A method, device and medium for auxiliary audit of documents

Abstract

The present application provides an auxiliary audit method, device and medium for documents, which belongs to the field of data processing technology. The method obtains document information to be audited from a user terminal; based on a pre-trained logistic regression model, at least one rejection prediction sub-information sequence corresponding to the document information to be audited is determined. The rejection prediction sub-information sequence is obtained based on the document sub-information whose first rejection prediction probability value in the document information to be audited is greater than a preset first probability threshold. Based on the logistic regression model and the rejection prediction sub-information sequence, the second rejection prediction probability value corresponding to the document information to be audited is determined. When the second rejection prediction probability value is greater than the second probability threshold, an audit prompt information is generated, and the audit prompt information is sent to the corresponding audit terminal. And based on the rejection reason text set corresponding to the rejection prediction sub-information sequence, a rejection analysis text is generated, and the rejection analysis text is stored in the cloud server.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an auxiliary audit method, device and medium for documents. Background Art

[0002] With the development of corporate business, the efficiency and accuracy of document review has become an important assessment indicator, especially for the group's financial shared center, the workload of financial personnel in reviewing documents has increased sharply. Excessive document review work makes it easy for financial personnel to make mistakes during the review process, or fail to respond to the document submitter in time when the review is rejected.

[0003] At present, when a document is rejected, the document submitter often asks the reviewer offline or online to find out the reason for the rejection and ask for a solution. This method consumes a lot of review time, resulting in a poor user experience of the document review system, and cannot meet the user's demand for the intelligent document review system.

[0004] Based on this, there is an urgent need for a technical solution that can reduce the amount of manual review by document reviewers, improve review efficiency, and improve the intelligence level of the document review system. Summary of the invention

[0005] The embodiments of the present application provide an auxiliary audit method, device and medium for documents, which are used to solve the current problems of large audit volume, low audit efficiency and low intelligence level of document audit systems among document audit workers.

[0006] On the one hand, an embodiment of the present application provides a method for assisting in reviewing a document, the method comprising:

[0007] Acquire the document information to be reviewed from the user terminal; the document information to be reviewed at least includes a header, document transaction content, and a document reviewer;

[0008] Based on a pre-trained logistic regression model, at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed is determined; the rejection prediction sub-information sequence is obtained based on document sub-information whose first rejection prediction probability value in the document information to be reviewed is greater than a preset first probability threshold; the document sub-information is obtained by comparing a preset historical keyword set with the document information to be reviewed;

[0009] Determine a second rejection prediction probability value corresponding to the to-be-reviewed document information based on the logistic regression model and the rejection prediction sub-information sequence;

[0010] When the second rejection prediction probability value is greater than a second probability threshold, generating audit prompt information, and sending the audit prompt information to a corresponding audit terminal; and

[0011] A rejection analysis text is generated according to the rejection reason text set corresponding to the rejection prediction sub-information sequence, and the rejection analysis text is stored in a cloud server.

[0012] In one implementation of the present application, before determining at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed based on a pre-trained logistic regression model, the method further includes:

[0013] Obtain some historical rejection document information and corresponding rejection reason text data sets;

[0014] Using a preset TF-IDF model, respectively determine the word frequency and inverse text frequency index of the words in the rejection reason text data set;

[0015] When the product value of the word frequency and the inverse text frequency index of the word is greater than a preset value, the word is used as a keyword of the rejection reason text data set;

[0016] Based on the obtained keywords and bag-of-words model, keyword feature vectors corresponding to the rejection reason text data of the rejection reason text data set are generated, so as to train the logistic regression model according to the keyword feature vectors and the historical rejection document information.

[0017] In one implementation of the present application, the logistic regression model is trained according to each of the keyword feature vectors and each of the historical rejected document information, specifically including:

[0018] Taking any one of the keyword feature vectors as a feature vector to be associated;

[0019] Calculating the cosine similarity between the feature vector to be associated and each of the other keyword feature vectors in sequence, and generating a set of associated feature vectors corresponding to the feature vector to be associated according to the comparison result of the cosine similarity and a first preset threshold value; one set of associated feature vectors corresponds to one preset historical rejection reason;

[0020] Determining the occurrence frequency of each keyword feature vector in each associated feature vector set according to each associated feature vector set obtained, so as to determine the occurrence probability value corresponding to each keyword feature vector according to the occurrence frequency;

[0021] The document sub-information of each of the historical rejected document information and its corresponding keyword feature vectors and the occurrence probability values ​​are added to the data dictionary as model training samples to train the logistic regression model.

