A document review method, device and medium

Abstract

The embodiment of the present application discloses a document review method, device and medium, which belongs to the field of data processing technology and solves the problem of low accuracy of document review and serious false alarm and omission. It includes: in the acquired document, determining the document main table information and multiple sub-table information, generating a document continuous sequence with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information; determining the field type in the document continuous sequence, encoding the document continuous sequence based on different field types, and obtaining a document vector; configuring the required field information based on the approval conclusion corresponding to the acquired document, and determining the field feature combination vector; training the preset model based on the document vector and the field feature combination vector to obtain a document review model; inputting the acquired document to be reviewed into the document review model, so as to output the review conclusion corresponding to the document to be reviewed through the document review model.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a document review method, device and medium. Background Art

[0002] In today's rapidly developing business environment, with the continuous expansion of enterprise scale and the increasing complexity of business, Enterprise Resource Planning (ERP) system has become an indispensable infrastructure to support efficient operation and decision-making of enterprises. However, although the ERP system has greatly optimized the internal management process of enterprises, its core audit link is still highly dependent on manual operation, which to a certain extent restricts the improvement of the overall operational efficiency of enterprises and the flexibility to respond to market changes.

[0003] At present, the audit process in enterprise management software generally faces multiple challenges. First, with the surge in business volume and the diversification of document types, the workload of manual audit has increased dramatically, which not only increases the work pressure of employees, but also may lead to the extension of the audit cycle and affect the timeliness of business processing. Secondly, manual audit relies on the professional knowledge, experience and judgment of auditors, while the knowledge structure and skill level of employees in the enterprise are uneven, making it difficult to ensure the efficiency of audit, and there is a large risk of subjective error, resulting in low accuracy of document audit and serious false positives and omissions. Summary of the invention

[0004] The embodiments of the present application provide a document review method, device and medium for solving the following technical problems: the knowledge structure and skill level of employees within an enterprise are uneven, which makes it difficult to ensure the review efficiency, and there is a large risk of subjective errors, resulting in a low accuracy rate of document review and serious false positives and omissions.

[0005] The present application embodiment adopts the following technical solutions:

[0006] The embodiment of the present application provides a document review method, which includes: determining the document main table information and multiple sub-table information in the acquired document, generating a document continuous sequence with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information; determining the field type in the document continuous sequence, encoding the document continuous sequence based on different field types, and obtaining a document vector; configuring the required field information based on the approval conclusion corresponding to the acquired document, and determining a field feature combination vector; training a preset model based on the document vector and the field feature combination vector, and obtaining a document review model; inputting the acquired document to be reviewed into the document review model, so as to output the review conclusion corresponding to the document to be reviewed through the document review model.

[0007] The embodiment of the present application uses an improved model for document review, and realizes the purpose of automatic approval of the entire approval process from collecting document data, performing preprocessing, using supervised learning encoding to decoding, and outputting approval conclusions and error fields when approval is passed or not, thereby improving the efficiency and accuracy of the review and liberating human resources. The automated review model is based on preset rules and algorithms, and can more accurately identify and process document information, reduce errors caused by human factors, and improve the accuracy and reliability of the review. The embodiment of the present application processes a complex document structure containing a main table and multiple sub-tables, and by constructing a continuous sequence of documents with a hierarchical relationship, it can more comprehensively capture the information in the document, thereby supporting more complex business review requirements.

[0008] In one implementation of the present application, before generating a continuous sequence of documents with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information, the method also includes: determining a first label set and a second label set in the document in the acquired document; wherein the first label set is related to the document user information, and the second label set is related to the document event; determining a first document set corresponding to the same user in the first label set; determining the second labels corresponding to the documents in the first document set, respectively, comparing the second labels with a preset label mapping table, and determining the inclusion relationship between the second labels; wherein the preset label mapping table includes multiple labels, and also includes the inclusion relationship between the multiple labels; based on the inclusion relationship, determining the association relationship between the main table information and the multiple sub-table information.

