Document auditing method and device, equipment and medium

Abstract

The invention relates to the technical field of document processing, and discloses a document auditing method, device and equipment and a medium, the method comprises the following steps: obtaining a to-be-audited document containing at least one of a text, an image and a chart, and analyzing the to-be-audited document to obtain corresponding to-be-audited content; obtaining an auditing rule document, extracting an auditing rule described in a natural language from the auditing rule document, and converting the auditing rule into a structured query statement; and extracting and auditing multi-dimensional elements of the to-be-audited content based on the structured query statement to obtain a corresponding auditing result, and outputting the auditing result. According to the technical scheme, the multi-modal document is automatically analyzed, the natural language rule is converted into the structured query statement, full-automatic analysis and auditing of the document are achieved on the basis of a multi-dimensional element extraction and auditing mechanism, manual work is not needed, and the document auditing efficiency and accuracy can be effectively improved.

Description

Technical Field

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

[0002] Technical documents, compliance reports, and other documents in various industries often require review operations. Currently, the usual way to review documents is manual review. Reviewers need to conduct corresponding reviews on massive amounts of documents according to requirements. However, this manual review method is not only time-consuming and labor-intensive, but also prone to errors and highly dependent on the reviewer's professional knowledge and experience. Summary of the Invention

[0003] In view of this, the present invention provides a document review method, apparatus, device and medium, which do not rely on manual labor and can effectively improve the efficiency and accuracy of document review.

[0004] In a first aspect, the present invention provides a document review method, comprising:

[0005] Obtaining a document to be reviewed that includes at least one of text, images, and charts, and parsing the document to be reviewed to obtain corresponding content to be reviewed;

[0006] Obtaining an audit rule document, extracting audit rules described in natural language from the audit rule document, and converting the audit rules into a structured query statement;

[0007] Based on the structured query statement, multi-dimensional elements of the content to be reviewed are extracted and reviewed, and corresponding review results are obtained and output.

[0008] In an optional implementation, parsing the document to be reviewed to obtain corresponding content to be reviewed includes:

[0009] Dividing the paragraphs or chapters contained in the document to be reviewed, identifying the logical structure between the title level and the main text, generating a document tree based on the division results and the identified structure, and semantically annotating the text in the document tree;

[0010] If the document to be reviewed also contains images and / or charts, text recognition is performed on the images and / or charts, and a semantic mapping is established between the recognized text and the text already existing in the document tree to obtain a multimodal joint semantic representation, and the document tree is updated based on the recognized text and the multimodal joint semantic representation to obtain the content to be reviewed.

[0011] In an optional embodiment, performing text recognition on the image and / or chart includes:

[0012] Performing OCR on the printed text in the image, locating the positional association of the recognized text in the image through layout analysis, and combining the recognized text into corresponding text based on the positional association; and / or,

[0013] The labels and data in the chart are extracted by computer vision technology, and a structured data matrix is ​​generated based on the extracted labels and data as the recognized text.

[0014] In an optional implementation, extracting and reviewing multi-dimensional elements of the content to be reviewed based on the structured query statement includes:

[0015] Extracting corresponding audit item information from the content to be audited based on the structured query statement, and retrieving corresponding audit basis information from a preset knowledge base;

[0016] The audit item information is audited for compliance factors based on the retrieved audit basis information.

[0017] In an optional implementation, the process of presetting the knowledge base includes:

[0018] Obtain an audit basis document, perform vectorization processing on each audit basis clause in the audit basis document to obtain multiple audit basis information, store the multiple audit basis information obtained in a designated knowledge base, and when the audit basis document is updated, synchronously update the audit basis information in the knowledge base through RAG technology.

[0019] In an optional implementation, after extracting corresponding audit item information from the content to be audited based on the structured query statement, the method further includes:

[0020] The consistency factor audit is performed by comparing the audit item information extracted from different locations.