[0022] In one implementation of the present application, training the logistic regression model specifically includes:

[0023] Inputting the associated feature vector sets and the corresponding occurrence probability values ​​corresponding to each of the document sub-information in the data dictionary into the to-be-trained logistic regression model in sequence to train the to-be-trained logistic regression model;

[0024] When the to-be-trained logistic regression model is being trained, the corresponding model parameter values ​​are updated by a gradient descent algorithm until it is determined that the corresponding model parameter values ​​make the function value of the logistic regression cost function less than a second preset threshold, thereby obtaining the trained logistic regression model.

[0025] In one implementation of the present application, the method further includes:

[0026] Obtain real-time updated information of each document to be reviewed and the corresponding rejection reason text;

[0027] The information of each document to be reviewed and the corresponding rejection reason text are added to the preset database to update the model training samples and retrain the logistic regression model.

[0028] In one implementation of the present application, generating a rejection analysis text according to a rejection reason text set corresponding to the rejection prediction sub-information sequence specifically includes:

[0029] Inputting the rejection reason text set into a preset analysis text generation model to combine the rejection reason text words; the analysis text generation model is trained based on a number of rejection reason text word samples and their corresponding rejection reason sentences;

[0030] The rejected analysis text is determined according to the output result of the analysis text generation model.

[0031] In one implementation of the present application, the method further includes:

[0032] In response to a rejection operation of the audit terminal, prompt information corresponding to the rejection analysis text is sent to the audit terminal; the prompt information includes a control for viewing the rejection analysis text;

[0033] In the case where the audit terminal does not add the rejection analysis text to the rejection reason text box of the rejection operation, sending the prompt information to the user terminal so that the user terminal can view the rejection analysis text;

[0034] When the review terminal adds a rejection reason description text in the rejection reason text box, the rejection reason description text is compared with the rejection analysis text, so as to send the rejection reason description text and / or the prompt information to the user terminal according to the text comparison result.

[0035] In one implementation of the present application, sending the rejection reason description text and / or the prompt information to the user terminal according to the text comparison result specifically includes:

[0036] Calculating the text feature vectors corresponding to the rejection reason description text and the rejection analysis text, and the corresponding text similarity respectively; the text similarity is cosine similarity;

[0037] When the text similarity is greater than a similarity threshold, sending the prompt information to the user terminal;

[0038] When the text similarity is less than or equal to the similarity threshold, the rejection reason description text is sent to the user terminal, and the rejection reason description text and the document information to be reviewed are added to a preset database to update the model training sample and retrain the logistic regression model.

[0039] On the other hand, an embodiment of the present application further provides an auxiliary audit device for documents, the device comprising:

[0040] at least one processor; and a memory in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0041] Acquire the document information to be reviewed from the user terminal; the document information to be reviewed at least includes a header, document transaction content, and a document reviewer;

[0042] Based on a pre-trained logistic regression model, at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed is determined; the rejection prediction sub-information sequence is obtained based on document sub-information whose first rejection prediction probability value in the document information to be reviewed is greater than a preset first probability threshold; the document sub-information is obtained by comparing a preset historical keyword set with the document information to be reviewed;

[0043] Determine a second rejection prediction probability value corresponding to the to-be-reviewed document information based on the logistic regression model and the rejection prediction sub-information sequence;

[0044] When the second rejection prediction probability value is greater than a second probability threshold, generating audit prompt information, and sending the audit prompt information to a corresponding audit terminal; and

[0045] A rejection analysis text is generated according to the rejection reason text set corresponding to the rejection prediction sub-information sequence, and the rejection analysis text is stored in a cloud server.

