Character recognition and automatic auditing method and system

Abstract

The invention discloses a character recognition and automatic auditing method and system, and the method comprises the steps: obtaining all to-be-audited document data including semi-structured and non-structured document data, guaranteeing the comprehensiveness of an auditing process, accurately recognizing the document content through an OCR technology, converting the document content into operable structured data, and carrying out the automatic auditing of the to-be-audited document data. Key information in a document is selected through two modes of rule extraction and model extraction, the efficiency and accuracy of information processing are improved, the information is compared with pre-established target structured text data, compliance check is automatically completed, the compliance of an auditing result is ensured, a confidence coefficient threshold value is updated in real time according to the complexity of the document, and the accuracy of information processing is improved. The method is suitable for document data with different complexities, the automatic auditing process remarkably improves the auditing efficiency, when a large amount of document data is processed, a large amount of time and manpower resources can be saved, the auditing result and the corresponding confidence rate can be output, a user can see the specific result of document analysis, and the auditing reliability and the credibility of the user are improved.

Description

Technical Field

[0001] The present application relates to the technical field of text recognition and automated review, and in particular to a text recognition and automated review method and system. Background Art

[0002] With the advancement of digitalization, enterprises and institutions are increasingly relying on automation technology to process and review large amounts of unstructured document data (such as images, tables, and text documents). Traditional technical solutions mainly rely on manual review, which involves manually reviewing paper documents one by one, including checking invoices, contracts, audit reports, etc. This is inefficient, time-consuming, and error-prone. Manual review methods require a lot of manpower, are time-consuming, and labor-intensive, and are difficult to cope with the processing needs of large-scale document data. Manual review methods are prone to human errors, especially when faced with a large number of repetitive tasks, and the error rate is high. Therefore, existing technical solutions are difficult to meet the needs of modern enterprises for efficient, accurate, flexible, and comprehensive data processing. Summary of the invention

[0003] The present application provides a text recognition and automated review method and system, aiming to solve the problem that existing technologies are difficult to meet the needs of modern enterprises for efficient, accurate, flexible and comprehensive data processing.

[0004] In a first aspect, a method for text recognition and automated review is provided, the method comprising:

[0005] Acquire document data to be reviewed, wherein the document data to be reviewed includes semi-structured document data and unstructured text data;

[0006] Obtaining a corresponding OCR engine according to the content of the document data to be reviewed, and using the corresponding OCR engine to recognize the document data to be reviewed to obtain document data;

[0007] According to the document data, a document extraction type is selected to extract target document data, wherein the target document data format is structured text data, wherein the document extraction type includes rule extraction and model extraction;

[0008] By comparing the target document data with pre-built target structured text data, determining whether the target document data is compliant;

[0009] According to the complexity of the document data to be reviewed, the confidence threshold of the target document data is updated in real time;

[0010] The compliance rate of the target document data is output according to whether the target document data is compliant and a confidence threshold of the target document data.

[0011] In the above solution, optionally, the semi-structured text data includes table document data, and the unstructured text data includes image document data and text document data.

[0012] In the above solution, optionally, obtaining a corresponding OCR engine according to the data content of the document to be reviewed includes:

[0013] The document data to be reviewed is table document data and the corresponding OCR engine is ABBYY FineReader;

[0014] The document data to be reviewed is image document data, and the corresponding OCR engine is Google Vision API or Microsoft Azure Computer Vision;

[0015] The document data to be reviewed is text document data and the corresponding OCR engine is Tesseract or ABBYY FineReader.

[0016] In the above solution, optionally, the rule extraction includes the following steps:

[0017] Construct character patterns. Use character-based pattern extraction to take the input semi-structured text data as a character sequence to construct character patterns.

[0018] The constructed character pattern is used to match the input file. If a match is found, the extracted target document data is stored.

[0019] In the above solution, optionally, the model extraction includes the following steps:

[0020] Calculate the features of the input corpus based on the Bert pre-trained model, and use Bert's built-in word segmenter to segment the input unstructured text data;

[0021] After summarizing the three-dimensional features produced by the Bert pre-training model, they are sent to the CRF layer for optimal decoding and the maximum score sequence is selected;

[0022] Use the selected model to perform calculations, extract the target structure field information, and generate the target document data.