[0021] In an optional implementation, before extracting and reviewing the multi-dimensional elements of the content to be reviewed based on the structured query statement, the method further includes:

[0022] Extracting audit tasks from the audit rule document and decomposing the audit tasks into atomic steps using the large model COT technology;

[0023] From the predefined multi-dimensional audit instruction templates, determine the audit instruction template corresponding to each of the atomic-level steps as the target instruction template, and use the target instruction model to execute the steps of extracting and auditing the multi-dimensional elements of the content to be audited based on the structured query statement.

[0024] The document review method provided by the present invention has the following advantages:

[0025] The technical solution of the present invention automatically parses multimodal documents, converts natural language rules into structured query statements, and based on a multi-dimensional factor extraction and review mechanism, realizes fully automatic parsing and review of documents without relying on manual labor, which can effectively improve the efficiency and accuracy of document review.

[0026] In a second aspect, the present invention further provides a document review device, comprising:

[0027] A parsing module is used to obtain a document to be reviewed that contains at least one of text, images, and charts, and parse the document to be reviewed to obtain corresponding content to be reviewed;

[0028] An extraction module is used to: obtain an audit rule document, extract the audit rules described in natural language from the audit rule document, and convert the audit rules into a structured query statement;

[0029] The audit module is used to extract and audit the multi-dimensional elements of the content to be audited based on the structured query statement, obtain the corresponding audit results and output them.

[0030] The document review device provided by the present invention has the following advantages:

[0031] The technical solution of the present invention automatically parses multimodal documents, converts natural language rules into structured query statements, and based on a multi-dimensional factor extraction and review mechanism, realizes fully automatic parsing and review of documents without relying on manual labor, which can effectively improve the efficiency and accuracy of document review.

[0032] In a third aspect, the present invention provides a computer device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, computer instructions being stored in the memory, and the processor executing the document review method of the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0033] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the document review method of the first aspect or any corresponding embodiment thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 This is a flowchart of a document review method provided by an embodiment of the present invention;

[0036] Figure 2 This is a structural block diagram of a document review system provided by an embodiment of the present invention;

[0037] Figure 3 Schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0039] In the prior art, the document review field has long faced challenges such as low efficiency of manual review. In order to solve this problem, the present invention provides a document review method, such as Figure 1 Specifically, it may include:

[0040] S11: Acquire a document to be reviewed that contains at least one of text, image, and chart, and parse the document to be reviewed to obtain corresponding content to be reviewed.

[0041] Among them, text is information presented in written language or symbolic form, which conveys content through words, numbers, punctuation, etc.; images convey information through visual elements (such as photos, paintings, and graphics), relying on color, shape, composition, etc.; charts are structured visual tools used to organize data or relationships, often combining text and graphics. The embodiments of the present invention support multimodal format input, which can be a single document containing multiple types of information (at least two of text, images, and charts), or a single document containing only a single type of information (one of text, images, and charts), such as supporting text (such as DOC / PDF), images (such as OCR scans), spreadsheets (such as XLS / XLSX), etc., covering mixed text and image document scenarios.

[0042] Parsing of documents to be reviewed refers to the process of multimodal data integration of mixed text and image documents. By parsing the documents to be reviewed, the content contained therein can be extracted as the content to be reviewed.

[0043] S12: Acquire an audit rule document, extract the audit rules described in natural language from the audit rule document, and convert the audit rules into a structured query statement.

[0044] Audit rule documents describe audit points in natural language (e.g., "Check whether the technical indicators in the document comply with the GB / T2887-2011 standard") and support custom audit dimensions (consistency, compliance) and priority configuration (e.g., setting priorities for different audit dimensions or different audit basis information, and executing audits corresponding to different audit dimensions or different audit basis information in descending order of priority). Structured query statement conversion refers to the process of converting natural language rules into machine-executable instructions. This can be achieved using semantic parsing algorithms, which can overcome the high maintenance cost of hard-coded rules. For example:

[0045] Input: "Check if all device models are in the vendor list";

[0046] Output: {"Dimension":"Consistency","Object":"Device Model","Comparison Library":"Supplier List","Matching Method":"Exact Match"}.