[0009] In one implementation of the present application, a continuous sequence of documents with a hierarchical relationship is generated based on the association relationship between the main table information and the multiple sub-table information, specifically including: constructing multiple sub-table sequences based on the multiple sub-table information; wherein the multiple sub-table sequences are respectively provided with a unique identifier; combining the main table information with each sub-table sequence, and the unique identifiers of each sub-table sequence, to obtain multiple independent document sequences; determining the order between the multiple sub-table sequences based on the timestamps of the multiple sub-table information, and sorting the multiple independent document sequences to obtain a continuous sequence of documents with a hierarchical relationship.

[0010] In one implementation of the present application, the field type in the continuous sequence of documents is determined, and the continuous sequence of documents is encoded respectively based on different field types to obtain a document vector, specifically including: when the field type is a numeric field, the numeric field is used as the first vector; when the field type is an enumeration field, the enumeration value is mapped to a dense vector space through a pre-trained embedding layer to obtain a second vector corresponding to the enumeration field; when the field type is a text field, the text field is segmented, and the text field after the segmentation is mapped to a dense vector space based on the pre-trained embedding layer to obtain a third vector corresponding to the text field; the symbol mark in the continuous sequence of documents is determined, and the fourth vector is determined based on the unique code corresponding to the symbol mark; based on the preset order of the continuous sequence of documents, the first vector, the second vector, the third vector and the fourth vector are sorted to obtain the document vector.

[0011] In one implementation of the present application, based on the approval conclusion corresponding to the acquired document, the required field information is configured to determine the field feature combination vector, specifically including: based on the acquired document, determining the approved document and the rejected document; in the approved document, determining the compliant field combination, and performing a first mark on the compliant field combination as a compliant field feature combination; in the rejected document, determining the non-compliant field combination, and performing a second mark on the non-compliant field combination as a non-compliant field feature combination; determining the field feature combination vector based on the marked field feature combination, and storing the table name, field name and field value corresponding to the marked field feature combination in the database.

[0012] In one implementation of the present application, a preset model is trained based on a document vector and a field feature combination vector to obtain a document review model, specifically including: encoding the document vector and the field feature combination vector through an encoder to obtain a reference hidden state corresponding to each time step; using an attention mechanism to interact the current hidden state with all reference hidden states to determine the attention weight distribution; based on the attention weight distribution, weighted summing the reference hidden states to obtain a semantic vector; taking the current hidden state and the semantic vector as input, predicting the probability distribution of the output approval conclusion through a softmax layer, until an output sequence that meets the training conditions is generated, and determining the generation of the document review model.

[0013] In one implementation of the present application, after training a preset model based on a document vector and a field feature combination vector, the method further includes: comparing historical verification results with approval results to determine data with inconsistent approval probability values; re-inputting the data with inconsistent approval probability values ​​into the model for fine-tuning through rank-preserving regression technology until the output approval probability value meets the preset conditions; re-testing the output probability value of the fine-tuned model after a preset time period, and when the output probability value does not meet the preset condition of the probability value, determining first extended data based on documents obtained within the preset time period; performing language type conversion on the first extended data to obtain second extended data, and determining reference similar data having a similarity with the second extended data greater than a preset similarity threshold based on a preset database; converting the reference similar data into a language type corresponding to the first extended data, and expanding the sample set through the converted data; adjusting the fine-tuned model based on the expanded sample set to obtain a document review model.

[0014] In one implementation of the present application, the acquired document to be reviewed is input into a document review model, so as to output the review conclusion corresponding to the document to be reviewed through the document review model, specifically including: inputting the acquired document to be reviewed into the document review model, and outputting the review probability score through the document review model; and, when the probability review score belongs to the review rejection score range, displaying the error fields reviewed.