[0046] On the other hand, the embodiment of the present application further provides a non-volatile computer storage medium for auxiliary review of documents, storing computer executable instructions, wherein the computer executable instructions are configured as follows:

[0047] Acquire the document information to be reviewed from the user terminal; the document information to be reviewed at least includes a header, document transaction content, and a document reviewer;

[0048] Based on a pre-trained logistic regression model, at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed is determined; the rejection prediction sub-information sequence is obtained based on document sub-information whose first rejection prediction probability value in the document information to be reviewed is greater than a preset first probability threshold; the document sub-information is obtained by comparing a preset historical keyword set with the document information to be reviewed;

[0049] Determine a second rejection prediction probability value corresponding to the to-be-reviewed document information based on the logistic regression model and the rejection prediction sub-information sequence;

[0050] When the second rejection prediction probability value is greater than a second probability threshold, generating audit prompt information, and sending the audit prompt information to a corresponding audit terminal; and

[0051] A rejection analysis text is generated according to the rejection reason text set corresponding to the rejection prediction sub-information sequence, and the rejection analysis text is stored in a cloud server.

[0052] Through the above technical solution, the present application can predict the rejection probability of documents through the logistic regression model, so as to classify the submitted documents, so that the auditors can clearly see the documents with rejection risks, and then focus on reviewing these documents. At the same time, the logistic regression model can automatically capture and learn key feature variables, improve the accuracy of intelligent prediction, and reduce the risk of missed judgments. This solves the current problems of large audit volume, low audit efficiency, and low intelligence level of document audit systems for document auditors.

[0053] In addition, this application can also provide a rejection analysis text for reviewers or document submitters to review, without the need to spend more manpower to generate or obtain the rejection analysis text, which can improve the user experience of the document review system. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0055] Figure 1 A schematic diagram of a flow chart of an auxiliary audit method for documents in an embodiment of the present application;

[0056] Figure 2 It is a structural schematic diagram of an auxiliary audit device for documents in an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0058] As for the intelligentization of the document review process, there are currently algorithms such as support vector machines, random forests, and decision trees that are applied to intelligent document review. However, these algorithms have low computational efficiency and a limited scope of applicable document review. In some cases, manual participation is still required in some review work, such as the reviewer does not give a reason for rejection, and the document submitter cannot know the reason for rejection in a timely manner. This makes the current document review system unable to meet the user's demand for an intelligent review process.

[0059] Based on this, the embodiments of the present application provide an auxiliary audit method, device and medium for documents, which are used to solve the current problems of large audit volume, low audit efficiency and low intelligence level of document audit systems for document audit workers.

[0060] The following describes in detail various embodiments of the present application in conjunction with the accompanying drawings.

[0061] The present application embodiment provides a method for assisting in reviewing documents, such as Figure 1 As shown, the method may include steps S101-S105:

[0062] S101, the server obtains the document information to be reviewed from the user terminal.

[0063] The document information to be reviewed shall at least include the header, document transaction content, and document reviewer.

[0064] It should be noted that the server, as the executor of the auxiliary audit method for documents, exists only as an example, and the executor is not limited to the server, and this application does not make any specific limitation on this.

[0065] The server may be a server running the document review system or connected to a server running the document review system, and the server may establish a connection with a user terminal for document submission and an audit terminal for document review. The user terminal may be a mobile phone, computer or other device of a document submitter, such as a laptop computer of a financial reimbursement staff, and may submit document information to be reviewed through the document review system.

[0066] The document information to be reviewed includes the table header, document transaction content and document reviewer. The table header includes the document type. The document transaction content refers to the specific content in the document information to be reviewed, such as the reimbursement amount, reimbursement item type, and reimbursement item quantity for financial reimbursement documents. The document reviewer can be the person who is responsible for reviewing the document information to be reviewed after the document submitter submits the document information to be reviewed.

[0067] S102: The server determines at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed based on a pre-trained logistic regression model.

[0068] The rejection prediction sub-information sequence is obtained based on the document sub-information whose first rejection prediction probability value in the document information to be reviewed is greater than the preset first probability threshold. The document sub-information is obtained by comparing the preset historical keyword set with the document information to be reviewed. The first probability threshold can be specified by the user, and this application does not make specific restrictions on this. Since the document information to be reviewed may have multiple different types of documents, one document information to be reviewed may contain multiple rejection prediction sub-information sequences.