[0023] In the above solution, optionally, determining whether the target document data is compliant by comparing the target document data with pre-built target structured text data includes the following steps:

[0024] Verify the consistency between the extracted document data to be reviewed and the existing data in the system, and check whether the extracted target document data complies with the preset business rules and logic. If so, it is determined to be compliant.

[0025] In the above solution, optionally, updating the confidence threshold of the target document data in real time according to the complexity of the document data to be reviewed includes:

[0026] The confidence threshold is adjusted in real time according to the quality of the input image corresponding to the document data to be reviewed, the complexity of the text, and the language type factors.

[0027] In the above solution, optionally, the confidence threshold adjustment method includes the following steps:

[0028] Collect OCR sample data with different characteristics, including images of different quality, texts of different complexity, and samples of different language types;

[0029] Manually label the collected sample data to ensure the accuracy of the labeling and provide a reliable benchmark for model training and threshold setting;

[0030] Developing algorithms to identify key features of input images, including image quality, text complexity, and language type;

[0031] Conduct statistical analysis on the labeled sample data to determine the recognition accuracy and error rate under different feature combinations;

[0032] According to the results of statistical analysis, the initial confidence threshold range is set for different feature combinations;

[0033] Use the annotated and analyzed data to train the OCR model, including a dynamic adjustment mechanism for the confidence threshold;

[0034] In the actual OCR recognition process, the confidence threshold is dynamically adjusted according to the characteristics of the input image;

[0035] Evaluate the recognition results of the OCR model to determine whether they meet the set confidence threshold;

[0036] Collect user feedback or automatically identify errors and analyze the characteristic conditions under which errors occur;

[0037] Adjust the threshold setting based on error feedback and recognition results.

[0038] In the above scheme, optionally, the method further comprises: matching and verifying the results of OCR recognition with entities, relationships and attributes in the knowledge graph, and verifying based on contextual semantic understanding;

[0039] The knowledge graph matching and verification includes the following steps:

[0040] Match and verify the OCR recognition results with the entities, relationships, and attributes in the knowledge graph;

[0041] Natural language processing technology is used to analyze and identify the contextual semantics of the text. By building a language model, the correct vocabulary and sentence structure that are more likely to appear in a specific context are predicted, thereby improving the accuracy of verification.

[0042] In a second aspect, a text recognition and automated review system is provided, the system comprising:

[0043] Acquisition module: used to acquire document data to be reviewed, wherein the document data to be reviewed includes semi-structured document data and unstructured text data;

[0044] Identification module: used to obtain the corresponding OCR engine according to the content of the document data to be reviewed, and use the corresponding OCR engine to recognize the document data to be reviewed to obtain document data;

[0045] Extraction module: used for selecting a document extraction type to extract target document data according to the document data, wherein the target document data format is structured text data, wherein the document extraction type includes rule extraction and model extraction;

[0046] Comparison module: used to determine whether the target document data is compliant by comparing the target document data with pre-built target structured text data;

[0047] Dynamic confidence threshold adjustment module: used to update the confidence threshold of the target document data in real time according to the complexity of the document data to be reviewed;

[0048] Output module: used for outputting the compliance rate of the target document data according to whether the target document data is compliant and the confidence threshold of the target document data.

[0049] Compared with the prior art, this application has at least the following beneficial effects:

[0050] Based on further analysis and research of existing technical problems, this application recognizes that existing solutions are difficult to meet the needs of modern enterprises for efficient, accurate, flexible and comprehensive data processing. By obtaining all document data to be audited, including semi-structured and unstructured data, the comprehensiveness of the audit process is ensured, and omissions that may occur in manual audits are avoided. The OCR technology is used to accurately identify the content of the document and convert it into operable structured data, which improves the accuracy of the audit and reduces human errors. Through rule extraction and model extraction, key information in the document is selected to improve the efficiency and accuracy of information processing. It is compared with the pre-built target structured text data to automatically complete the compliance check to ensure the compliance of the audit results. The confidence threshold is updated in real time according to the complexity of the document, making the audit process more flexible and adaptable to document data of different complexities. The automated audit process significantly improves the audit efficiency, especially when processing a large amount of document data. Compared with manual audit, it can save a lot of time and human resources, output the audit results and the corresponding confidence rate, so that users can see the specific results of document parsing and perform manual intervention when necessary, which improves the reliability of the audit and the trust of users.