[0015] An embodiment of the present application provides a document review device, comprising: 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: determine document main table information and multiple sub-table information in the acquired document, and generate a document continuous sequence with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information; determine the field type in the document continuous sequence, and encode the document continuous sequence based on different field types to obtain a document vector; configure the required field information based on the approval conclusion corresponding to the acquired document, and determine a field feature combination vector; train a preset model based on the document vector and the field feature combination vector to obtain a document review model; input the acquired document to be reviewed into the document review model, so as to output the review conclusion corresponding to the document to be reviewed through the document review model.

[0016] A non-volatile computer storage medium provided by an embodiment of the present application stores computer executable instructions, and the computer executable instructions are configured to: determine document main table information and multiple sub-table information in an acquired document, and generate a document continuous sequence with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information; determine the field type in the document continuous sequence, and encode the document continuous sequence based on different field types to obtain a document vector; configure the required field information based on the approval conclusion corresponding to the acquired document, and determine a field feature combination vector; train a preset model based on the document vector and the field feature combination vector to obtain a document review model; input the acquired document to be reviewed into the document review model, so as to output the review conclusion corresponding to the document to be reviewed through the document review model.

[0017] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects: The embodiments of the present application apply the improved model to document review, realizing the purpose of automatic approval of the entire approval process from collecting document data, preprocessing, using supervised learning encoding to decoding, and outputting approval conclusions and error fields when approval is passed or not, thereby improving the efficiency and accuracy of the review and liberating human resources. The automated review model is based on preset rules and algorithms, and can more accurately identify and process document information, reduce errors caused by human factors, and improve the accuracy and reliability of the review. The embodiments of the present application can handle complex document structures containing a main table and multiple sub-tables. By constructing a continuous sequence of documents with a hierarchical relationship, it can more comprehensively capture the information in the document, thereby supporting more complex business review requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:

[0019] Figure 1 A flow chart of a document review method provided in an embodiment of the present application;

[0020] Figure 2 A schematic diagram of the structure of a document review device provided in an embodiment of the present application.

[0021] Reference numerals:

[0022] 200: document review device, 201: processor, 202: memory. DETAILED DESCRIPTION

[0023] The embodiments of the present application provide a document review method, device and medium.

[0024] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0025] The technical solution proposed in the embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0026] Figure 1 A flow chart of a document review method provided in an embodiment of the present application, such as Figure 1 As shown, the audit method includes the following steps:

[0027] Step 101: in the acquired document, determine the document main table information and multiple sub-table information, and generate a continuous sequence of documents with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information.

[0028] In one embodiment of the present application, in the acquired document, a first tag set and a second tag set in the document are determined; wherein the first tag set is related to the document user information, and the second tag set is related to the document event. A first document set corresponding to the same user is determined in the first tag set. The second tags corresponding to the documents in the first document set are determined, and the second tags are compared with a preset tag mapping table to determine the inclusion relationship between the second tags; wherein the preset tag mapping table includes multiple tags, and also includes the inclusion relationship between the multiple tags. Based on the inclusion relationship, the association relationship between the main table information and the multiple sub-table information is determined.

[0029] Specifically, in the acquired document, a first tag set in the document is determined, and the first tag set includes tags related to the document user information. These tags are usually used to identify user-dimensional information such as the submitter, approver, and owner of the document. For example, tags may include "user ID", "user name", "user department", etc. In the acquired document, a second tag set in the document is determined, and the second tag set includes tags related to document events. These tags describe the event-dimensional information such as the specific content, type, and status of the document. For example, tags may include "purchase order number", "payment amount", "approval status", etc.

[0030] Furthermore, after determining the first tag set, all documents belonging to the same user are grouped into one set, namely the first document set, according to tags such as "user ID" or "user name". For each document in the first document set, extract its corresponding second tag, i.e., the tag related to the document event. Then, compare these tags with the preset tag mapping table. The preset tag mapping table in the embodiment of the present application is a predefined table, which contains multiple tags and the inclusion relationship between these tags. For example, if there is a "payment" event, it may include multiple specific types of payments such as "advance payment", "delivery payment", and "final payment". These specific types of payment tags are defined as sub-tags of the "payment" tag in the preset tag mapping table and have an inclusion relationship.