[0047] S13: Extract and review the multi-dimensional elements of the content to be reviewed based on the structured query statement, obtain and output the corresponding review results.

[0048] Multi-dimensional element extraction and review refers to the process of simultaneously addressing audit requirements such as compliance and consistency. Specifically, the current audit requirements can be determined based on structured query data, and then multi-dimensional elements can be extracted and reviewed for the audit requirements according to the current audit requirements, thereby solving the problem of incomplete coverage of single-dimensional audits. Specifically, consistency elements are to extract data contradictions across chapters / documents (such as "the voltage range mentioned in the previous text is 0-220V, and the actual measured value in the later text is 250V"), and structured output of difference content and location information; compliance elements are to extract non-conformities based on the standard terms of the knowledge base (such as "the frequency of data backup is not mentioned, which violates Article 5.4.1 of GB / T XXX") and automatically calculate the compliance rate. In addition, the audit results can be generated by the audit dimension in the form of Key-Value pairs in JSON format to automatically summarize the extraction results.

[0049] In a specific implementation, when processing a construction project acceptance report, the document containing construction instructions, inspection data sheets, and on-site photos is first parsed to obtain the content contained therein as the content to be reviewed; the review rule "verify whether the inspection data meets the requirements of the GB50204 standard" is converted to generate a structured statement; according to the structured statement, the inspection data in the content to be reviewed is extracted and verified to see whether it meets the requirements of the GB50204 standard, and the corresponding results are obtained and output.

[0050] The technical solution of the present invention automatically parses multimodal documents, converts natural language rules into structured query statements, and based on a multi-dimensional factor extraction and review mechanism, realizes fully automatic parsing and review of documents without relying on manual labor, which can effectively improve the efficiency and accuracy of document review.

[0051] The present invention provides a document review method, which parses the document to be reviewed to obtain the corresponding content to be reviewed, and may include:

[0052] Dividing the paragraphs or chapters contained in the document to be reviewed, identifying the logical structure between the title level and the main text, generating a document tree based on the division results and the identified structure, and semantically annotating the text in the document tree;

[0053] If the document to be reviewed also contains images and / or charts, text recognition is performed on the images and / or charts, and a semantic mapping is established between the recognized text and the text already existing in the document tree to obtain a multimodal joint semantic representation, and the document tree is updated based on the recognized text and the multimodal joint semantic representation to obtain the content to be reviewed.

[0054] Among them, the document tree refers to the tree-like data structure formed by hierarchical division of document paragraphs through natural language processing technology, which is used to represent the semantic association relationship of document content. Semantic annotation refers to the entity type tagging of text content in the document tree. Specifically, named entity recognition technology can be used to perform entity annotation on key information such as technical parameters and standard names. It can include text entities (such as "standard name" and "technical parameters"), digital entities (such as "value" and "percentage"). Semantic tags can also be added to the text in the document tree, such as "Chapter 1-Technical Indicators-Parameter A" to further improve the completeness of the document tree information. Text recognition refers to the extraction of text information in images or charts. Specifically, it can be implemented by optical character recognition technology combined with layout analysis algorithms, etc., to convert unstructured image data into processable text information. Multimodal joint semantic representation refers to the association of image recognition text with document tree text (obtained based on the text in the document to be reviewed) through semantic mapping relationships. Specifically, it can be implemented by a cross-modal alignment model based on the attention mechanism to build a unified semantic framework for mixed text and image content.

[0055] In a specific implementation method, the document to be reviewed is first structurally parsed, and the paragraph separation marks (such as line breaks, indents, etc.) and chapter title features (such as numbering, font bolding, font size changes, etc.) in the document are identified through natural language processing algorithms or rule engines, thereby dividing the document into paragraphs or chapter units; then the subordinate relationships between the titles (such as the nested structure of first-level titles and second-level titles) and the logical associations between the titles and the corresponding text contents are analyzed to construct a logical structure reflecting the document organization framework; then, based on the above division and recognition results, a document tree is generated, in which the root node represents the document as a whole, the child nodes correspond to chapters or paragraphs according to the hierarchy, and the content ownership is reflected through the parent-child relationship; finally, semantic analysis technology is used to annotate the text nodes in the document tree with semantics and entities (such as text entities, digital entities, etc.), thereby forming a document representation model with both structured and semantic features, providing a data basis for subsequent intelligent review.