[0051] This application text recognition and automated audit method greatly improves the efficiency, accuracy and flexibility of the audit by combining advanced OCR technology, intelligent extraction, compliance checking and dynamic threshold update technologies, and meets the modern enterprise's needs for efficient, accurate, flexible and comprehensive data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a text recognition and automated review method provided for one embodiment of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0054] In the description of this application: unless otherwise specified, the meaning of "plurality" is two or more. The terms "first", "second", "third", etc. in this application are intended to distinguish the objects referred to, and do not have special meanings in terms of technical connotations (for example, they should not be understood as emphasizing the importance or order, etc.). Expressions such as "including", "comprising", "having", etc. also mean "not limited to" (certain units, components, materials, steps, etc.).

[0055] In one embodiment, Figure 1 As shown, a text recognition and automatic review method is provided, comprising the following steps:

[0056] Acquire document data to be reviewed, wherein the document data to be reviewed includes semi-structured document data and unstructured text data;

[0057] Obtaining a corresponding OCR engine according to the content of the document data to be reviewed, and using the corresponding OCR engine to recognize the document data to be reviewed to obtain document data;

[0058] According to the document data, a document extraction type is selected to extract target document data, wherein the target document data format is structured text data, wherein the document extraction type includes rule extraction and model extraction;

[0059] By comparing the target document data with pre-built target structured text data, determining whether the target document data is compliant;

[0060] According to the complexity of the document data to be reviewed, the confidence threshold of the target document data is updated in real time;

[0061] The compliance rate of the target document data is output according to whether the target document data is compliant and a confidence threshold of the target document data.

[0062] In this embodiment, the semi-structured text data includes table document data, and the unstructured text data includes picture document data and text document data.

[0063] In this embodiment, obtaining a corresponding OCR engine according to the content of the document to be reviewed includes:

[0064] The document data to be reviewed is table document data and the corresponding OCR engine is ABBYY FineReader;

[0065] The document data to be reviewed is image document data, and the corresponding OCR engine is Google Vision API or Microsoft Azure Computer Vision;

[0066] The document data to be reviewed is text document data and the corresponding OCR engine is Tesseract or ABBYY FineReader.

[0067] In this embodiment, the rule extraction includes the following steps:

[0068] Construct character patterns. Use character-based pattern extraction to take the input semi-structured text data as a character sequence to construct character patterns.

[0069] The constructed character pattern is used to match the input file. If a match is found, the extracted target document data is stored.

[0070] In this embodiment, the model extraction includes the following steps:

[0071] Calculate the features of the input corpus based on the Bert pre-trained model, and use Bert's built-in word segmenter to segment the input unstructured text data;

[0072] After summarizing the three-dimensional features produced by the Bert pre-training model, they are sent to the CRF layer for optimal decoding and the maximum score sequence is selected;

[0073] Use the selected model to perform calculations, extract the target structure field information, and generate the target document data.

[0074] In this embodiment, the step of comparing the target document data with the pre-built target structured text data to determine whether the target document data is compliant includes the following steps:

[0075] Verify the consistency between the extracted document data to be reviewed and the existing data in the system, and check whether the extracted target document data complies with the preset business rules and logic. If so, it is determined to be compliant.

[0076] In this embodiment, the confidence threshold of the target document data is updated in real time according to the complexity of the document data to be reviewed, including:

[0077] The confidence threshold is adjusted in real time according to the quality of the input image corresponding to the document data to be reviewed, the complexity of the text, and the language type factors.

[0078] In this embodiment, the confidence threshold adjustment method includes the following steps:

[0079] Collect OCR sample data with different characteristics, including images of different quality, texts of different complexity, and samples of different language types;

[0080] Manually label the collected sample data to ensure the accuracy of the labeling and provide a reliable benchmark for model training and threshold setting;

[0081] Developing algorithms to identify key features of input images, including image quality, text complexity, and language type;

[0082] Conduct statistical analysis on the labeled sample data to determine the recognition accuracy and error rate under different feature combinations;

[0083] According to the results of statistical analysis, the initial confidence threshold range is set for different feature combinations;

[0084] Use the annotated and analyzed data to train the OCR model, including a dynamic adjustment mechanism for the confidence threshold;

[0085] In the actual OCR recognition process, the confidence threshold is dynamically adjusted according to the characteristics of the input image;

[0086] Evaluate the recognition results of the OCR model to determine whether they meet the set confidence threshold;

[0087] Collect user feedback or automatically identify errors and analyze the characteristic conditions under which errors occur;

[0088] Adjust the threshold setting based on error feedback and recognition results.