[0069] The above-mentioned first rejection prediction probability value can be understood as the probability of a document being rejected when the document sub-information appears in the document information to be reviewed. The first probability threshold is used to screen part of the document sub-information, that is, when the first rejection prediction probability value is greater than the first probability threshold, it is added to the rejection prediction sub-information sequence. For example, the first rejection prediction probability value of the document sub-information corresponding to a certain expense item is 80%, and the above-mentioned pre-set first probability threshold is 60%. The document sub-information corresponding to the expense item is added to the rejection prediction sub-information sequence; the first rejection prediction probability value of the document sub-information corresponding to another document itinerary selection is 30%, and the above-mentioned pre-set first probability threshold is 60%. The document sub-information corresponding to the document itinerary selection is not added to the rejection prediction sub-information sequence. The specific value of the above-mentioned first probability threshold is only for example, and the first probability threshold can be set during actual use. This application does not make any specific limitation on this specific value.

[0070] This application uses a logistic regression model to process the document information to be reviewed, in which the rejection prediction sub-information sequence can be understood as the pre-divided document sub-information that complies with the above rules in the document information to be reviewed. Document sub-information includes sub-information such as the amount, item, date, sensitive words in the summary, etc. The specific rejection prediction sub-information sequence includes feature vectors such as the expense item should be selected as travel expenses, there are duplicate invoices, and the itinerary is not a closed loop. The document sub-information corresponds to a rejection probability (first rejection prediction probability value) as shown in the following table. The above-mentioned first probability threshold is set by the user, and this application does not make specific restrictions on this.

[0071]

[0072] In the embodiment of the present application, before determining at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed based on the pre-trained logistic regression model, the method further includes:

[0073] The server obtains some historical rejection document information and the corresponding rejection reason text data set. Then, the word frequency and inverse text frequency index of the words in the rejection reason text data set are determined by the preset TF-IDF model. And when the product value of the word frequency and the inverse text frequency index of the word is greater than the preset value, the word is used as the keyword of the rejection reason text data set. Based on the obtained keywords and the bag-of-words model, the keyword feature vectors corresponding to the rejection reason text data of the rejection reason text data set are generated, so as to train the logistic regression model according to the keyword feature vectors and the historical rejection document information.

[0074] In other words, the server can obtain the training samples of the logistic regression model from the preset database pre-set by the user, including several historical rejection document information and corresponding rejection reason text data sets. The rejection reason text data set is composed of the historical rejection reason text data of several historical rejection document information. The historical rejection reason text data can be pre-generated by the user, or filled in by the document reviewer in actual use. This application does not make specific restrictions on this. Using the TF-IDF model, the server can first count the frequency of each word appearing in the rejection reason text data set in the same rejection reason text data, and calculate the inverse text frequency index of each word in all rejection reason text data. According to the product of the word frequency and the inverse text frequency index, the TF-IDF value of each word can be obtained, and then the product value is compared with the preset value. The server selects words greater than the preset value as keywords in the rejection reason text data set. Among them, the preset value is set by the user during actual use, and this application does not make specific restrictions on this.

[0075] After obtaining the keywords of the rejection reason text data set, the server can also use the Bag-of-Words model (BOW) to obtain the keyword feature vectors of the keywords contained in each rejection reason text data. For example, if a rejection reason text data is "the amount exceeds the reimbursement index", the keywords include "exceed" and "index", and its keyword feature vector is {1, 2}.

[0076] Subsequently, the server trains the logistic regression model based on the feature vectors of each keyword and the information of each historical rejected document, including:

[0077] The server uses any keyword feature vector among the keyword feature vectors as the feature vector to be associated. The cosine similarity between the feature vector to be associated and the other keyword feature vectors is calculated in sequence, so as to generate a set of associated feature vectors corresponding to the feature vector to be associated based on the comparison result of the cosine similarity and the first preset threshold. A set of associated feature vectors corresponds to a preset historical rejection reason. Based on the obtained sets of associated feature vectors, the frequency of occurrence of each keyword feature vector in each set of associated feature vectors is determined, so as to determine the probability of occurrence value corresponding to each keyword feature vector according to the frequency of occurrence. The document sub-information of each historically rejected document information and its corresponding keyword feature vectors and probability of occurrence values ​​are added to the data dictionary as model training samples to train the logistic regression model.