[0031] Furthermore, by comparing the second tag with the preset tag mapping table, tags with inclusion relationships are identified. These inclusion relationships often reflect the master-slave relationship in the document structure, that is, some documents (or some parts of the documents) are the main table information, and other documents (or other parts of the documents) are the sub-table information of these main table information.

[0032] In one embodiment of the present application, multiple sub-table sequences are constructed based on multiple sub-table information; wherein the multiple sub-table sequences are respectively provided with unique identifiers. The main table information is combined with each sub-table sequence, and the unique identifiers of each sub-table sequence are respectively combined to obtain multiple independent document sequences. Based on the timestamps of the multiple sub-table information, the order between the multiple sub-table sequences is determined, and the multiple independent document sequences are sorted to obtain a continuous sequence of documents with a hierarchical relationship.

[0033] Specifically, in actual business, documents often contain information from a main table and multiple related sub-tables. First, this structured data is "flattened". This means integrating the key information of the main table and sub-tables into a continuous sequence, and distinguishing each sub-table record from other records through specific tags (such as [start_subtable] and [end_subtable]) to ensure that the model can understand the hierarchical relationship between the data.

[0034] Specifically, for example, a unique identifier can be assigned to each sub-table group, and each sub-table information set forms a sub-table sequence. The main table information is combined with each sub-table sequence and its unique identifier to generate one or more independent document sequences for each main table information. Each document sequence contains the main table information and all its associated sub-table sequences and their unique identifiers. In most cases, the records in the sub-table sequence are sorted according to the timestamp to reflect their actual occurrence order in the business process. Based on the timestamp, the generated multiple independent document sequences are sorted to obtain a continuous sequence of documents with a hierarchical relationship.

[0035] For example, if the document is a difference rate reimbursement form, the main table includes the name of the person making the reimbursement, the department to which the person belongs, and the reason for the reimbursement. The corresponding first sub-table includes the starting and ending places corresponding to the destination, and the corresponding second sub-table includes the starting and ending places corresponding to the return. The main table and the first sub-table form a first independent document sequence, and the main table and the second sub-table form a second independent document sequence. The first independent document sequence and the second independent document sequence are sorted to obtain the continuous sequence of the document.

[0036] Step 102: Determine the field type in the continuous sequence of documents, encode the continuous sequence of documents based on different field types, and obtain a document vector.

[0037] In one embodiment of the present application, when the field type is a numeric field, the numeric field is used as the first vector. When the field type is an enumeration field, the enumeration value is mapped to a dense vector space through a pre-trained embedding layer to obtain a second vector corresponding to the enumeration field. When the field type is a text field, the text field is segmented, and the text field after the segmentation is mapped to a dense vector space based on the pre-trained embedding layer to obtain a third vector corresponding to the text field. The symbol mark in the continuous sequence of the document is determined, and the fourth vector is determined based on the unique code corresponding to the symbol mark. Based on the preset order of the continuous sequence of the document, the first vector, the second vector, the third vector and the fourth vector are sorted to obtain a document vector.

[0038] Specifically, all fields, including numeric, enumerated, and textual types, need to be converted into a vector form that the model can understand. The embodiment of the present application is implemented through word embedding technology. For numeric fields, they can be used directly; enumerated and textual fields are converted into dense vectors through pre-trained embedding layers. In addition, special tags (such as headers and separators) should also be given unique codes to ensure that the model can recognize the data structure and can be correctly input into the model's entry.

[0039] Specifically, for numeric fields, their values ​​can be directly used as one or more elements of the vector. Enumeration fields usually contain a set of predefined options, such as the "status" field may contain options such as "pending", "processing", and "completed". Use a pre-trained embedding layer to convert these enumeration values ​​into vectors. For text fields, first perform word segmentation on them to decompose the text into a series of words or tags. Then, these word segments are mapped to a dense vector space through a pre-trained embedding layer similar to the enumeration field. Since the text may contain multiple words, the mapped vectors are summed or averaged to obtain a vector representing the entire text field. In the continuous sequence of documents, there are also some symbolic tags, which are assigned unique codes and mapped to the vector space through the embedding layer. Based on the preset order of the continuous sequence of documents, the four types of vectors obtained above are sorted and combined into a single vector, namely the document vector.