[0089] In this embodiment, the method further includes: matching and verifying the results of OCR recognition with entities, relationships and attributes in the knowledge graph, and verifying based on contextual semantic understanding;

[0090] The knowledge graph matching and verification includes the following steps:

[0091] Match and verify the OCR recognition results with the entities, relationships, and attributes in the knowledge graph;

[0092] Natural language processing technology is used to analyze and identify the contextual semantics of the text. By building a language model, the correct vocabulary and sentence structure that are more likely to appear in a specific context are predicted, thereby improving the accuracy of verification.

[0093] This embodiment sets up a dynamic confidence threshold instead of a fixed threshold to judge the reliability of the recognition result. The threshold will be adjusted in real time according to factors such as the quality of the input image, the complexity of the text, the language type, etc. A large number of samples with different characteristics (such as input image quality, text complexity, language type, etc.) are collected, and these samples are manually and accurately labeled, and the labeled sample data is analyzed to calculate the recognition accuracy and error rate under different feature combinations. According to the results of data labeling and analysis, the initial confidence threshold range is set for different feature combinations.

[0094] For example, if the input image is detected to be of high quality, simple text, and a common language type, the threshold can be increased to a higher value; conversely, if the image is blurry, the text is complex, and the language is rare, the threshold can be lowered.

[0095] If it is found that the threshold is set too high or too low in certain cases, resulting in an increase in false positives or missed positives, adjust the threshold range and adjustment strategy accordingly.

[0096] By continuously training the model with new data to enable it to dynamically set thresholds more accurately according to input features, a dynamic confidence threshold can be effectively set to improve the reliability and accuracy of OCR recognition results.

[0097] In this embodiment, a knowledge graph containing rich domain knowledge and semantic relationships is constructed.

[0098] Match and verify the results of OCR recognition with the entities, relationships, and attributes in the knowledge graph.

[0099] Verification based on context semantic understanding: Instead of simply comparing the recognition results of individual characters or words, natural language processing techniques are used to analyze the context semantics of the recognized text. By constructing a language model, the correct words and sentence structures that are more likely to appear in a specific context are predicted, thereby improving the accuracy of verification.

[0100] Suppose the following text is recognized by OCR:

[0101] "The weather is very clear today. I plan to go for a walk."

[0102] Through the verification mechanism of context semantic understanding:

[0103] Analyze the part "very clear". Based on common language expressions and context (weather is usually described as "sunny"), it can be judged that "clear" is the wrong recognition result and should be corrected to "sunny".

[0104] Another example:

[0105] "I like to eat pingguo. It is very sweet."

[0106] From the context, "pingguo" does not conform to the normal expression. Combining common fruit names, it can be inferred that it should be "apple".

[0107]

[0108] This embodiment ensures the comprehensiveness of the audit process by obtaining all document data to be audited, including semi-structured and unstructured data, avoids omissions that may occur in manual audits, uses OCR technology to accurately identify document content and converts it into actionable structured data, thereby improving the accuracy of audits and reducing human errors. It selects key information in documents through rule extraction and model extraction, improves the efficiency and accuracy of information processing, compares it with pre-built target structured text data, automatically completes compliance checks, ensures the compliance of audit results, updates confidence thresholds in real time according to the complexity of the documents, makes the audit process more flexible, and adapts to document data of different complexities. The automated audit process significantly improves audit efficiency, especially when processing large amounts of document data. Compared with manual audits, it can save a lot of time and human resources, outputs audit results and corresponding confidence rates, enables users to see the specific results of document parsing, and perform manual intervention when necessary, thereby improving the reliability of audits and user trust.

[0109] The text recognition and automated audit method of this embodiment greatly improves the efficiency, accuracy and flexibility of the audit by combining advanced OCR technology, intelligent extraction, compliance checking and dynamic threshold updating technologies, and meets the needs of modern enterprises for efficient, accurate, flexible and comprehensive data processing.

[0110] In one embodiment, when a large university conducts an annual financial audit, it needs to review a large number of paper invoices, contracts, and audit reports one by one to ensure the accuracy and compliance of financial data. The traditional manual review method is not only time-consuming and labor-intensive, but also prone to errors. In order to solve this problem, the university adopted this technical solution to perform automated document review.