[0056] When the document to be reviewed contains images and / or charts, the text content in the image or chart (such as annotated text, data labels, legend descriptions, etc.) is first extracted using optical character recognition (OCR) technology, and its structural information (such as table rows and columns, flowchart nodes) is identified in combination with visual analysis (such as target detection and layout segmentation); then, semantic matching algorithms (such as word vector similarity calculation and knowledge graph association) are used to associate the identified text with existing nodes in the document tree (such as adjacent paragraphs and chapter titles), and construct a cross-modal semantic mapping relationship (for example, establishing a logical link between the "temperature change curve" in the chart and the "experimental results analysis" section in the text, and the "growth rate" in the text corresponds to the broken line trend in the chart), forming a multimodal joint representation that integrates the semantics of text, image, and chart. On this basis, the document tree is dynamically updated, including creating exclusive nodes for images / charts and embedding their identification text and metadata (such as location, type), adjusting the semantic annotations of adjacent nodes to reflect multimodal associations (such as adding "data support" tags to relevant paragraphs), and optimizing the hierarchical structure of the document tree (for example, adding supplementary explanatory charts as child nodes of the corresponding chapters). Ultimately, a content model to be reviewed is generated that contains a unified semantic representation of text, images, and charts, ensuring the logical consistency and traceability between non-text elements and the document body, and providing a structured basis for subsequent multimodal content review (such as text and image consistency verification) or interactive retrieval (such as locating relevant charts through text queries).

[0057] The embodiments of the present invention realize the automatic association processing of image and text data, solve the problem of low efficiency of manual parsing of cross-modal information in mixed graphic and text documents, can automatically extract structured data in charts and establish a semantic mapping relationship with text content, and provide a complete multimodal semantic foundation for subsequent review.

[0058] In addition, noise filtering can be performed on the document to be reviewed (either before or after obtaining the document tree, both within the scope of protection of the present invention), including automatic removal of watermarks, headers and footers, repeated paragraphs and other irrelevant information, to increase the proportion of valid data. Specifically, the watermark area in the document is first detected using computer vision technology, and a deep learning-based image restoration algorithm (such as a generative adversarial network) is used to intelligently fill in and reconstruct the texture of the watermark pixels while maintaining the integrity of the background content. For header and footer information, the layout analysis algorithm is combined to identify the fixed format area of ​​the document, and metadata such as page numbers and dates at specific locations are located through OCR. Header and footer content that is not related to the main text is filtered out based on semantic analysis. To address the problem of repeated paragraphs, text fingerprint technology (such as SimHash or MinHash) is used to calculate the feature hash value of each paragraph, and repeated or highly similar text blocks are identified through similarity comparison. The most appropriate version is retained based on contextual coherence analysis. In this process, a noise feature library is dynamically constructed and common noise patterns (such as copyright statements in specific formats, standardized disclaimers, etc.) are continuously updated to improve the filtering accuracy of subsequent document processing. The final output of the purified document will increase the proportion of valid data by more than 30%, providing high-quality input for subsequent semantic analysis and content review. At the same time, the process is fully automated and the integrity of the original semantics of the document can be maintained without human intervention.

[0059] The present invention provides a document review method, which performs text recognition on the image and / or chart, and may include:

[0060] Performing OCR on the printed text in the image, locating the positional association of the recognized text in the image through layout analysis, and combining the recognized text into corresponding text based on the positional association; and / or,

[0061] The labels and data in the chart are extracted by computer vision technology, and a structured data matrix is ​​generated based on the extracted labels and data as the recognized text.

