Intelligent document auditing method and device, computer equipment and storage medium
By using a trained field extraction model and preset rules to perform semantic recognition and matching verification on financial and insurance documents, the problems of low efficiency and low accuracy in existing technologies are solved, and efficient and accurate document review is achieved.
Patent Information
- Application Number
- CN202411183249.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-11-29
AI Technical Summary
Existing technologies for document review in the financial and insurance industries suffer from low efficiency, incomplete coverage, and low accuracy. Manual review is time-consuming, engineering matching is limited, and large language models are unstable in recognizing specific clauses, failing to meet real-time requirements.
The trained field extraction model is used for semantic recognition. The content to be reviewed is extracted through text segmentation, detection, recognition, encoding and entity recognition units. The review is carried out in combination with preset verification rules and matching rules. Convolutional neural networks and attention mechanisms are used to improve text understanding and semantic matching.
It improved the efficiency and accuracy of document review, reduced the possibility of risk leakage, optimized the review process, saved manpower, and achieved more efficient compliance checks.
Smart Images

Figure CN119047484A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the fields of artificial intelligence and financial technology, and in particular to an intelligent document review method, apparatus, computer equipment and storage medium. Background Art
[0002] With the rapid development of the financial market, regulatory filing is a key link in ensuring compliance in the financial, insurance and other industries. However, there are a series of problems in the current filing process: the terms do not meet regulatory requirements; the format of the filing materials is not standardized; lax review or system defects lead to high regulatory return rates and serious risk leakage.
[0003] The current solutions mainly include manual review, engineering matching, and the use of large language models. However, these solutions have the following problems:
[0004] 1) Manual review has the disadvantages of being time-consuming, inefficient, and labor-intensive;
[0005] 2) Engineering matching is limited by the limitations of regular expressions and cannot fully cover all complex clause formats and semantic verification requirements;
[0006] 3) Large language models are unstable in identifying specific financial or insurance terms, have slow processing speeds, and cannot meet the needs of real-time review. In addition, secondary training is required to improve the accuracy and adaptability of the model, which is costly and complicated.
[0007] In summary, current document review has problems such as low efficiency, lack of comprehensive coverage, and low accuracy. Summary of the invention
[0008] The purpose of the embodiments of the present application is to propose an intelligent document review method, device, computer equipment and storage medium to solve the technical problems in related technologies such as low document review efficiency and lack of comprehensive coverage, which in turn leads to low review accuracy.
[0009] In order to solve the above technical problems, the present application embodiment provides an intelligent document review method, which adopts the following technical solutions:
[0010] Receiving a document to be reviewed, and inputting the document to be reviewed into a trained field extraction model for semantic recognition and extraction to obtain content to be reviewed;
[0011] Verify the validity of the fields of the content to be reviewed according to the preset verification rules to obtain the valid review content;
[0012] The audit reference information is called, and the valid audit content is matched and verified with the audit reference information based on a preset matching rule to obtain an audit result.
[0013] Furthermore, the field extraction model includes a text segmentation unit, a text detection unit, a text recognition unit, an encoding unit, an entity recognition unit and an output unit; the step of inputting the document to be reviewed into the trained field extraction model for semantic recognition and extraction to obtain the content to be reviewed includes:
[0014] Inputting the document to be reviewed into the text segmentation unit for element recognition to obtain text areas and non-text areas;
[0015] Performing text detection on the text area and the non-text area by the text detection unit to obtain a global text feature containing text position information;
[0016] Inputting the global features of the text into the text recognition unit for recognition to obtain a text recognition result;
[0017] Performing syntactic analysis on the text recognition result to obtain semantic phrases, and inputting the semantic phrases into the encoding unit to perform attention calculation to obtain semantic relationship vectors;
[0018] Extracting audit information from the semantic relationship vector by the entity recognition unit to obtain an audit entity;
[0019] The audit entity is outputted according to a preset template through the output unit to obtain the content to be audited.
[0020] Furthermore, the text segmentation unit includes a convolutional neural network layer, a region candidate network layer, an interest region pooling layer, an attention layer, and a classification layer; the step of inputting the document to be reviewed into the text segmentation unit for element recognition to obtain the text region and the non-text region includes:
[0021] Extracting features of the document to be reviewed by using the convolutional neural network to obtain a document feature representation;
[0022] Inputting the document feature representation into the region candidate network layer, and obtaining candidate region features through the region candidate network layer;
[0023] Inputting the document feature representation and the candidate region feature into the region of interest pooling layer for pooling operation to obtain document pooling features;
[0024] Performing attention calculation on the document pooling feature through the attention layer to obtain an attention enhancement feature, and fusing the attention enhancement feature with the document pooling feature to obtain an enhanced pooling fusion feature;
[0025] The enhanced pooling fusion features are classified by the classification layer to obtain the text area and non-text area corresponding to the document to be reviewed.
[0026] Furthermore, the encoding unit includes a vector embedding layer, an average pooling layer, a multi-layer perceptron layer and an encoder layer; the step of inputting the semantic phrase into the encoding unit for attention calculation to obtain the semantic relationship vector includes:
[0027] Inputting the semantic phrase into the vector embedding layer for embedding to obtain a coding sequence vector;
[0028] Performing a pooling operation on the encoding sequence vector through the average pooling layer to obtain a pooled semantic code;
[0029] Inputting the pooled semantic code and the code sequence vector into the multi-layer perceptron layer for weighted calculation to obtain a phrase sequence vector;
[0030] The encoder layer performs attention calculation on the phrase sequence vector to obtain a semantic relationship vector.
[0031] Furthermore, the entity recognition unit includes a fully connected layer and a softmax layer; the step of extracting audit information from the semantic relationship vector through the entity recognition unit to obtain the audit entity includes:
[0032] The semantic relationship vectors are integrated and connected through the fully connected layer to obtain a fully connected semantic vector;
[0033] The fully connected semantic vector is input into the softmax layer for classification, and the audit entity is output.