[0111] In order to verify the performance of this technical solution in document review, we designed a series of experiments to evaluate its efficiency, accuracy, and robustness by comparing the traditional manual review method with the automated review method of this technical solution.

[0112] The experimental dataset includes a large number of paper invoices, contracts and other documents from different industries and in different formats. These documents cover a variety of fonts, sizes, tilt angles and lighting conditions to test the generalization ability and robustness of this technical solution.

[0113] Data preparation: Preprocess the experimental data set, including image scanning, denoising, binarization and other steps, for subsequent OCR processing and recognition.

[0114] Model training: Use deep learning technology combined with labeled image data to train text detection and text recognition models.

[0115] OCR processing: Use the trained model to perform OCR processing on the documents in the experimental dataset to extract key information.

[0116] Compare and verify: Compare the key information extracted by OCR with the preset structured format, and start the verification mechanism for verification.

[0117] Traditional manual review method: On average, each reviewer can review about 200 documents per day.

[0118] The automated review method of this technical solution can process about 1,200 documents per day, which improves the review efficiency by about 600%.

[0119] Recognition accuracy:

[0120] In the test data set, the recognition accuracy of this technical solution reached 98.5%. Even in images with complex fonts, sizes, tilt angles and lighting conditions, the recognition accuracy remained above 95%.

[0121] Warehouse entry results and error markings:

[0122] Total characters: 770;

[0123] Number of characters recognized incorrectly: 3;

[0124] Recognition accuracy = (total number of characters - number of characters with recognition errors) / total number of characters * 100% = 99.61%;

[0125] By comparing the experimental data, we can clearly see the significant advantages of this technical solution in document review. It not only improves the review efficiency and reduces costs, but also ensures the accuracy and compliance of financial data. At the same time, its strong robustness enables this technical solution to adapt to various complex document review scenarios, bringing significant economic and social benefits to enterprises.

[0126] This embodiment converts paper documents into editable electronic text based on OCR technology, reducing the workload of manual input and improving work efficiency. At the same time, compared with manual input, OCR technology can more accurately convert characters in images into text, reducing errors caused by human factors.

[0127] OCR technology can quickly convert text in paper documents into electronic text, thus realizing the digitization and durability of documents. This not only facilitates the management and transmission of documents, but also reduces the space occupied by paper documents and the risk of damage.

[0128] OCR technology can quickly process a large number of images or documents in batches and realize automatic text extraction. This is very useful for enterprises and institutions that need to process a large number of documents, extract information and generate reports, greatly improving work efficiency and accuracy.

[0129] In one embodiment, a text recognition and automated review system is provided, comprising the following program modules: an acquisition module: used to acquire document data to be reviewed, wherein the document data to be reviewed includes semi-structured document data and unstructured text data;

[0130] Identification module: used to obtain the corresponding OCR engine according to the content of the document data to be reviewed, and use the corresponding OCR engine to recognize the document data to be reviewed to obtain document data;

[0131] Extraction module: used for selecting a document extraction type to extract target document data according to the document data, wherein the target document data format is structured text data, wherein the document extraction type includes rule extraction and model extraction;

[0132] Comparison module: used to determine whether the target document data is compliant by comparing the target document data with pre-built target structured text data;

[0133] Dynamic confidence threshold adjustment module: used to update the confidence threshold of the target document data in real time according to the complexity of the document data to be reviewed;

[0134] Output module: used for outputting the compliance rate of the target document data according to whether the target document data is compliant and the confidence threshold of the target document data.

[0135] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for text recognition and automated review, characterized in that: The method comprises: Acquire document data to be reviewed, wherein the document data to be reviewed includes semi-structured document data and unstructured text data; Obtaining a corresponding OCR engine according to the content of the document data to be reviewed, and using the corresponding OCR engine to recognize the document data to be reviewed to obtain document data; According to the document data, a document extraction type is selected to extract target document data, wherein the target document data format is structured text data, wherein the document extraction type includes rule extraction and model extraction; By comparing the target document data with pre-built target structured text data, determining whether the target document data is compliant; According to the complexity of the document data to be reviewed, the confidence threshold of the target document data is updated in real time; The compliance rate of the target document data is output according to whether the target document data is compliant and a confidence threshold of the target document data.