[0062] Among them, OCR recognition refers to the technology of converting printed text in an image into editable text, which can be implemented using open source tools such as Tesseract or PaddleOCR. Layout analysis refers to the analysis of the spatial distribution relationship of text in an image, which can be implemented using layout segmentation algorithms based on deep learning, etc., and a logical sequence relationship is established by identifying the coordinate information of text blocks. Computer vision technology refers to the extraction of feature information of visual elements from charts, which can be implemented using contour detection and template matching methods in the OpenCV library to locate the positions of coordinate axis labels and data points. Structured data matrix refers to the storage of numerical relationships in charts in the form of a two-dimensional table, which can be implemented using JSON or XML format for data encapsulation to facilitate logical verification in subsequent review processes.

[0063] Specifically, for the image part of the document, optical character recognition (OCR) is first used to perform high-precision recognition of the printed text in the image to obtain the original character information; at the same time, the text area in the image is detected and located through a deep learning-based layout analysis algorithm (such as CNN or Transformer architecture, etc.), and the spatial position relationship of each character block is identified (including line spacing, character alignment, relative position of text blocks, etc.), and based on these spatial position associations (such as horizontally adjacent characters forming words, vertically aligned text blocks forming paragraphs), the discrete recognized characters are intelligently combined into complete text paragraphs with semantic coherence. For the charts in the document, computer vision techniques (such as target detection and semantic segmentation) are used to automatically extract various label elements in the chart (including axis labels, legend descriptions, data point annotations, etc.) and their corresponding numerical data. By analyzing the visual associations of chart elements (such as the correspondence between labels and data lines, and the topological structure of table rows and columns), this information is reconstructed into a structured data matrix (such as converting a line chart into a two-dimensional array of [time point, value], or converting a pie chart into a mapping table of [category, percentage]), and finally forming a machine-readable standardized text representation; these two types of processing processes can work independently or collaboratively to jointly realize the conversion of non-text visual content into standardized text data that can be processed by subsequent semantic analysis modules, providing a unified semantic input basis for multimodal document understanding.

[0064] The embodiments of the present application realize the automatic parsing of image and chart data in documents, eliminate the time cost of manually parsing chart information, and avoid numerical transcription errors caused by visual fatigue.

[0065] The present invention provides a document review method, which extracts and reviews multi-dimensional elements of the content to be reviewed based on the structured query statement, and may include:

[0066] Extracting corresponding audit item information from the content to be audited based on the structured query statement, and retrieving corresponding audit basis information from a preset knowledge base;

[0067] The audit item information is audited for compliance factors based on the retrieved audit basis information.

[0068] Among them, structured query statements refer to the conversion of audit rules described in natural language into executable query instructions, which can be implemented specifically using SQL statements or graph query languages, and are used to accurately locate specific data items that need to be verified from the content to be audited. A knowledge base refers to a database that stores structured audit basis information, which can be generated by vectorizing national standards, industry specifications, policy approvals and other documents. Each rule clause is converted into vector data to support semantic retrieval, and is dynamically updated through RAG technology to ensure the timeliness of the audit basis. Compliance element review refers to matching and verifying the extracted audit item information with the standard clauses in the knowledge base, which can be implemented specifically through semantic similarity calculation or logical reasoning, and is used to identify non-conformities and generate compliance conclusions.

[0069] Specifically, audit item information is extracted from the content to be audited through structured query statements, such as the technical parameters, process descriptions or data indicators involved in the document. The audit basis information in the knowledge base is quickly retrieved through vectorized storage, such as converting regulatory clauses into vectors and then semantically matching them with the audit items. When the audit basis document is updated, the knowledge base automatically synchronizes the latest clauses through RAG technology. For example, new industry standards will be incorporated into the knowledge base in real time. During the compliance factor review process, the system determines whether the audit items meet the requirements based on the clauses in the knowledge base, such as whether the detection technical solution meets the constraints of specific regulations, and outputs the specific reasons for non-compliance.

[0070] The embodiments of the present application solve the problem of compliance audit delay caused by the inefficiency of manual retrieval of evidence in traditional audit methods. At the same time, it ensures the authority of the audit basis through a dynamic update mechanism, significantly improving the accuracy and timeliness of compliance audits of complex documents.