[0034] Furthermore, the step of verifying the validity of the fields of the content to be reviewed according to the preset verification rules to obtain the valid review content includes:
[0035] Extracting the text field of the content to be reviewed according to a preset regular expression to obtain the review field and the corresponding field content;
[0036] Verify the audit field and the field content based on preset verification rules to obtain a verification result;
[0037] The content to be reviewed is modified based on the verification result to obtain valid review content.
[0038] Furthermore, the step of matching and verifying the valid audit content with the audit reference information based on a preset matching rule to obtain the audit result includes:
[0039] Extract all valid audit fields and corresponding audit field values in the valid audit content;
[0040] Classify the valid audit fields to obtain field classification results;
[0041] Acquire a corresponding field reference value from the audit reference information according to the valid audit field;
[0042] According to the field classification result, the audit field value is matched with the field reference value according to the corresponding preset matching rule to obtain a matching degree;
[0043] An audit result is obtained based on the matching degree.
[0044] In order to solve the above technical problems, the embodiment of the present application also provides an intelligent document review device, which adopts the following technical solution:
[0045] The recognition and extraction module is used to receive the document to be reviewed, input the document to be reviewed into the trained field extraction model for semantic recognition and extraction, and obtain the content to be reviewed;
[0046] The validity verification module is used to verify the field validity of the content to be reviewed according to the preset verification rules to obtain the valid review content;
[0047] The audit module is used to call the audit reference information, match and verify the effective audit content with the audit reference information based on a preset matching rule, and obtain the audit result.
[0048] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:
[0049] The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the intelligent document review method as described above when executing the computer-readable instructions.
[0050] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:
[0051] The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps of the intelligent document review method described above are implemented.
[0052] Compared with the prior art, this application has the following beneficial effects:
[0053] The present application provides an intelligent document review method, which inputs a received document to be reviewed into a trained field extraction model for semantic recognition and extraction to obtain the content to be reviewed, verifies the field validity of the content to be reviewed to obtain the valid review content, and finally matches and verifies the valid review content with the review reference information based on preset matching rules; the present application performs semantic recognition on the document to be reviewed through the trained field extraction model, so that the field extraction model can better understand the context and semantics of the text, comprehensively cover the review content, reduce the possibility of risk leakage, and improve the efficiency and accuracy of recognition. At the same time, by verifying the validity of the field, it helps to optimize the review process and further improve the review efficiency. Finally, matching and verification are performed through preset matching rules, which can save manpower investment and improve the overall review efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0055] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;
[0056] Figure 2 is a flow chart of an embodiment of the intelligent document review method according to the present application;
[0057] Figure 3 yes Figure 2 A flowchart of a specific implementation of step S201;
[0058] Figure 4 is a schematic diagram of the structure of an embodiment of an intelligent document review device according to the present application;
[0059] Figure 5 It is a structural diagram of an embodiment of a computer device according to the present application. DETAILED DESCRIPTION
[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.
[0061] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0062] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.
[0063] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
[0064] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.
[0065] This application provides an intelligent document review method that can be applied to Figure 1 In the system architecture 100 shown, the system architecture 100 may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.
[0066] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.
[0067] Terminal devices 101, 102, 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 (Moving Picture Experts Group Audio Layer IV) players, laptop computers, desktop computers, etc.
[0068] The server 105 may be a server that provides various services, such as a background server that provides support for web pages displayed on the terminal devices 101 , 102 , and 103 .
[0069] It should be noted that the intelligent document review method provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the intelligent document review device is generally set in the server / terminal device.
[0070] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.
[0071] Continue to refer Figure 2 , shows a flow chart of an embodiment of the intelligent document review method according to the present application, comprising the following steps:
[0072] Step S201, receiving a document to be reviewed, inputting the document to be reviewed into a trained field extraction model for semantic recognition and extraction, and obtaining content to be reviewed.
[0073] In the financial field, the documents to be reviewed include but are not limited to contract terms, insurance terms, financial product instructions, compliance statements, risk assessment reports, regulatory filings, legal opinions, etc. Among them, financial product instructions are financial product instructions issued by financial institutions such as banks and securities companies, including product characteristics, risk warnings, etc.; compliance statements are statements submitted by enterprises or individuals to prove that their business activities comply with specific laws and regulations; risk assessment reports are reports that assess the risks of specific businesses or projects, including risk identification, assessment and control measures; regulatory filings are various filings submitted by enterprises or individuals in accordance with regulatory requirements, such as financial statements, business reports, etc.; legal opinions are opinions on specific legal issues, which may involve contract review, compliance assessment, etc.
[0074] In this embodiment, the document to be reviewed can be received via a wired connection or a wireless connection. It should be noted that the wireless connection may include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.
[0075] The formats of the documents to be reviewed include but are not limited to text, PDF, tables, and images.
[0076] It should be emphasized that in order to further ensure the privacy and security of the documents to be reviewed, the above documents to be reviewed can also be stored in a node of a blockchain.
[0077] The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.
[0078] In this embodiment, the field extraction model is trained, and the training process of the field extraction model includes the following steps:
[0079] A large number of historical audit documents are obtained as sample data sets. The sample data sets also include label data corresponding to each historical audit document. The label data is the audit entity category, which is used to classify the extracted audit content. Among them, the label data includes insurance name, subject matter, rate, rate adjustment coefficient, etc. The label data can be annotated using methods such as BIO, BIOE, BIOES and BMES. Taking BIOE as an example, BIOE tags include five types of labels: B, I, E, S, and O. Among them, B represents the beginning of an entity and is used for the first word of the entity; I represents the middle part of an entity and is used for the words in the middle of the entity; E represents the end of an entity and is used for the last word of the entity; S indicates that an entity consists of only a single word; O indicates that a word does not belong to any entity. Taking the title "Property Insurance Rice Seedling Planting Insurance Clauses" as an example, the named entities are property insurance, rice seedling planting, and insurance clauses, and the labels are as follows: property insurance is B-INSURANCE-NAME, rice is B-subject, seedling planting is E-subject, and insurance clauses are B-INSURANCE_POLICY.