[0078] That is to say, the rejection reason text data set includes multiple rejection reason text data, each rejection reason text data corresponds to a keyword feature vector, and the server can randomly specify a keyword feature vector as the feature vector to be associated, and then obtain the associated feature vector set by calculating the cosine similarity between the feature vector to be associated and other keyword feature vectors. Specifically, when the cosine similarity between the feature vector to be associated and another keyword feature vector is greater than the first preset threshold, it means that the feature vector to be associated is similar to another keyword feature vector, which means that the rejection reason is associated, that is, the two feature vectors are added to the associated feature vector set. The first preset threshold can be set according to actual use, and the present application does not make specific restrictions on this. Since the rejection reasons are associated, the keyword feature vectors in the same associated feature vector set can correspond to the same or several rejection reasons, that is, the same rejection reason may correspond to multiple keyword feature vectors.

[0079] The server can also count the number of occurrences of the keyword feature vector in different associated feature vector sets, and calculate the ratio of the number of occurrences and the associated feature vector set as the occurrence probability value corresponding to the keyword feature vector, that is, the occurrence probability of the rejection reason corresponding to the keyword feature vector, which also corresponds to the probability value of the document being rejected due to the rejection reason. The server can also use the above occurrence probability value as the first rejection prediction probability value corresponding to the document sub-information based on the corresponding relationship between the rejection reason and the document sub-information. The rejection reason is such as document sub-information A is wrong, or document sub-information B is wrong, etc.

[0080] Afterwards, the server may generate a data dictionary including the document sub-information, the keyword feature vector corresponding to the document sub-information, and the occurrence probability value of the keyword feature vector for training the logistic regression model. The training of the logistic regression model specifically includes:

[0081] The server inputs the associated feature vector sets and the corresponding occurrence probability values ​​corresponding to each document sub-information in the data dictionary into the logistic regression model to be trained in sequence, so as to train the logistic regression model to be trained. When the logistic regression model to be trained is trained, the corresponding model parameter values ​​are updated by the gradient descent algorithm until the corresponding model parameter values ​​are determined to make the function value of the logistic regression cost function less than the second preset threshold, thereby obtaining a trained logistic regression model.

[0082] The server uses the logistic regression algorithm to build a model based on the feature variables and the document sub-information (feature value) data dictionary. The modeling data set is split into a training set and a corresponding test set for training and evaluation. The specific implementation logic example is as follows: First, assume that there are n keyword feature vectors X=[x1, x2,…xn], where xi is the i-th keyword feature vector. Logistic regression assumes that the probability function is hθ(x)=g(θ^T*x), where hθ(x) is the probability of occurrence of the keyword feature variable x, and θ is the parameter value of the model. g(z) is the sigmoid function g(z)=1 / (1+e^(-z)), z is θ^T*x, and θ^T represents the transpose of θ. In addition, a logistic regression cost function is used, and Log is used to measure the difference between the actual value and the predicted value, J(θ)=-1 / m*Σ[y*log(hθ(x))+(1-y)*log(1-hθ(x))], J(θ) is the cost function, m is the number of training samples, y is the actual label value (0 or 1), and hθ(x) is the estimated value of the model prediction: In order to minimize the cost function J(θ), a gradient descent algorithm is used to estimate the optimal parameter value θ. By iteratively updating the parameter θ, the model is continuously adjusted until it reaches the optimal value, even if the function value of the logistic regression cost function is less than the second preset threshold, a trained logistic regression model is obtained. The second preset threshold is set by the user, and this application does not make any specific restrictions on this.

[0083] Through the above-trained logistic regression model, the server can calculate the first rejection prediction probability value of each document sub-information in the document information to be reviewed, and then generate the above-mentioned rejection prediction sub-information sequence by comparing the result with the first probability threshold.

[0084] S103, the server determines a second rejection prediction probability value corresponding to the document information to be reviewed based on the logistic regression model and the rejection prediction sub-information sequence.

[0085] Through the above-mentioned logistic regression model, the weighted average of the first rejection prediction probability values ​​corresponding to each rejection prediction sub-information sequence can be calculated. The weight can be set according to the different rejection reasons during the training process of the above-mentioned logistic regression model, that is, different weights are set for different associated feature vector sets. The second rejection prediction probability value is calculated according to the sum of the product values ​​of each weight and the first rejection prediction probability value.