[0040] Step 103: Based on the approval conclusion corresponding to the acquired document, the required field information is configured to determine the field feature combination vector.

[0041] In one embodiment of the present application, based on the acquired documents, approved documents and rejected documents are determined. In the approved documents, compliant field combinations are determined, and the compliant field combinations are first marked as compliant field feature combinations. In the rejected documents, non-compliant field combinations are determined, and the non-compliant field combinations are second marked as non-compliant field feature combinations. Based on the marked field feature combinations, field feature combination vectors are determined, and the table names, field names, and field values ​​corresponding to the marked field feature combinations are stored in the database.

[0042] Specifically, each time a document is approved or rejected, the fields that need to be recorded in the document are configured and recorded, the table name, field name, and field value of the field are stored in the database, the audit conclusion (compliant / non-compliant) of each record is marked, and the audit opinion is recorded. These marks serve as the target variables for model learning, guiding the model to learn which feature combinations point to compliant behavior and which feature combinations may indicate potential problems.

[0043] Specifically, documents are divided into two categories according to their approval status: approved documents and rejected documents. In approved documents, it is necessary to determine which field combinations are compliant. For example, if a purchase document requires three fields: supplier name, purchase quantity, and purchase amount, and the values ​​of these fields need to meet specific conditions (such as the purchase amount is not 0, and the supplier name is in the predefined supplier list), then the combination of these fields is considered compliant. Further, the compliant field combinations are determined and first marked to distinguish these combinations as compliant.

[0044] Furthermore, in the rejected documents, the non-compliant field combinations that caused the documents to be rejected are determined. This can be done by viewing the rejection reason or approval note fields, which provide detailed information about the rejection of the document. For the determined non-compliant field combinations, a second mark is performed to distinguish these combinations as non-compliant. Based on the marked field combinations, the field feature combination vector is further determined. This usually involves encoding the field's tag, name, value and other information into a vector form to facilitate subsequent analysis or machine learning processing. Finally, the table name, field name, field value and tag information corresponding to the marked field feature combination are stored in the database.

[0045] Step 104: Based on the document vector and the field feature combination vector, the preset model is trained to obtain a document review model.

[0046] In one embodiment of the present application, the document vector and the field feature combination vector are encoded by an encoder to obtain a reference hidden state corresponding to each time step. Using the attention mechanism, the current hidden state interacts with all reference hidden states to determine the attention weight distribution. Based on the attention weight distribution, the reference hidden state is weighted and summed to obtain a semantic vector. The current hidden state and the semantic vector are used as inputs, and the probability distribution of the output approval conclusion is predicted through the softmax layer until an output sequence that meets the training conditions is generated, and the document review model generation is determined.

[0047] Specifically, the embodiment of the present application improves the encoding part in the encoding and decoding model, encodes the input into a sequence of vectors, and generates a bounded range, so that the information contained in the input vector sequence can be fully utilized when focusing on the key points. Repeating the above steps can input the codes corresponding to the fields into the model.

[0048] Specifically, the role of the encoder in the embodiment of the present application is to process these vector inputs and convert them into a series of hidden states. These hidden states capture the important features of the input data and are used for subsequent predictions. In a recurrent neural network, the encoder inputs the vector in time steps to generate a sequence of hidden states. The attention mechanism allows the model to focus on different parts of the input when generating outputs. The embodiment of the present application uses an attention mechanism to interact the current hidden state with all reference hidden states. The result of the interaction is an attention weight distribution, which represents the correlation between the current output and each element in the input sequence. According to the attention weight distribution, all reference hidden states are weighted and summed to obtain a semantic vector. The semantic vector in the embodiment of the present application is a weighted representation of the input sequence, reflecting the input part that the current output should focus on.