[0080] The sample data set is randomly divided into a training set and a test set according to a preset ratio, and the pre-built network model is trained using the training set and the test set to finally obtain a trained field extraction model.
[0081] In some optional implementations, the field extraction model includes a text segmentation unit, a text detection unit, a text recognition unit, an encoding unit, an entity recognition unit and an output unit. The document to be reviewed is input into the field extraction model, and the document to be reviewed is processed in sequence by the text segmentation unit, the text detection unit, the text recognition unit, the encoding unit, the entity recognition unit and the output unit to obtain a semantic recognition result, i.e., the content to be reviewed.
[0082] In some optional implementations of this embodiment, see Figure 3 As shown, the above step of inputting the document to be reviewed into the trained field extraction model for semantic recognition and extraction to obtain the content to be reviewed includes:
[0083] Step S301: Input the document to be reviewed into a text segmentation unit for element recognition to obtain text areas and non-text areas.
[0084] In this embodiment, the elements of the document to be reviewed include but are not limited to titles, subtitles, text content, paragraphs, charts and graphs, tables, etc.; wherein titles, subtitles, text content, paragraphs, etc. belong to the text area, and charts and graphs, tables, etc. belong to the non-text area. The text segmentation unit recognizes the elements of the document to be reviewed and outputs the recognition result, i.e., marks the text area and non-text area of the document to be reviewed.
[0085] Step S302: Perform text detection on the text area and the non-text area by a text detection unit to obtain a global text feature including text position information.
[0086] The text detection unit mainly uses text detection algorithms to locate text areas. Text detection algorithms include PSENet algorithm, PANNet algorithm, MSR (Multi-Scale Shape Regression for Scene Text Detection) algorithm and DB (Differentiable Binarization) algorithm, etc. Taking the PSENet algorithm as an example, the PSEnet algorithm is implemented using the PSEnet network. The architecture of the PSEnet network is resnet+FPN (Feature Pyramid Networks). Features are extracted through resnet, and features of different layers are sent to FPN for feature fusion to achieve a progressive expansion algorithm of the segmentation results.
[0087] Among them, the resnet in the psenet network is a residual network based on the bottleneck layer. Specifically, the resnet50 network is used. Resnet50 has 50 conv2d layers. The features of the output of Conv2, Conv3, Conv4 and Conv5 layers are extracted to construct a feature pyramid. The top-down and lateral connections are used, that is, the 4-layer feature P2, P3, P4, and P5 are extracted to obtain 4 256-channel features. The features of the 4 channels are fused, and P3, P4, and P5 are upsampled by 2 times, 4 times, and 8 times respectively. Feature cascade is performed on the feature layer P2, and finally a 1024-dimensional fused feature vector is obtained. The fused feature vector is passed through the convolution layer and the BN layer (Batch Layer). Normalization, batch normalization) and ReLU layers (activation function layer) are used to reduce the dimension of the fused feature vector to a feature of 256, and then this feature is input into a 1×1 convolutional layer to obtain n segmentation results. The segmentation results are arranged in ascending order according to the kernel size. The BFE-based expansion algorithm is used to start from the smallest kernel size and gradually enlarge it until it is finally enlarged to the size of the original text to locate the text and obtain the text line positioning box. The text line positioning box is marked in the document waiting for review, and the text global feature containing the text position information is output.
[0088] Step S303: input the global features of the text into a text recognition unit for recognition to obtain a text recognition result.
[0089] In this embodiment, the text recognition unit can perform text recognition on tables, paragraphs, titles, etc. of the document to be reviewed.
[0090] In some optional implementations, the text recognition unit adopts a trained MASTER model, the MASTER model includes an encoder and a decoder, the encoder includes a feature extraction layer and an encoding layer, the encoding layer includes a residual sublayer, a multi-head global attention sublayer and a feature fusion sublayer; the decoder includes an embedding layer, a decoding layer and a linear & softmax layer, the decoding layer includes a masked multi-head attention sublayer, a residual connection sublayer, a multi-head attention sublayer, a residual connection sublayer, a feedforward network sublayer and a residual connection sublayer connected in sequence.
[0091] The text recognition unit performs text recognition on the global features of the text to obtain the corresponding text recognition results.
[0092] In some optional implementations, the text area may also directly use an OCR (Optical Character Recognition) model for text recognition.
[0093] Step S304, performing syntactic analysis on the text recognition result to obtain semantic phrases, and inputting the semantic phrases into the encoding unit to perform attention calculation to obtain semantic relationship vectors.
[0094] Among them, syntactic analysis includes component syntactic analysis and dependency syntactic analysis. In this embodiment, component syntactic analysis is used to traverse each sentence in the text recognition result to obtain a syntactic tree. The labels marked on the nodes in the generated syntactic tree correspond to the components contained in each sentence, among which ADVP is an adverbial phrase, NP is a noun phrase, VP is a verb phrase, VBD is the past tense of the verb, PP is a prepositional phrase, S is a simple declarative sentence, SBAR is a clause, WHNP is a WH noun phrase, and CC is a conjunction.
[0095] There are two main types of component syntactic analysis tag sets, phrase tags and clause tags. Phrase tags include ADVP, NP, VP, PP, etc., and clause tags include S, SBAR, etc. This embodiment uses phrase tags to determine semantic phrases, thereby achieving the purpose of segmenting sentences and ensuring that the audit entity is identified.
[0096] In some optional implementations, the encoding unit includes a connected vector embedding layer, an average pooling layer, a multi-layer perceptron layer, and an encoder layer. The above step of inputting the semantic phrase into the encoding unit for attention calculation to obtain the semantic relationship vector includes:
[0097] The semantic phrase is input into the vector embedding layer for embedding to obtain the encoding sequence vector;
[0098] The coding sequence vector is pooled through the average pooling layer to obtain the pooled semantic coding;
[0099] The pooled semantic code and the code sequence vector are input into the multi-layer perceptron layer for weighted calculation to obtain the phrase sequence vector;
[0100] The attention calculation is performed on the phrase sequence vector through the encoder layer to obtain the semantic relationship vector.