[0071] The present invention provides a document review method, and the process of presetting the knowledge base may include:

[0072] Obtain an audit basis document, perform vectorization processing on each audit basis clause in the audit basis document to obtain multiple audit basis information, store the multiple audit basis information obtained in a designated knowledge base, and when the audit basis document is updated, synchronously update the audit basis information in the knowledge base through RAG technology.

[0073] Vectorization processing refers to converting the audit basis clauses described in natural language into numerical vectors that can be recognized by computers. This can be achieved by encoding the text using a pre-trained language model, such as generating semantic vectors through BERT or similar models. This processing enables the semantic information of the text clauses to be stored in a structured manner, facilitating subsequent retrieval through vector similarity matching. RAG technology refers to retrieval enhancement generation technology, which can be achieved by combining document change detection with incremental vector updates. When the content of the audit basis document is increased, decreased, or modified, the changed part is identified through text difference comparison, and only the changed part is re-vectorized, and the corresponding vector data in the knowledge base is updated synchronously. This processing can achieve partial updates of the knowledge base and avoid the computing resource consumption caused by full reconstruction.

[0074] In one specific implementation, when building a knowledge base, the acquired audit basis document is first split into independent clauses. Each clause is vectorized to form a structured semantic vector, which is then stored in the knowledge base as audit basis information. When the audit basis document version is updated or other changes occur, the change detection mechanism in RAG technology automatically identifies new, deleted, or modified clauses, and only performs vectorized updates on the affected parts. For example, if the numerical threshold in a clause is changed from "not higher than 10%" to "not higher than 15%," the clause is located and a new semantic vector is generated to overwrite the original data. This allows the audit basis information in the knowledge base to be synchronized with the latest document in real time, eliminating the need for manual intervention in rule base maintenance.

[0075] In the embodiment of the present application, when regulations or standards are revised, the knowledge base can automatically update the terms to avoid delays caused by manual maintenance; the local vector update mechanism reduces computational overhead, allowing the system to quickly adapt to high-frequency document change scenarios.

[0076] The present invention provides a document review method, which, after extracting corresponding review item information from the content to be reviewed based on the structured query statement, may further include:

[0077] The consistency factor audit is performed by comparing the audit item information extracted from different locations.

[0078] Consistency factor review refers to the logical verification of similar review items scattered in different paragraphs, chapters or cross-modal content of a document. Specifically, it can be implemented by combining semantic similarity calculation with a rule reasoning engine, etc., to detect numerical contradictions or logical conflicts of the same review item in different contexts.

[0079] Specifically, the consistency factor review classifies the extracted review item information by type and cross-validates the data of the same review item in different locations. For example, in a technical document, if the voltage range is defined as 0-220V in the previous text and the actual measured value is 250V in the later text, the contradictions are identified through numerical comparison, and the chapter position where the difference content is located is located based on the semantic annotations in the document tree.

[0080] The embodiments of the present application can automatically detect logical contradictions in data across chapters and modalities within a document, such as identifying conflicts between parameter definitions and measured values ​​in a technical solution, or conflicts between clause references and conclusions in a compliance report, and output the difference content and its location information in a structured manner, thereby reducing the workload of manual comparison and improving the completeness and accuracy of the audit dimension.

[0081] The present invention provides a document review method, which may further include:

[0082] Extracting audit tasks from the audit rule document and decomposing the audit tasks into atomic steps using the large model COT technology;

[0083] From the predefined multi-dimensional audit instruction templates, determine the audit instruction template corresponding to each of the atomic-level steps as the target instruction template, and use the target instruction model to execute the steps of extracting and auditing the multi-dimensional elements of the content to be audited based on the structured query statement.

