of credit sample data, the preset distributed word vector dictionary is used to map the message domain content of each message domain in the letter of credit sample data to obtain the feature vector corresponding to each message domain;

[0130] For any message domain, the feature vector of the message domain is input into the bidirectional recurrent neural network to obtain the central feature label word extracted by the bidirectional recurrent neural network;

[0131] Calculate the weight of each central feature tag word in the message domain.

[0132] Further, when the letter of credit data is received, the content of each message field in the letter of credit data is intercepted, including:

[0133] Based on the preset noise channel model, intelligent correction processing is performed on the letter of credit data;

[0134] According to the pre-agreed correspondence between the message domain and the data structure of the letter of credit data, the message domain content of each message domain is intercepted from the letter of credit data.

[0135] Furthermore, the clause features include the central feature tag words contained in the telegram domain;

[0136] Based on the pre-generated verification rule matching engine library, the clause feature specifications of each message field and the logical relationship between each message field, each message field of the letter of credit data is reviewed according to the clause features, including:

[0137] According to the verification rule matching engine library, the clause feature specifications of each message domain are compared with the coverage of the central feature tag words of the message domain;

[0138] According to the logical relationship between each message domain in the verification rule matching engine library, the content association of the central feature tag word of the message domain is determined;

[0139] The audit results of each electronic message field of the letter of credit data are determined based on the coverage and content relevance.

[0140] Furthermore, the method also includes:

[0141] Obtaining the existence weights corresponding to the clause features of each message field output by the neural network model based on the letter of credit data to be reviewed and the association weights of the clause features of each message field;

[0142] The verification rule matching engine library is updated based on the existence weight and association weight.

[0143] Of course, the storage medium containing computer executable instructions provided in the embodiment of the present application, whose computer executable instructions are not limited to the above method operations, can also execute related operations in the letter of credit review method provided in any embodiment of the present application.

[0144] Through the above description of the implementation method, the technicians in the relevant field can clearly understand that the present application can be implemented with the help of software and necessary general hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application can be essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of each embodiment of the present application.

[0145] It is worth noting that in the embodiment of the above-mentioned search device, the various units and modules included are only divided according to functional logic, but are not limited to the above-mentioned division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application.

[0146] Note that the above are only preferred embodiments of the present application and the technical principles used. Those skilled in the art will understand that the present application is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A method for reviewing a letter of credit, characterized in that: The method comprises: When receiving the letter of credit data to be reviewed, intercepting the content of each message field in the letter of credit data; For any electronic message domain, input the electronic message domain content into a pre-trained neural network model to obtain the clause features of the electronic message domain content output by the neural network model, wherein the neural network model is a model integrating a directed graph and a bidirectional recurrent neural network; Based on the clause feature specifications of each message field in the pre-generated verification rule matching engine library and the logical relationship between each message field, each message field of the letter of credit data is reviewed according to the clause features; The credit verification rule matching engine library is pre-generated based on the credit sample data and the neural network model, including: Obtain the letter of credit sample data, and input the letter of credit sample data into a bidirectional recurrent neural network, and for each letter of credit sample data, obtain the central feature label words of each message domain output by the bidirectional recurrent neural network and the weight of each central feature label word; A directed graph structure is formed based on the central feature label words of each message field, and logical reasoning is performed based on the directed graph structure to determine the logical relationship between the clause features in each message field; The verification rule matching engine library is generated according to the weight of each central feature tag word and the logical relationship between the clause features in each electronic message domain.

2. The method according to claim 1, characterized in that The step of inputting the letter of credit sample data into a bidirectional recurrent neural network and obtaining, for each piece of letter of credit sample data, a central feature label word of each message domain output by the bidirectional recurrent neural network and a weight of each central feature label word comprises: For any letter of credit sample data, use a preset distributed word vector dictionary to map the message domain content of each message domain in the letter of credit sample data to obtain a feature vector corresponding to each message domain; For any message domain, input the feature vector of the message domain into the bidirectional recurrent neural network to obtain the central feature label word extracted by the bidirectional recurrent neural network; The weight of each of the central feature tag words in the electronic message domain is calculated.

3. The method according to claim 1, characterized in that When the letter of credit data is received, intercepting the content of each message field in the letter of credit data includes: Based on a preset noise channel model, intelligent correction processing is performed on the letter of credit data; According to the pre-agreed correspondence between the message domain and the data structure of the letter of credit data, the message domain content of each message domain is intercepted from the letter of credit data.

4. The method according to claim 1, characterized in that: The clause features include the central feature tag words contained in the electronic text domain; The method of matching the clause feature specifications of each message field in the pre-generated verification rule matching engine library and the logical relationship between each message field and verifying each message field of the letter of credit data according to the clause feature includes: Compare the coverage of the central feature tag words of each message domain in the verification rule matching engine library according to the clause feature specifications of each message domain; Determine the content association of the central feature tag words of the message domain according to the logical relationship between the message domains in the verification rule matching engine library; The audit result of each electronic message field of the letter of credit data is determined based on the coverage and the content association.

5. The method according to claim 1, characterized in that The method further comprises: Obtaining the existence weights corresponding to the clause features of each message field output by the neural network model based on the letter of credit data to be reviewed and the association weights of the clause features of each message field; The verification rule matching engine library is updated based on the existence weight and the association weight.

6. A letter of credit review device, characterized in that: The device comprises: An interception module, for intercepting the content of each message field in the letter of credit data upon receiving the letter of credit data to be reviewed; A clause feature output module, for any telegram domain, inputting the content of the telegram domain into a pre-trained neural network model, and obtaining the clause features of the content of the telegram domain output by the neural network model, wherein the neural network model is a model integrating a directed graph and a bidirectional recurrent neural network; An audit module, for auditing each message field of the letter of credit data according to the terms and conditions, based on the pre-generated audit rule matching engine library, the terms and conditions of each message field and the logical relationship between the message fields; An engine library generation module, used to pre-generate the letter of credit rule matching engine library based on letter of credit sample data and the neural network model; The engine library generation module includes: A bidirectional recurrent neural network processing unit, used to obtain the letter of credit sample data, and input the letter of credit sample data into the bidirectional recurrent neural network, and for each letter of credit sample data, obtain the central feature label words of each message domain output by the bidirectional recurrent neural network and the weight of each central feature label word; A directed graph reasoning unit, used to form a directed graph structure based on the central feature label words of each electronic message field, and to perform logical reasoning based on the directed graph structure to determine the logical relationship between the clause features in each electronic message field; The engine library generation unit is used to generate the verification rule matching engine library according to the weight of each central feature tag word and the logical relationship between the clause features in each electronic message domain.

7. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the letter of credit review method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for reviewing a letter of credit as described in any one of claims 1 to 5 is implemented.