[0101] In this embodiment, it is assumed that the input document to be reviewed is d, and the document d contains n sentences (s1, s2, ..., s n ), m words (w1,w2,...,w m ), where the i-th sentence contains l semantic phrases, represented by s i =(c i,1 ,c i,2 ,...,c i,l ), sentence s i In the example, the jth semantic phrase consisting of k words is denoted as c i,j =(w i,j,1 ,...w i,j,k ).
[0102] Before inputting the semantic phrase sequence into the vector embedding layer, the text recognition results need to be cleaned to remove irrelevant characters, such as special punctuation marks, blank characters, line breaks, spaces, etc.
[0103] After cleaning the data, the text recognition results are divided into phrases, with a [cls] symbol added before each phrase and a [sep] symbol added at the end of the sentence. The [cls] symbol is used to output a vector to represent the semantics of the phrase, and the [sep] symbol is used to indicate the end of the phrase.
[0104] Before inputting the semantic phrase into the vector embedding layer, the semantic phrase needs to be vectorized to obtain a text sequence. The text sequence of document d is ([cls], w1, w2, ..., w4, [sep], [cls], w5, ..., w7, [sep], [cls], w8, ..., w m ), the text sequence contains the [cls] symbol and the [sep] symbol, and the text sequence is input into the vector embedding layer to obtain the encoding sequence vector, which is expressed as:
[0105]
[0106] Get the encoding sequence vector corresponding to the semantic phrase, output the encoding sequence vector through the encoding sequence vector, and use the semantic phrase c i,j For example, the corresponding semantic phrase encoding is:
[0107] Among them, the coding sequence vector is pooled through the average pooling layer to obtain the pooled semantic code; the pooled semantic code is pooled through the average pooling layer to generate the document code of the document to be reviewed; the pooled semantic code and the coding sequence vector are input into the multi-layer perceptron layer to construct a weighted representation matrix; based on the pooled semantic code and the weighted representation matrix, the semantic relationship vector is calculated.
[0108] In some optional implementations, the encoder layer is used to capture the relationship between semantic phrases, including a multi-head attention mechanism layer and a feed-forward neural network layer.
[0109] Specifically, the phrase sequence vector is input into the multi-head attention mechanism layer for multi-head attention mechanism learning to obtain the multi-head attention vector. The sum of the multi-head attention vector and the phrase sequence vector is normalized to obtain the intermediate variable. The calculation formula is as follows:
[0110] H=LN(C+MultiHead(C))
[0111] Among them, H is an intermediate variable, Multihead(C) is the multi-head attention vector calculated by the multi-head attention mechanism in the Transformer encoder, and LN represents the hierarchical normalization operation.
[0112] In this embodiment, the intermediate variables are calculated through the feedforward neural network layer to obtain the semantic relationship vector, and the calculation formula is as follows:
[0113] C'=LN(H+FFN(H))
[0114] Among them, FFN(·) is a feedforward neural network consisting of two linear transformations, the middle activation function is ReLU, and C' is the sequence of semantic relationship vectors passed through the Transformer encoder.
[0115] In this embodiment, by performing attention calculations in the encoding unit, the ability of the encoding unit to capture contextual information can be improved, thereby improving the recognition efficiency and generalization ability of the model.
[0116] Step S305: extracting audit information from the semantic relationship vector through the entity recognition unit to obtain the audit entity.
[0117] The semantic relationship vector is classified into entities through the entity recognition unit to obtain the audit entity.
[0118] In some optional implementations, the entity recognition unit includes a fully connected layer and a softmax layer; the semantic relationship vectors are integrated and connected through the fully connected layer to obtain a fully connected semantic vector; the fully connected semantic vector is input into the softmax layer for classification, and the audit entity is output.
[0119] The fully connected layer receives the semantic relationship vector as input, which can further extract and combine features to capture higher-level relationships and patterns, thereby providing a more abstract feature representation for classification tasks; the Softmax layer is a normalization function that converts the fully connected semantic vector into a probability distribution. For each classification category, the Softmax layer outputs a value between 0 and 1, and then determines the entity category based on the probability distribution, thereby outputting the audit entity. The audit entity includes insurance name, subject matter, rate, and rate adjustment coefficient, etc.
[0120] In this embodiment, entity classification and recognition are performed by combining a fully connected layer with a softmax layer, which can provide a powerful feature extraction capability, help extract useful information from the context of the entity, and improve the accuracy of information extraction.
[0121] Step S306: Output the audit entity according to the preset template through the output unit to obtain the content to be audited.
[0122] The formats of different audit entities may be different. In order to facilitate subsequent verification, the audit entity can be output according to the preset template, so as to achieve standardized processing of the audit content. The preset template can be stored in a key-value pair format. Specifically, the audit field is used as the key and the field content corresponding to the audit field is used as the value corresponding to the key; a list can also be created according to the mapping relationship between the audit field and the corresponding field content for storage.
[0123] By using the field extraction model to extract fields from documents to be reviewed, the extraction efficiency can be improved. The model can also support field recognition and extraction of multiple elements, such as fields / tables / titles, further improving the extraction accuracy.
[0124] Step S202, verify the validity of the fields of the content to be reviewed according to preset verification rules to obtain valid review content.
[0125] Field validity verification includes but is not limited to text structure analysis, audit field consistency, consistency of results for the same field, rate upper and lower limit verification, and whether the product name structure is consistent or contains errors.