[0049] Furthermore, the current hidden state and semantic vector are used as inputs, and the probability distribution of the output approval conclusion is predicted through the softmax layer. The softmax layer in the embodiment of the present application converts the input into a probability distribution so that the sum of the probabilities of all possible outputs (such as "passed" and "rejected") is 1. For example, in the output of the softmax layer, the model may give an approval conclusion of "passed" with a probability of 0.8 and a "rejected" probability of 0.2. Based on this probability distribution, the final approval conclusion can be determined. The training data in the embodiment of the present application contains known document inputs and corresponding approval conclusions. The model optimizes its parameters by minimizing the difference between the predicted output and the actual output (such as using a cross entropy loss function).

[0050] In one embodiment of the present application, the historical verification results are compared with the approval results to determine the data with inconsistent approval probability values. Through the rank-preserving regression technology, the data with inconsistent approval probability values ​​are input into the model again for fine-tuning until the output approval probability value meets the preset conditions. After a preset time period, the output probability value of the fine-tuned model is re-detected, and when the output probability value does not meet the preset condition of the probability value, the first extended data is determined based on the documents obtained within the preset time period. The first extended data is converted to a language type to obtain the second extended data, and reference similar data whose similarity with the second extended data is greater than a preset similarity threshold is determined based on a preset database. The reference similar data is converted to the language type corresponding to the first extended data, and the sample set is expanded by the converted data. The fine-tuned model is adjusted based on the expanded sample set to obtain a document review model.

[0051] Specifically, in order to improve the reliability of the document review model prediction, especially the calibration of the probability output of approval and failure, the historical verification results are compared with the actual approval results. The inconsistent data after the actual comparison is re-entered into the model for fine-tuning, and the rank-preserving regression technology is used to ensure that the probability value output by the model better reflects the accuracy during the review. In addition, the fine-tuned model is re-evaluated on a monthly basis to address the overfitting problem of this model.

[0052] Furthermore, the model's prediction results (i.e., approval probability values) on historical data are compared with the actual approval results to identify samples where the model predictions do not match the actual situation, i.e., data with inconsistent approval probability values. These inconsistent data reflect the deficiencies or deviations of the model in certain situations. For data with inconsistent approval probability values, the model is adjusted using rank-preserving regression techniques to reduce the inconsistency between the model predictions and the actual approval results. During the fine-tuning process, the model relearns the features in these inconsistent data and adjusts its parameters so that the actual approval situation is more accurately reflected in future predictions.

[0053] Furthermore, after a period of time, the output probability value of the fine-tuned model is retested to evaluate the stability and accuracy of the model after long-term operation. If the output probability value of the model still does not meet the preset conditions (such as the prediction accuracy is lower than the threshold), further measures need to be taken.

[0054] Further, when the output probability value does not meet the preset conditions, the first extended data is determined based on the new documents obtained within the preset time period. The first extended data is converted into a language type, and then the reference similar data whose similarity with the converted second extended data is greater than the preset similarity threshold is searched in the preset database. For example, if the first extended data is a Chinese document, it is first converted into a unified English format. Then, the reference data whose similarity with these converted documents is greater than 80% is searched in the preset English document database, and these reference data will be used as additional training samples. The reference similar data found is converted back to the language type corresponding to the first extended data, and the original sample set is expanded by these converted data. Doing so can increase the amount of training data of the model and improve its adaptability to different languages ​​and scenarios. Finally, the fine-tuned model is further trained and adjusted using the expanded sample set. Through this process, the model can learn more features and patterns of samples, thereby improving the accuracy and stability of its approval probability value.

[0055] Step 105: input the acquired documents to be reviewed into the document review model, so as to output the review conclusion corresponding to the documents to be reviewed through the document review model.