[0086] S104: When the second rejection prediction probability value is greater than the second probability threshold, the server generates audit prompt information and sends the audit prompt information to the corresponding audit terminal.

[0087] The second probability threshold can be set according to actual use, and this application does not make specific restrictions on this. When the second rejection prediction probability value is greater than the second probability threshold, the server can generate prompt information for prompting the review terminal to focus on the review of the corresponding reviewer, and mark the document sub-information corresponding to the second rejection prediction probability value to generate prompt information.

[0088] S105, the server generates a rejection analysis text according to the rejection reason text set corresponding to the rejection prediction sub-information sequence, and stores the rejection analysis text to the cloud server.

[0089] The cloud server can be connected to the server to store the rejection analysis text.

[0090] In an embodiment of the present application, the method further includes:

[0091] The server obtains the real-time updated information of each document to be reviewed and the corresponding rejection reason text. The information of each document to be reviewed and the corresponding rejection reason text are added to the preset database to update the model training samples and retrain the logistic regression model.

[0092] That is to say, if the document information to be reviewed is rejected, the server can retrain the logistic regression model based on the rejected document information to be reviewed and the rejection reason text corresponding to the document information to be reviewed obtained by processing the logistic regression model to ensure that the logistic regression model remains optimal.

[0093] In the embodiment of the present application, the above-mentioned generation of the rejection analysis text according to the rejection reason text set corresponding to the rejection prediction sub-information sequence specifically includes:

[0094] The server inputs the rejection reason text set into a preset analysis text generation model to combine the rejection reason text words. The analysis text generation model is trained based on a number of rejection reason text word samples and their corresponding rejection reason sentences. According to the output result of the analysis text generation model, the rejection analysis text is determined.

[0095] The analysis text generation model can be a neural network model, which is trained by a number of rejection reason text word samples and the rejection reason sentences corresponding to each sample. It can recognize the rejection reason text and output the corresponding rejection reason sentence, and use the output rejection reason sentence as the rejection analysis text.

[0096] In addition, this application can:

[0097] In response to the rejection operation of the audit terminal, a prompt message corresponding to the rejection analysis text is sent to the audit terminal. The prompt message includes a control for viewing the rejection analysis text. In the case where the audit terminal does not add the rejection analysis text to the rejection reason text box of the rejection operation, the prompt message is sent to the user terminal so that the user terminal can view the rejection analysis text. In the case where the audit terminal adds a rejection reason description text in the rejection reason text box, the rejection reason description text is compared with the rejection analysis text, so as to send the rejection reason description text and / or prompt message to the user terminal according to the text comparison result.

[0098] That is, the audit terminal can perform a rejection operation. After obtaining the second rejection prediction probability value, the audit terminal user can view the pending document information with the audit prompt information, and click the rejection control to reject the pending document information if the rejection condition is met. At this time, the server can obtain the rejection analysis text and generate a control that can view the rejection analysis text. For example, the user can view the rejection analysis text by clicking, sliding, etc. on the control.

[0099] If the audit terminal does not actively add the rejection analysis text when rejecting the document information to be reviewed, the server can send a prompt message to the user terminal so that the user terminal can access the cloud server to view the rejection analysis text. If the audit terminal actively adds the rejection reason description text, the rejection analysis text can be further compared with the rejection reason description text, and then the rejection reason description text can be sent and / or prompt information can be given.

[0100] According to the text comparison result, a rejection reason description text and / or prompt information is sent to the user terminal, specifically including:

[0101] The server calculates the text feature vectors corresponding to the rejection reason description text and the rejection analysis text, and the corresponding text similarity. The text similarity is cosine similarity. When the text similarity is greater than the similarity threshold, a prompt message is sent to the user terminal. When the text similarity is less than or equal to the similarity threshold, the rejection reason description text is sent to the user terminal, and the rejection reason description text and the document information to be reviewed are added to the preset database to update the model training samples and retrain the logistic regression model.

[0102] In other words, when the rejection reason description text is not similar to the rejection analysis text given by the above-mentioned logistic regression model, the server can send the rejection reason description text to the user terminal, and perform feedback training of the logistic regression model through the rejection reason description text to ensure the freshness of the logistic regression model.