[0126] Among them, text structure analysis involves rate table files, which require upper and lower limits of rate adjustment coefficients. For example, the file states that "rate adjustment coefficients shall not be higher than 1.5 or lower than 0.5", so it is necessary to extract the sizes of all rate adjustment coefficients in the form of tables, paragraphs, and titles in the file and determine whether they meet the requirements; the consistency of audit fields means that the audit fields are consistent in all internal materials, for example, the subject field, the subject mentioned in any text content of clauses, rates, actuarial reports, feasibility analysis reports, and government documents must be consistent; the consistency of results of the same field means that the results obtained by different calculation methods of the same calculation must be consistent, for example If the base rate = expense surcharge rate, then the values of the audited entity's pure risk loss rate, base rate, and surcharge rate identified in the actuarial report are used to calculate the base rate = risk surcharge rate + profit surcharge rate + pure risk loss rate to see if the two results are consistent; the rate upper and lower limit verification is used to verify whether the rate upper and lower limits meet the requirements, for example, to determine whether the extracted rate-related values meet the policy requirements of 0.75-1.25 and commercial requirements of 0.5-1.5; whether the product name structure is consistent or there are errors is used to verify whether the product name structure in the audited document meets the specifications, for example, to conduct a standardization check on the names of clauses, rate tables, actuarial reports, and feasibility reports.
[0127] In some optional implementations, the step of verifying the validity of the fields of the content to be reviewed according to the preset verification rules to obtain the valid review content includes:
[0128] Extract the text fields of the content to be reviewed according to the preset regular expression to obtain the review fields and the corresponding field contents;
[0129] Verify the audit field and the field content based on the preset verification rules to obtain the verification result;
[0130] Based on the verification results, the content to be reviewed is modified to obtain valid review content.
[0131] Among them, the preset regular expression represents a regular expression for audit field matching. The preset regular expression can match the corresponding audit field and extract the corresponding field content.
[0132] In this embodiment, the preset verification rules include but are not limited to whether the text structure meets the requirements, the consistency of the audit fields, the consistency of the results of the same fields, whether the upper and lower limits of the rates meet the regulations, and whether the upper and lower limits of the product name structure meet the specifications.
[0133] According to the verification results, the interference fields such as historical fields, intermediate values of calculations and non-standard objects in the audit field are determined, and the interference fields are eliminated to obtain the valid audit content.
[0134] Among them, the intermediate value of the operation is a temporary value generated in the calculation process before the final result is obtained in the clauses involving complex calculations, such as insurance premiums, compensation amounts, etc. For example, in the calculation of the base rate, the description of the original file may be "The base rate is calculated as follows: (1-90%) = 10%, plus other additional ratios It is 10% + 20% = 30%", then 10% is the intermediate value of the operation.
[0135] By verifying the validity of fields, you can improve data quality, ensure data consistency, and thus improve audit efficiency.
[0136] Step S203, based on the preset matching rules, the effective audit content is matched and verified with the audit reference information to obtain the audit result.
[0137] In this embodiment, the effective audit content is matched with the audit reference information based on the preset matching rules, and whether the effective audit content meets the compliance conditions is determined according to the matching results, that is, the audit result is to determine whether the effective audit content meets the compliance conditions.
[0138] Among them, the preset matching rules include semantic fuzzy matching, field precise matching and operation rule matching. Semantic fuzzy matching can be audited from the semantic level, and can identify synonyms, match the same concepts, handle polysemous words, semantic matching, etc., to avoid misjudgment and improve the accuracy of audit results. Field precise matching means that the field values of the same field must be completely consistent, and can be compared through string matching. Operation rule matching means calculating a given field to determine whether the calculation result meets the compliance requirements.
[0139] Based on preset matching rules, compliance conditions include terminology consistency, concept consistency, polysemous word matching, field consistency, format compliance, field completeness, field value accuracy, calculation correctness, logical consistency, and formula accuracy.
[0140] In some optional implementations, the above step of matching and verifying the effective audit content with the audit reference information based on the preset matching rule to obtain the audit result includes:
[0141] Extract all valid audit fields and corresponding audit field values in the valid audit content;
[0142] Classify the valid audit fields and obtain the field classification results;
[0143] Obtain the corresponding field reference value from the audit reference information according to the valid audit field;
[0144] According to the field classification results, the audit field value is matched with the field reference value according to the corresponding preset matching rules to obtain the matching degree;
[0145] The audit results are obtained based on the matching degree.
[0146] Among them, the valid review fields include insurance name, subject matter, rate, upper and lower limits of rates, rate adjustment coefficient, etc.
[0147] In this embodiment, the audit fields are classified in order to determine the matching rules corresponding to the audit fields. The field classification results include semantic matching fields, precise matching fields, and operation matching fields. The audit field values corresponding to the semantic matching fields are matched through semantic fuzzy matching rules, the audit field values corresponding to the precise matching fields are matched through field precise matching rules, and the audit field values corresponding to the operation matching fields are matched through operation rule matching rules, such as rate calculation rules.
[0148] It should be understood that the same audit field can be matched using different matching rules.
[0149] In some optional implementations, semantic fuzzy matching can be performed by calculating the similarity between the audit field value and the field reference value as the matching degree. Field exact matching can be performed by using the Jaro Winkler algorithm (string similarity comparison algorithm) to calculate the matching degree.
[0150] The preset matching rules of this application support multi-dimensional verification, including semantic matching, field matching and operation matching, which can improve the coverage of verification points, while improving the verification accuracy and reducing the possibility of risk leakage.
[0151] In some optional implementations, after obtaining the audit result, the following steps are also included:
[0152] If there are errors in the review results, the location of the error will be accurately marked in the document to be reviewed, and annotations will be added to explain the cause of the error. At the same time, GPT (Generative Pre-trained Transformer) will be used to summarize and facilitate manual inspection and correction.
[0153] This application uses a trained field extraction model to perform semantic recognition on the documents to be reviewed, so that the field extraction model can better understand the context and semantics of the text, comprehensively cover the review content, reduce the possibility of risk leakage, and improve the efficiency and accuracy of recognition. At the same time, by verifying the validity of the fields, it helps to optimize the review process and further improve the review efficiency. Finally, matching verification is performed through preset matching rules, which can save manpower investment and improve the overall review efficiency and accuracy.