[0056] In one embodiment of the present application, the acquired documents to be reviewed are input into the document review model, and the review probability score is output through the document review model. And, when the probability review score belongs to the review rejection score range, the error fields found out during the review are displayed.

[0057] Specifically, the document to be reviewed is input into the document review model. The model encodes each field of the document and then generates a hidden state or representation vector through the encoder. Next, the model uses the attention mechanism to focus on the fields or features that have a greater impact on the approval result. Finally, an audit probability score is output through a softmax layer or a similar mechanism. After obtaining the audit probability score, it is compared with a preset audit rejection score interval. If the audit probability score falls within this interval, it is considered that the document needs to be rejected. At this point, the model may need to further analyze the fields in the document and identify the erroneous fields that caused the rejection. These erroneous fields may be due to incomplete data, incorrect format, out-of-range values, or other violations of audit rules.

[0058] Figure 2This is a schematic diagram of the structure of a document review device provided in an embodiment of the present application. Figure 2 As shown, the document review device 200 includes: at least one processor 201; and a memory 202 that is communicatively connected to the at least one processor 201; wherein the memory 202 stores instructions that can be executed by the at least one processor 201, and the instructions are executed by the at least one processor 201 so that the at least one processor 201 can: determine the document main table information and multiple sub-table information in the acquired document, and generate a document continuous sequence with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information; determine the field type in the document continuous sequence, and encode the document continuous sequence based on different field types to obtain a document vector; based on the approval conclusion corresponding to the acquired document, configure the required field information and determine the field feature combination vector; based on the document vector and the field feature combination vector, train the preset model to obtain a document review model; input the acquired document to be reviewed into the document review model to output the review conclusion corresponding to the document to be reviewed through the document review model.

[0059] A non-volatile computer storage medium provided by an embodiment of the present application stores computer executable instructions, and the computer executable instructions are configured to: determine document main table information and multiple sub-table information in an acquired document, and generate a document continuous sequence with a hierarchical relationship based on the association relationship between the main table information and the multiple sub-table information; determine the field type in the document continuous sequence, and encode the document continuous sequence based on different field types to obtain a document vector; configure the required field information based on the approval conclusion corresponding to the acquired document, and determine a field feature combination vector; train a preset model based on the document vector and the field feature combination vector to obtain a document review model; input the acquired document to be reviewed into the document review model, so as to output the review conclusion corresponding to the document to be reviewed through the document review model.

[0060] 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, equipment, and non-volatile computer storage 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 embodiment.

[0061] 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 embodiments of the present application may have various changes and variations. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A document review method, characterized in that: The method comprises: In the acquired documents, the document main table information and a plurality of sub-table information are determined, and based on the association relationship between the main table information and the plurality of sub-table information, a document continuous sequence with a hierarchical relationship is generated; Determine the field types in the continuous sequence of documents, and encode the continuous sequence of documents based on different field types to obtain a document vector; Based on the approval conclusion corresponding to the acquired document, the required field information is configured to determine a field feature combination vector; Based on the document vector and the field feature combination vector, a preset model is trained to obtain a document review model; Input the acquired document to be reviewed into the document review model, so as to output the review conclusion corresponding to the document to be reviewed through the document review model; The step of generating a continuous sequence of documents having a hierarchical relationship based on the association relationship between the main table information and the plurality of sub-table information specifically includes: Based on the plurality of sub-table information, a plurality of sub-table sequences are constructed; wherein the plurality of sub-table sequences are respectively provided with unique identifiers; Combining the main table information with each of the sub-table sequences and the unique identifiers of each of the sub-table sequences to obtain multiple independent document sequences; Based on the timestamps of the multiple sub-table information, the order between the multiple sub-table sequences is determined, and the multiple independent document sequences are sorted to obtain the continuous sequence of documents with a hierarchical relationship.