[0103] Through the above technical solution, the present application can predict the rejection probability of documents through the logistic regression model, so as to classify the submitted documents, so that the auditors can clearly see the documents with rejection risks, and then focus on reviewing these documents. At the same time, the logistic regression model can automatically capture and learn key feature variables, improve the accuracy of intelligent prediction, and reduce the risk of missed judgments. This solves the current problems of large audit volume, low audit efficiency, and low intelligence level of document audit systems for document auditors.

[0104] In addition, this application can also provide a rejection analysis text for reviewers or document submitters to review, without the need to spend more manpower to generate or obtain the rejection analysis text, which can improve the user experience of the document review system.

[0105] Figure 2 A schematic diagram of the structure of an auxiliary audit device for documents provided in an embodiment of the present application, such as Figure 2 As shown, the device includes:

[0106] At least one processor; and a memory in communication with the at least one processor. The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0107] Obtain the document information to be reviewed from the user terminal. The document information to be reviewed includes at least the header, the document transaction content, and the document reviewer. Based on the pre-trained logistic regression model, determine at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed. The rejection prediction sub-information sequence is obtained based on the document sub-information whose first rejection prediction probability value in the document information to be reviewed is greater than the preset first probability threshold. The document sub-information is obtained by comparing the document information to be reviewed with a preset historical keyword set. Based on the logistic regression model and the rejection prediction sub-information sequence, determine the second rejection prediction probability value corresponding to the document information to be reviewed. When the second rejection prediction probability value is greater than the second probability threshold, generate an audit prompt message, and send the audit prompt message to the corresponding audit terminal. And generate a rejection analysis text based on the rejection reason text set corresponding to the rejection prediction sub-information sequence, and store the rejection analysis text to the cloud server.

[0108] The embodiment of the present application also provides a non-volatile computer storage medium for auxiliary review of documents, which stores computer executable instructions, and the computer executable instructions are set to:

[0109] Obtain the document information to be reviewed from the user terminal. The document information to be reviewed includes at least the header, the document transaction content, and the document reviewer. Based on the pre-trained logistic regression model, determine at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed. The rejection prediction sub-information sequence is obtained based on the document sub-information whose first rejection prediction probability value in the document information to be reviewed is greater than the preset first probability threshold. The document sub-information is obtained by comparing the document information to be reviewed with a preset historical keyword set. Based on the logistic regression model and the rejection prediction sub-information sequence, determine the second rejection prediction probability value corresponding to the document information to be reviewed. When the second rejection prediction probability value is greater than the second probability threshold, generate an audit prompt message, and send the audit prompt message to the corresponding audit terminal. And generate a rejection analysis text based on the rejection reason text set corresponding to the rejection prediction sub-information sequence, and store the rejection analysis text to the cloud server.

[0110] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0111] The devices, media and methods provided in the embodiments of the present application correspond one to one, and therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0112] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0113] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for assisting in reviewing documents, characterized in that: The method comprises: Acquire the document information to be reviewed from the user terminal; the document information to be reviewed at least includes a header, document transaction content, and a document reviewer; Based on a pre-trained logistic regression model, at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed is determined; the rejection prediction sub-information sequence is obtained based on document sub-information whose first rejection prediction probability value in the document information to be reviewed is greater than a preset first probability threshold; the document sub-information is obtained by comparing a preset historical keyword set with the document information to be reviewed; Determine a second rejection prediction probability value corresponding to the to-be-reviewed document information based on the logistic regression model and the rejection prediction sub-information sequence; When the second rejection prediction probability value is greater than a second probability threshold, generating audit prompt information, and sending the audit prompt information to a corresponding audit terminal; and Generate a rejection analysis text according to the rejection reason text set corresponding to the rejection prediction sub-information sequence, and store the rejection analysis text in a cloud server; Wherein, before determining at least one rejection prediction sub-information sequence corresponding to the document information to be reviewed based on the pre-trained logistic regression model, the method further includes: Obtain some historical rejection document information and corresponding rejection reason text data sets; Using a preset TF-IDF model, respectively determine the word frequency and inverse text frequency index of the words in the rejection reason text data set; When the product value of the word frequency and the inverse text frequency index of the word is greater than a preset value, the word is used as a keyword of the rejection reason text data set; Based on the obtained keywords and bag-of-words model, generate keyword feature vectors corresponding to the rejection reason text data of the rejection reason text data set, so as to train the logistic regression model according to the keyword feature vectors and the historical rejection document information; The logistic regression model is trained according to each of the keyword feature vectors and each of the historical rejected document information, specifically including: Taking any one of the keyword feature vectors as a feature vector to be associated; Calculating the cosine similarity between the feature vector to be associated and each of the other keyword feature vectors in sequence, and generating a set of associated feature vectors corresponding to the feature vector to be associated according to the comparison result of the cosine similarity and a first preset threshold value; one set of associated feature vectors corresponds to one preset historical rejection reason; Determining the occurrence frequency of each keyword feature vector in each associated feature vector set according to each associated feature vector set obtained, so as to determine the occurrence probability value corresponding to each keyword feature vector according to the occurrence frequency; The document sub-information of each of the historical rejected document information and the corresponding keyword feature vectors and the occurrence probability values ​​are added to the data dictionary as model training samples to train the logistic regression model; Wherein, determining the second rejection prediction probability value corresponding to the document information to be reviewed specifically includes: Different weights are set for different sets of associated feature vectors, and the second rejection prediction probability value is calculated based on the sum of the product values ​​of each weight and the first rejection prediction probability value.