[0154] In some optional implementations, the text segmentation unit includes a convolutional neural network layer, a region candidate network layer, an interest region pooling layer, an attention layer, and a classification layer; the step of inputting the document to be reviewed into the text segmentation unit for element recognition to obtain the text region and the non-text region includes:
[0155] The feature extraction of the document to be reviewed is performed through a convolutional neural network to obtain the document feature representation;
[0156] The document feature representation is input into the region candidate network layer, and the candidate region features are obtained through the region candidate network layer;
[0157] The document feature representation and the candidate region feature are input into the interest region pooling layer for pooling operation to obtain the document pooling feature;
[0158] The attention layer calculates the attention of the document pooling feature to obtain the attention enhancement feature, and then fuses the attention enhancement feature with the document pooling feature to obtain the fusion feature;
[0159] The fused features are classified through the classification layer to obtain the text area and non-text area corresponding to the document to be reviewed.
[0160] Among them, the convolutional neural network is used for feature extraction. In some optional implementation methods, the convolutional neural network includes multiple convolutional sublayers, residual sublayers, multi-head global attention sublayers, multiple pooling sublayers and fully connected layers; the convolutional sublayer is used to extract features of the document to be reviewed to obtain convolutional features; the residual sublayer can adopt a convolutional structure to expand the convolutional features and enhance the multi-scale expression ability of the features. Specifically, the residual sublayer is used to perform residual connection on the convolutional features to obtain enhanced features; the multi-head global attention sublayer can better learn the features of different aspects of the world, so as to obtain features with a larger receptive field, which can meet the changing reality and have good robustness. Specifically, the enhanced features are input into the multi-head global attention sublayer, and the attention is calculated by the multi-head global attention sublayer to obtain the global attention features; after the enhanced features and the global attention features are fused, the enhanced features obtained are pooled through the pooling sublayer to reduce the feature dimension to obtain the pooled features; the pooled features are connected and output through the fully connected layer to obtain the document feature representation. For example, if the document to be reviewed is a text file, the document features are represented as text data features; if the document to be reviewed is an image file, the document features are represented as a document feature map.
[0161] The region candidate network layer is mainly used to generate detection frames, which can greatly improve the generation speed of detection frames. The structure of the region candidate network can be divided into two branches. One uses Softmax to divide the detection frames into two categories: Positive and Negative. The other is used to calculate the offset of the detection frame to obtain a more accurate detection frame. The final Proposal layer is responsible for combining the Positive detection frame and the offset to obtain the final detection frame, while eliminating the detection frame that is too small or exceeds the boundary.
[0162] The region of interest pooling layer is used to collect the input document feature representation and candidate region features, and extract pooled features (proposal feature maps) after integrating this information.
[0163] The attention layer is connected to the region candidate network layer and the interest region pooling layer respectively, and is used to perform attention calculation on the candidate region features and the document pooling features to obtain the attention enhancement features. The enhanced pooling fusion features after fusing the attention enhancement features and the pooling features are input into the subsequent classification layer to discriminate the target category, thereby outputting the element label, that is, the recognition result.
[0164] The classification layer uses the enhanced pooling fusion features to calculate the category of each enhanced pooling fusion feature (such as title, subtitle, text, image, table, etc.) through the full connect layer and the softmax activation function, and outputs the cls_prob probability vector; at the same time, the bounding box regression is used again to obtain the position offset bbox_pred of each pooling feature, which is used to regress a more accurate target detection box, and determine the text area and non-text area according to the target detection box.
[0165] By using the text segmentation unit to identify elements of the document to be reviewed, the identification object can be accurately located, high-precision detection can be achieved, the accuracy of recognition can be improved, and multiple objects can be identified at the same time, improving the efficiency of recognition.
[0166] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.
[0167] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).
[0168] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.
[0169] Further references Figure 4 , as a response to the above Figure 2 The present application provides an embodiment of an intelligent document review device, and the device embodiment is Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.
[0170] like Figure 4As shown, the intelligent document review device 400 described in this embodiment includes: an identification and extraction module 401, a validity verification module 402, and a review module 403. Among them:
[0171] The recognition and extraction module 401 is used to receive the document to be reviewed, input the document to be reviewed into the trained field extraction model to perform semantic recognition and extraction, and obtain the content to be reviewed;
[0172] The validity check module 402 is used to check the field validity of the content to be reviewed according to the preset verification rules to obtain the valid review content;
[0173] The audit module 403 is used to call the audit reference information, and match and verify the effective audit content with the audit reference information based on a preset matching rule to obtain an audit result.
[0174] It should be emphasized that in order to further ensure the privacy and security of the documents to be reviewed, the above documents to be reviewed can also be stored in a node of a blockchain.
[0175] Based on the above-mentioned intelligent document review device 400, semantic recognition is performed on the document to be reviewed through the trained field extraction model, so that the field extraction model can better understand the context and semantics of the text, comprehensively cover the review content, reduce the possibility of risk leakage, and improve the efficiency and accuracy of recognition. At the same time, by verifying the validity of the field, it helps to optimize the review process and further improve the review efficiency. Finally, matching verification is performed through preset matching rules, which can save manpower investment and improve the overall review efficiency and accuracy.
[0176] In some optional implementations, the field extraction model includes a text segmentation unit, a text detection unit, a text recognition unit, an encoding unit, an entity recognition unit, and an output unit; the recognition and extraction module 401 includes an element segmentation submodule, a text detection submodule, a text recognition submodule, an attention submodule, an entity recognition submodule, and an output submodule, wherein:
[0177] The element segmentation submodule is used to input the document to be reviewed into the text segmentation unit for element recognition to obtain text areas and non-text areas;
[0178] The text detection submodule is used to perform text detection on the text area and the non-text area through the text detection unit to obtain a global text feature containing text position information;
[0179] The text recognition submodule is used to input the global features of the text into the text recognition unit for recognition to obtain a text recognition result;
[0180] The attention submodule is used to perform syntactic analysis on the text recognition result to obtain semantic phrases, and input the semantic phrases into the encoding unit to perform attention calculation to obtain semantic relationship vectors;
[0181] The entity recognition submodule is used to extract audit information from the semantic relationship vector through the entity recognition unit to obtain the audit entity;
[0182] The output submodule is used to output the audit entity according to a preset template through the output unit to obtain the content to be audited.