[0084] Among them, large-model COT technology refers to a task decomposition method based on thought chain reasoning. Specifically, a pre-trained language model can be used to perform multi-step reasoning on the audit task to generate an atomic step sequence. The atomic step refers to the smallest audit operation unit that cannot be further divided. For example, "compliance audit" can be broken down into a four-step process of "standard clause retrieval → indicator extraction → logical comparison → conclusion generation". Multi-dimensional audit instruction templates (Prompt project templates) refer to pre-defined standardized audit operation instructions. A Prompt project template library can be pre-set, in which multi-dimensional audit instruction templates are pre-defined, such as: consistency check: "Analyze whether there is any contradiction between parameter A of Chapter X and Chapter Y in the document, and list the differences"; compliance check: "Based on the GB / T XXX standard in the knowledge base, determine whether the technical solution in the document complies with Article X, Clause Y, and explain the reasons."

[0085] Specifically, a template library that combines natural language descriptions with structured query statements can be used, which can decouple audit rules from execution logic.

[0086] Specifically, the audit tasks that need to be performed are first parsed from the audit rule document, and the Chain-of-Thought (COT) reasoning technology of the large model is used to logically decompose these tasks and break them down into indivisible atomic-level audit steps. Then, the semantic matching algorithm is used to retrieve the audit instruction template that matches each atomic step from the pre-built multi-dimensional audit instruction template library. These templates not only contain specific audit dimension definitions (such as compliance and consistency), but also preset structured query logic and judgment criteria under each dimension. After the audit instruction template is matched, the content to be audited is dynamically bound to the query conditions in the target instruction template to generate an executable composite audit instruction stream. Multi-level audit operations including text semantic verification, data logic check, and cross-modal correlation audit are automatically executed in parallel or serial manner. Finally, a multi-dimensional audit report with problem location and modification suggestions is output, realizing intelligent processing of the entire process from original rules to specific execution.

[0087] This application implements automated process orchestration for complex audit tasks, addressing the high cost of rule maintenance and insufficient cross-modal data processing capabilities of traditional methods when handling multi-dimensional audit requirements. This is specifically reflected in the improved logical coverage integrity brought about by the improved accuracy of audit task decomposition, and the improved rule update efficiency brought about by the standardization of instruction templates, enabling the system to automatically handle complex audit scenarios involving mixed text and image data.

[0088] The document review method provided by the embodiment of the present invention has the following characteristics:

[0089] Improve audit efficiency and reduce manual reliance: The large model automatically extracts information on audit elements such as consistency and compliance, reducing reliance on professional experience and improving audit efficiency.

[0090] Enhanced multimodal document parsing: Automatically processes images, text, and table data based on OCR models and associates semantics, solving the problem of inefficient manual parsing.

[0091] Dynamically adapt rules to shorten response cycles: Rapidly update audit rules based on large model prompt engineering and RAG technology.

[0092] The embodiment of the present invention also provides a document review device, such as Figure 2 Specifically, it may include:

[0093] The parsing module 11 is used to obtain a document to be reviewed that contains at least one of text, images, and charts, and parse the document to be reviewed to obtain corresponding content to be reviewed;

[0094] The extraction module 12 is used to: obtain an audit rule document, extract the audit rules described in natural language from the audit rule document, and convert the audit rules into a structured query statement;

[0095] The review module 13 is used to extract and review multi-dimensional elements of the content to be reviewed based on the structured query statement, obtain corresponding review results and output them.

[0096] The device fault handling apparatus in this embodiment is presented in the form of a functional module, where the module refers to an application-specific integrated circuit (ASIC), a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0097] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0098] The embodiment of the present invention also provides a computer device having the above Figure 2 The document review system shown.

[0099] See also Figure 3 , Figure 3 is a structural diagram of a computer device provided by an optional embodiment of the present invention, such as Figure 3 As shown, the computer device includes: one or more processors 10, memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components utilize different buses to communicate with each other and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in the memory or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Equally, multiple computer devices can be connected, and each device provides part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 3 A processor 10 is taken as an example.

[0100] The processor 10 may be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0101] The memory 20 stores instructions that can be executed by at least one processor 10, so as to enable at least one processor 10 to execute the method shown in the above embodiment.

[0102] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created based on the use of a computer device for displaying a small program landing page, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0103] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0104] The computer device further includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Figure 3 The bus connection is taken as an example.