2. A document review method according to claim 1, characterized in that: Before generating a continuous sequence of documents with a hierarchical relationship based on the association relationship between the main table information and the plurality of sub-table information, the method further includes: In the acquired document, a first tag set and a second tag set in the document are determined; wherein the first tag set is related to the document user information, and the second tag set is related to the document event; Determining a first document set corresponding to the same user in the first tag set; Determine the second tags corresponding to the documents in the first document set respectively, compare the second tags with the preset tag mapping table, and determine the inclusion relationship between the second tags; wherein the preset tag mapping table includes multiple tags and also includes the inclusion relationship between the multiple tags; Based on the inclusion relationship, the association relationship between the main table information and the plurality of sub-table information is determined.

3. A document review method according to claim 1, characterized in that: The determining of the field types in the continuous sequence of documents, encoding the continuous sequence of documents based on different field types, and obtaining a document vector specifically includes: In the case where the field type is a numeric field, taking the numeric field as the first vector; In the case where the field type is an enumeration field, mapping the enumeration value into a dense vector space through a pre-trained embedding layer to obtain a second vector corresponding to the enumeration field; In the case where the field type is a text field, performing word segmentation processing on the text field, and mapping the text field after the word segmentation processing to a dense vector space based on the pre-trained embedding layer to obtain a third vector corresponding to the text field; Determine the symbol marks in the continuous sequence of the document, and determine a fourth vector based on the unique code corresponding to the symbol marks; Based on the preset order of the continuous sequence of the documents, the first vector, the second vector, the third vector and the fourth vector are sorted to obtain the document vector.

4. A document review method according to claim 1, characterized in that: The step of configuring the required field information based on the approval conclusion corresponding to the acquired document and determining the field feature combination vector specifically includes: Based on the obtained documents, determine the approved documents and the rejected documents; In the approved document, a compliance field combination is determined, and the compliance field combination is first marked as a compliance field feature combination; Determining a non-compliant field combination in the rejection document, and marking the non-compliant field combination as a non-compliant field feature combination; A field feature combination vector is determined based on the marked field feature combination, and the table name, field name and field value corresponding to the marked field feature combination are stored in a database.

5. A document review method according to claim 1, characterized in that: The step of training a preset model based on the document vector and the field feature combination vector to obtain a document review model specifically includes: Encoding the document vector and the field feature combination vector through an encoder to obtain a reference hidden state corresponding to each time step; Use the attention mechanism to interact the current hidden state with all reference hidden states to determine the attention weight distribution; Based on the attention weight distribution, performing weighted summation on the reference hidden state to obtain a semantic vector; The current hidden state and the semantic vector are used as inputs, and the probability distribution of the output approval conclusion is predicted through the softmax layer until an output sequence that meets the training conditions is generated, thereby determining the generation of the document review model.

6. A document review method according to claim 1, characterized in that: After training the preset model based on the document vector and the field feature combination vector, the method further includes: Compare historical verification results with approval results to identify data with inconsistent approval probability values; By using the rank-preserving regression technology, the data with inconsistent approval probability values ​​are input into the model again for fine-tuning until the output approval probability value meets the preset conditions; After a preset time period, the output probability value of the fine-tuned model is re-tested, and when the output probability value does not meet the preset probability value condition, the first expansion data is determined based on the documents obtained within the preset time period; Performing language type conversion on the first extended data to obtain second extended data, and determining reference similar data having a similarity with the second extended data greater than a preset similarity threshold based on a preset database; Converting the reference similar data into a language type corresponding to the first extended data, and expanding the sample set by using the converted data; The fine-tuned model is adjusted based on the expanded sample set to obtain the document review model.

7. A document review method according to claim 1, characterized in that: The step of inputting the acquired document to be reviewed into the document review model so as to output the review conclusion corresponding to the document to be reviewed through the document review model specifically includes: Inputting the acquired document to be reviewed into the document review model, and outputting the review probability score through the document review model; Also, when the probability review score falls within the review rejection score range, the error fields found during the review will be displayed.

8. A document review device, characterized in that: The device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute the method according to any one of claims 1 to 7.

9. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 7.

Images