[0183] By using the field extraction model to extract fields from documents to be reviewed, the extraction efficiency can be improved. The model can also support field recognition and extraction of multiple elements, such as fields / tables / titles, further improving the extraction accuracy.
[0184] In some optional implementations, the text segmentation unit includes a convolutional neural network layer, a region candidate network layer, an interest region pooling layer, an attention layer, and a classification layer; and the element segmentation submodule is further used for:
[0185] Extracting features of the document to be reviewed by using the convolutional neural network to obtain a document feature representation;
[0186] Inputting the document feature representation into the region candidate network layer, and obtaining candidate region features through the region candidate network layer;
[0187] Inputting the document feature representation and the candidate region feature into the region of interest pooling layer for pooling operation to obtain document pooling features;
[0188] Performing attention calculation on the document pooling feature through the attention layer to obtain an attention enhancement feature, and fusing the attention enhancement feature with the document pooling feature to obtain an enhanced pooling fusion feature;
[0189] The enhanced pooling fusion features are classified by the classification layer to obtain the text area and non-text area corresponding to the document to be reviewed.
[0190] By using the text segmentation unit to identify elements of the document to be reviewed, the identification object can be accurately located, high-precision detection can be achieved, the accuracy of recognition can be improved, and multiple objects can be identified at the same time, thereby improving the efficiency of recognition.
[0191] In some optional implementations, the encoding unit includes a vector embedding layer, an average pooling layer, a multi-layer perceptron layer, and an encoder layer; the encoding unit is further used to:
[0192] Inputting the semantic phrase into the vector embedding layer for embedding to obtain a coding sequence vector;
[0193] Performing a pooling operation on the encoding sequence vector through the average pooling layer to obtain a pooled semantic code;
[0194] Used to input the pooled semantic code and the code sequence vector into the multi-layer perceptron layer for weighted calculation to obtain a phrase sequence vector;
[0195] Used to perform attention calculation on the phrase sequence vector through the encoder layer to obtain a semantic relationship vector.
[0196] By performing attention calculations on the encoding unit, the ability of the encoding unit to capture contextual information can be improved, thereby improving the recognition efficiency and generalization ability of the model.
[0197] In some optional implementations, the entity recognition unit includes a fully connected layer and a softmax layer; the entity recognition submodule is further used to:
[0198] The semantic relationship vectors are integrated and connected through the fully connected layer to obtain a fully connected semantic vector;
[0199] The fully connected semantic vector is input into the softmax layer for classification, and the audit entity is output.
[0200] Entity classification and recognition through the fully connected layer combined with the softmax layer can provide powerful feature extraction capabilities, help extract useful information from the context of the entity, and improve the accuracy of information extraction.
[0201] In some optional implementations, the validity verification module 402 includes an extraction submodule, a verification submodule, and a correction submodule, wherein:
[0202] The extraction submodule is used to extract the text field of the content to be reviewed according to a preset regular expression to obtain the review field and the corresponding field content;
[0203] The verification submodule is used to verify the audit field and the field content based on the preset verification rules to obtain the verification result;
[0204] The correction submodule is used to correct the content to be reviewed based on the verification result to obtain valid review content.
[0205] By verifying the validity of fields, you can improve data quality, ensure data consistency, and thus improve audit efficiency.
[0206] In some optional implementations, the audit module 403 includes a content extraction submodule, a classification submodule, an acquisition submodule, a matching submodule, and an audit submodule, wherein:
[0207] The content extraction submodule is used to extract all valid audit fields and corresponding audit field values in the valid audit content;
[0208] The classification submodule is used to classify the valid audit fields to obtain field classification results;
[0209] The acquisition submodule is used to acquire the corresponding field reference value from the audit reference information according to the valid audit field;
[0210] The matching submodule is used to match the audit field value with the field reference value according to the field classification result and the corresponding preset matching rule to obtain a matching degree;
[0211] The audit submodule is used to obtain the audit result based on the matching degree.
[0212] The preset matching rules of this application support multi-dimensional verification, including semantic matching, field matching and operation matching, which can improve the coverage of verification points, while improving the verification accuracy and reducing the possibility of risk leakage.
[0213] To solve the above technical problems, the present application also provides a computer device. Figure 5 , Figure 5 This is a basic structural block diagram of the computer device in this embodiment.
[0214] The computer device 5 includes a memory 51, a processor 52, and a network interface 53 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 5 with components 51-53, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.
[0215] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.
[0216] The memory 51 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 51 can be an internal storage unit of the computer device 5, such as a hard disk or memory of the computer device 5. In other embodiments, the memory 51 can also be an external storage device of the computer device 5, such as a plug-in hard disk equipped on the computer device 5, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. Of course, the memory 51 can also include both the internal storage unit of the computer device 5 and its external storage device. In this embodiment, the memory 51 is generally used to store the operating system and various application software installed on the computer device 5, such as computer-readable instructions of the intelligent document review method, etc. In addition, the memory 51 can also be used to temporarily store various types of data that have been output or are to be output.
[0217] The processor 52 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 52 is generally used to control the overall operation of the computer device 5. In this embodiment, the processor 52 is used to run computer-readable instructions stored in the memory 51 or process data, such as computer-readable instructions for running the intelligent document audit method.
[0218] The network interface 53 may include a wireless network interface or a wired network interface. The network interface 53 is generally used to establish a communication connection between the computer device 5 and other electronic devices.
[0219] This embodiment implements the steps of the intelligent document audit method in the above embodiment when the processor executes computer-readable instructions stored in the memory, and performs semantic recognition on the document to be audited through the trained field extraction model, so that the field extraction model can better understand the context and semantics of the text, comprehensively cover the audit content, reduce the possibility of risk leakage, and improve the efficiency and accuracy of recognition. At the same time, by verifying the validity of the field, it helps to optimize the audit process and further improve the audit efficiency. Finally, matching verification is performed through preset matching rules, which can save manpower investment and improve the overall audit efficiency and accuracy.