[0105] The input device 30 can receive input digital or character information and generate key signal input related to user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a trackpad, a touch pad, an indicator stick, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 can include a display device, an auxiliary lighting device (e.g., an LED), and a tactile feedback device (e.g., a vibration motor). The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display, and a plasma display. In some optional embodiments, the display device can be a touch screen.

[0106] The embodiment of the present invention also provides a computer-readable storage medium. The above-mentioned method according to the embodiment of the present invention can be implemented in hardware, firmware, or implemented as a computer code that can be recorded in a storage medium, or implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and downloaded through a network and will be stored in a local storage medium, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state drive, etc.; further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the method shown in the above embodiment is implemented.

[0107] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A document review method, characterized in that: include: Obtaining a document to be reviewed that includes at least one of text, images, and charts, and parsing the document to be reviewed to obtain corresponding content to be reviewed; Obtaining an audit rule document, extracting audit rules described in natural language from the audit rule document, and converting the audit rules into a structured query statement; Based on the structured query statement, multi-dimensional elements of the content to be reviewed are extracted and reviewed, and corresponding review results are obtained and output.

2. The method according to claim 1, characterized in that The document to be reviewed is parsed to obtain the corresponding content to be reviewed, including: Dividing the paragraphs or chapters contained in the document to be reviewed, identifying the logical structure between the title level and the main text, generating a document tree based on the division results and the identified structure, and semantically annotating the text in the document tree; If the document to be reviewed also contains images and / or charts, text recognition is performed on the images and / or charts, and a semantic mapping is established between the recognized text and the text already existing in the document tree to obtain a multimodal joint semantic representation, and the document tree is updated based on the recognized text and the multimodal joint semantic representation to obtain the content to be reviewed.

3. The method according to claim 2, characterized in that Performing text recognition on the image and / or chart, including: Performing OCR on the printed text in the image, locating the positional association of the recognized text in the image through layout analysis, and combining the recognized text into corresponding text based on the positional association; and / or, The labels and data in the chart are extracted by computer vision technology, and a structured data matrix is ​​generated based on the extracted labels and data as the recognized text.

4. The method according to claim 1, wherein Extracting and reviewing multi-dimensional elements of the content to be reviewed based on the structured query statement includes: Extracting corresponding audit item information from the content to be audited based on the structured query statement, and retrieving corresponding audit basis information from a preset knowledge base; The audit item information is audited for compliance factors based on the retrieved audit basis information.

5. The method according to claim 4, characterized in that The process of presetting the knowledge base includes: Obtain an audit basis document, perform vectorization processing on each audit basis clause in the audit basis document to obtain multiple audit basis information, store the multiple audit basis information obtained in a designated knowledge base, and when the audit basis document is updated, synchronously update the audit basis information in the knowledge base through RAG technology.

6. The method according to claim 4, characterized in that After extracting corresponding audit item information from the content to be audited based on the structured query statement, the method further includes: The consistency factor audit is performed by comparing the audit item information extracted from different locations.

7. The method according to any one of claims 1 to 6, characterized in that Before extracting and reviewing the multi-dimensional elements of the content to be reviewed based on the structured query statement, the method further includes: Extracting audit tasks from the audit rule document and decomposing the audit tasks into atomic steps using the large model COT technology; From the predefined multi-dimensional audit instruction templates, determine the audit instruction template corresponding to each of the atomic-level steps as the target instruction template, and use the target instruction model to execute the steps of extracting and auditing the multi-dimensional elements of the content to be audited based on the structured query statement.

8. A document review device, characterized in that: include: A parsing module is used to obtain a document to be reviewed that contains at least one of text, images, and charts, and parse the document to be reviewed to obtain corresponding content to be reviewed; An extraction module is used to: obtain an audit rule document, extract the audit rules described in natural language from the audit rule document, and convert the audit rules into a structured query statement; The audit module is used to extract and audit the multi-dimensional elements of the content to be audited based on the structured query statement, obtain the corresponding audit results and output them.

9. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the document review method according to any one of claims 1 to 7 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the document review method according to any one of claims 1 to 7.

Images