2. The auxiliary audit method for documents according to claim 1, characterized in that: The logistic regression model is trained, specifically including: Inputting the associated feature vector sets and the corresponding occurrence probability values ​​corresponding to each of the document sub-information in the data dictionary into the to-be-trained logistic regression model in sequence to train the to-be-trained logistic regression model; When the to-be-trained logistic regression model is being trained, the corresponding model parameter values ​​are updated by a gradient descent algorithm until it is determined that the corresponding model parameter values ​​make the function value of the logistic regression cost function less than a second preset threshold, thereby obtaining the trained logistic regression model.

3. The auxiliary audit method for documents according to claim 1, characterized in that: The method further comprises: Obtain real-time updated information of each document to be reviewed and the corresponding rejection reason text; The information of each document to be reviewed and the corresponding rejection reason text are added to the preset database to update the model training samples and retrain the logistic regression model.

4. The auxiliary audit method for documents according to claim 1, characterized in that: Generate a rejection analysis text according to the rejection reason text set corresponding to the rejection prediction sub-information sequence, specifically including: Inputting the rejection reason text set into a preset analysis text generation model to combine the rejection reason text words; the analysis text generation model is trained based on a number of rejection reason text word samples and their corresponding rejection reason sentences; The rejected analysis text is determined according to the output result of the analysis text generation model.

5. The auxiliary audit method for documents according to claim 4, characterized in that: The method further comprises: In response to a rejection operation of the audit terminal, prompt information corresponding to the rejection analysis text is sent to the audit terminal; the prompt information includes a control for viewing the rejection analysis text; In the case where the audit terminal does not add the rejection analysis text to the rejection reason text box of the rejection operation, sending the prompt information to the user terminal so that the user terminal can view the rejection analysis text; When the review terminal adds a rejection reason description text in the rejection reason text box, the rejection reason description text is compared with the rejection analysis text to send the rejection reason description text and / or the prompt information to the user terminal based on the text comparison result.

6. The auxiliary audit method for documents according to claim 1, characterized in that: According to the text comparison result, sending the rejection reason description text and / or the prompt information to the user terminal specifically includes: Calculating the text feature vectors corresponding to the rejection reason description text and the rejection analysis text, and the corresponding text similarity respectively; the text similarity is cosine similarity; When the text similarity is greater than a similarity threshold, sending the prompt information to the user terminal; When the text similarity is less than or equal to the similarity threshold, the rejection reason description text is sent to the user terminal, and the rejection reason description text and the document information to be reviewed are added to a preset database to update the model training sample and retrain the logistic regression model.

7. An auxiliary audit device for documents, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute an auxiliary audit method for documents as described in any one of claims 1 to 6.

8. A non-volatile computer storage medium for assisting the audit of documents, storing computer executable instructions, characterized in that: The computer executable instructions can execute an auxiliary audit method for documents as described in any one of claims 1 to 6 above.

Images