[0220] The present application also provides another implementation mode, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor, so that the at least one processor executes the steps of the intelligent document review method as described above, and performs semantic recognition on the document to be reviewed through the trained field extraction model, so that the field extraction model can better understand the context and semantics of the text, comprehensively cover the review content, reduce the possibility of risk leakage, and improve the efficiency and accuracy of recognition. At the same time, by verifying the validity of the field, it helps to optimize the review process and further improve the review efficiency. Finally, matching verification is performed through preset matching rules, which can save manpower investment and improve the overall review efficiency and accuracy.
[0221] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.
[0222] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.
Claims
1. An intelligent document review method, characterized in that: The steps include: Receiving a document to be reviewed, and inputting the document to be reviewed into a trained field extraction model for semantic recognition and extraction to obtain content to be reviewed; Verify the validity of the fields of the content to be reviewed according to the preset verification rules to obtain the valid review content; The audit reference information is called, and the valid audit content is matched and verified with the audit reference information based on a preset matching rule to obtain an audit result.
2. The intelligent document review method according to claim 1, characterized in that: The field extraction model includes a text segmentation unit, a text detection unit, a text recognition unit, an encoding unit, an entity recognition unit and an output unit; the step of inputting the document to be reviewed into the trained field extraction model for semantic recognition and extraction to obtain the content to be reviewed includes: Inputting the document to be reviewed into the text segmentation unit for element recognition to obtain text areas and non-text areas; Performing text detection on the text area and the non-text area by the text detection unit to obtain a global text feature containing text position information; Inputting the global features of the text into the text recognition unit for recognition to obtain a text recognition result; Performing syntactic analysis on the text recognition result to obtain semantic phrases, and inputting the semantic phrases into the encoding unit to perform attention calculation to obtain semantic relationship vectors; Extracting audit information from the semantic relationship vector by the entity recognition unit to obtain an audit entity; The audit entity is outputted according to a preset template through the output unit to obtain the content to be audited.
3. The intelligent document review method according to claim 2, characterized in that: The text segmentation unit includes a convolutional neural network layer, a region candidate network layer, an interest region pooling layer, an attention layer, and a classification layer; the step of inputting the document to be reviewed into the text segmentation unit for element recognition to obtain text regions and non-text regions includes: Extracting features of the document to be reviewed by using the convolutional neural network to obtain a document feature representation; Inputting the document feature representation into the region candidate network layer, and obtaining candidate region features through the region candidate network layer; Inputting the document feature representation and the candidate region feature into the region of interest pooling layer for pooling operation to obtain document pooling features; Performing attention calculation on the document pooling feature through the attention layer to obtain an attention enhancement feature, and fusing the attention enhancement feature with the document pooling feature to obtain an enhanced pooling fusion feature; The enhanced pooling fusion features are classified by the classification layer to obtain the text area and non-text area corresponding to the document to be reviewed.
4. The intelligent document review method according to claim 2, characterized in that: The encoding unit includes a vector embedding layer, an average pooling layer, a multi-layer perceptron layer and an encoder layer; the step of inputting the semantic phrase into the encoding unit for attention calculation to obtain a semantic relationship vector includes: Inputting the semantic phrase into the vector embedding layer for embedding to obtain a coding sequence vector; Performing a pooling operation on the encoding sequence vector through the average pooling layer to obtain a pooled semantic code; Inputting the pooled semantic code and the code sequence vector into the multi-layer perceptron layer for weighted calculation to obtain a phrase sequence vector; The encoder layer performs attention calculation on the phrase sequence vector to obtain a semantic relationship vector.
5. The intelligent document review method according to claim 2, characterized in that: The entity recognition unit includes a fully connected layer and a softmax layer; the step of extracting audit information from the semantic relationship vector through the entity recognition unit to obtain the audit entity includes: The semantic relationship vectors are integrated and connected through the fully connected layer to obtain a fully connected semantic vector; The fully connected semantic vector is input into the softmax layer for classification, and the audit entity is output.
6. The intelligent document review method according to claim 1, characterized in that: The step of verifying the field validity of the content to be reviewed according to the preset verification rules to obtain the valid review content includes: Extracting the text field of the content to be reviewed according to a preset regular expression to obtain the review field and the corresponding field content; Verify the audit field and the field content based on preset verification rules to obtain a verification result; The content to be reviewed is modified based on the verification result to obtain valid review content.
7. The intelligent document review method according to claim 1, characterized in that: The step of matching and verifying the valid audit content with the audit reference information based on a preset matching rule to obtain an audit result includes: Extract all valid audit fields and corresponding audit field values in the valid audit content; Classify the valid audit fields to obtain field classification results; Acquire a corresponding field reference value from the audit reference information according to the valid audit field; According to the field classification result, the audit field value is matched with the field reference value according to the corresponding preset matching rule to obtain a matching degree; An audit result is obtained based on the matching degree.
8. An intelligent document review device, characterized in that: include: The recognition and extraction module is used to receive the document to be reviewed, input the document to be reviewed into the trained field extraction model for semantic recognition and extraction, and obtain the content to be reviewed; The validity verification module is used to verify the field validity of the content to be reviewed according to the preset verification rules to obtain the valid review content; The audit module is used to call the audit reference information, match and verify the effective audit content with the audit reference information based on a preset matching rule, and obtain the audit result.
9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the intelligent document review method as claimed in any one of claims 1 to 7 when executing the computer-readable instructions.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the intelligent document review method according to any one of claims 1 to 7 are implemented.
Citation Information
Cited By
Official document element extraction and auditing system based on large model
CN119808759A
A document element extraction and review system based on a large model
CN119808759B
Marine operation task intelligent auditing method, device, equipment and medium
CN120013489A
Contract auditing-oriented custom text rule generation and auditing method and system and storage medium
CN120725002A
Work log intelligent auditing and optimizing method and system based on artificial intelligence
CN121009902A