2. The method according to claim 1, characterized in that The semi-structured text data includes table document data, and the unstructured text data includes picture document data and text document data.

3. The method according to claim 2, characterized in that Obtaining a corresponding OCR engine according to the data content of the document to be reviewed includes: The document data to be reviewed is table document data and the corresponding OCR engine is ABBYY FineReader; The document data to be reviewed is image document data, and the corresponding OCR engine is Google Vision API or Microsoft Azure Computer Vision; The document data to be reviewed is text document data and the corresponding OCR engine is Tesseract or ABBYY FineReader.

4. The method according to claim 1, characterized in that: The rule extraction comprises the following steps: Construct character patterns. Use character-based pattern extraction to take the input semi-structured text data as a character sequence to construct character patterns. The constructed character pattern is used to match the input file. If a match is found, the extracted target document data is stored.

5. The method according to claim 1, characterized in that The model extraction comprises the following steps: Calculate the features of the input corpus based on the Bert pre-trained model, and use Bert's built-in word segmenter to segment the input unstructured text data; After summarizing the three-dimensional features produced by the Bert pre-training model, they are sent to the CRF layer for optimal decoding and the maximum score sequence is selected; Use the selected model to perform calculations, extract the target structure field information, and generate the target document data.

6. The method according to claim 1, characterized in that Determining whether the target document data is compliant by comparing the target document data with pre-built target structured text data includes the following steps: Verify the consistency between the extracted document data to be reviewed and the existing data in the system, and check whether the extracted target document data complies with the preset business rules and logic. If so, it is determined to be compliant.

7. The method according to claim 1, characterized in that The step of updating the confidence threshold of the target document data in real time according to the complexity of the document data to be reviewed includes: The confidence threshold is adjusted in real time according to the quality of the input image corresponding to the document data to be reviewed, the complexity of the text, and the language type factors.

8. The method according to claim 7, characterized in that The confidence threshold adjustment method comprises the following steps: Collect OCR sample data with different characteristics, including images of different qualities, texts of different complexity, and samples of different language types; Manually label the collected sample data to ensure the accuracy of the labeling and provide a reliable benchmark for model training and threshold setting; Developing algorithms to identify key features of input images, including image quality, text complexity, and language type; Conduct statistical analysis on the labeled sample data to determine the recognition accuracy and error rate under different feature combinations; According to the results of statistical analysis, the initial confidence threshold range is set for different feature combinations; Use the annotated and analyzed data to train the OCR model, including a dynamic adjustment mechanism for the confidence threshold; In the actual OCR recognition process, the confidence threshold is dynamically adjusted according to the characteristics of the input image; Evaluate the recognition results of the OCR model to determine whether they meet the set confidence threshold; Collect user feedback or automatically identify errors and analyze the characteristic conditions under which errors occur; Adjust the threshold setting based on error feedback and recognition results.

9. The method according to claim 1, characterized in that: The method further includes: matching and verifying the results of OCR recognition with entities, relationships and attributes in the knowledge graph, and verifying based on contextual semantic understanding; The knowledge graph matching and verification includes the following steps: Match and verify the OCR recognition results with the entities, relationships, and attributes in the knowledge graph; Natural language processing technology is used to analyze and identify the contextual semantics of the text. By building a language model, the correct vocabulary and sentence structure that are more likely to appear in a specific context are predicted, thereby improving the accuracy of verification.

10. A text recognition and automated review system, characterized in that: include: Acquisition module: used to acquire document data to be reviewed, wherein the document data to be reviewed includes semi-structured document data and unstructured text data; Identification module: used to obtain the corresponding OCR engine according to the content of the document data to be reviewed, and use the corresponding OCR engine to recognize the document data to be reviewed to obtain document data; Extraction module: used for selecting a document extraction type to extract target document data according to the document data, wherein the target document data format is structured text data, wherein the document extraction type includes rule extraction and model extraction; Comparison module: used to determine whether the target document data is compliant by comparing the target document data with pre-built target structured text data; Dynamic confidence threshold adjustment module: used to update the confidence threshold of the target document data in real time according to the complexity of the document data to be reviewed; Output module: used for outputting the compliance rate of the target document data according to whether the target document data is compliant and the confidence threshold of the target document data.

Images