Intelligent document review method and system based on large language model

Abstract

The invention discloses an intelligent document review method and system based on a large language model, and the method comprises the steps: inputting a document, carrying out the analysis of the document, and carrying out the slicing of the analyzed document; converting the slice text into vectors, and constructing a clause dependency graph; inputting map risk characteristics, and performing risk judgment based on a power grid and a coal mine industry system; generating a first edition report based on a risk judgment result, checking whether the generated content is compliant and standard through a filter, and forcibly rewriting if the generated content is not compliant with the rule; and based on a coordinate reverse mapping technology, mapping the risk points back to the coordinates of the document, and generating an annotation layer. According to the method, the forgetting problem caused by the context length limitation of the Transform model is effectively solved, and cross-section logic conflicts which cannot be found by traditional linear scanning are effectively solved; through a coordinate reverse mapping technology, pixel-level highlighting of the risk point on the original file is realized, the maintenance workload is greatly reduced, and the error possibility is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent document review technology, and in particular to an intelligent document review method and system based on a large language model. Background Technology

[0002] In existing technologies, major enterprises primarily rely on the following two technical approaches when processing various unstructured and semi-structured documents (such as business agreements, technical specifications, bidding documents, product manuals, supply chain materials, internal process documents, etc.):

[0003] 1. Manual review method: Business personnel or project managers manually check the document's logical consistency, the completeness of key clauses, and compliance by referring to the internal knowledge base (OA system, Wiki), industry standards, and historical experience.

[0004] II. Keyword and Regular Expression Matching: Relies on a predefined keyword library (such as "breach of contract liability" and "overcurrent protection") and complex regular expression chains for matching.

[0005] However, traditional regular expressions or keyword matching cannot understand complex legal semantics. Furthermore, with updates to laws and regulations or changes in internal corporate systems (such as revisions to the Production Safety Law or industry red line standards), a large number of regular expressions and logic codes need to be rewritten manually, resulting in a huge workload for maintenance and a high risk of errors. Summary of the Invention

[0006] This invention provides an intelligent document review method and system based on a large language model to solve the technical problems of high false alarm rate and difficulty in quickly finding the source of the fault in the prior art.

[0007] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent document review method based on a large language model, comprising:

[0008] S10, Input document, parse the document, and slice the parsed document;

[0009] S20, convert the sliced ​​text into vectors and construct the term dependency graph;

[0010] S30, Input risk characteristics of the graph and make risk assessment based on the regulations of the power grid and coal mining industries;

[0011] S40: Based on the risk assessment results, an initial report is generated. The generated content is checked for compliance and standards through filters. If it does not comply with the rules, it is forced to be rewritten.

[0012] S50, based on coordinate reverse mapping technology, maps risk points back to the coordinates of the document to generate an annotation layer.

[0013] Optionally, step S10 includes:

[0014] S101, Input PDF / WORD / OFD documents and remove watermark noise from the documents;

[0015] S102, perform OCR recognition on the document after removing watermark noise, and perform sliding window slicing to divide the data stream into semantic retrieval stream and graph construction stream.

[0016] Optionally, step S20 includes:

[0017] S201, slice the text T i Convert to an embedding vector and query the Tok-k regulation L in the vector database. k and system R k ;

[0018] S202, based on graph construction flow, employs a dual-driven model of model and rules to extract entities and relationships, and constructs a term dependency graph. .

[0019] Optionally, step S30 includes:

[0020] S301, Input Risk characteristics of the map;

[0021] S302, Target Clauses Based on Regulatory Analysis of Power Grid and Coal Mining Industries;

[0022] S303, based on the analysis results, generates suggestions through a step-by-step approach of risk positioning, legal source tracing, and text reshaping.

[0023] Secondly, the present invention provides an intelligent document review system based on a large language model, comprising:

[0024] The input layer is used to distinguish and extract content related to the main text, tables, headers and footers, and seals.

[0025] The data processing layer is used to process the extracted relevant content;

[0026] The knowledge and hint layer is used to receive search results, connect to the vector database, and query external regulations, internal systems, and historical negative samples.

[0027] The core inference layer is used to deploy large Decoder-only models and to mount industry-specific fine-tuning weights.

[0028] The output and interaction layer is used to generate visual annotations at the corresponding coordinate positions in the original document. It is also used to output an interpretable review report containing risk classification, legal basis and customized rewriting suggestions based on coordinate reverse mapping technology, and provides a human proofreading feedback interface to realize dynamic correction and incremental learning of the knowledge base.

[0029] Optionally, the data processing layer includes:

[0030] The segmentation-pagination engine is used to divide the parsed unstructured text into slices that maintain semantic integrity. It is also used to record the page index, spatial coordinates and hierarchical topology of each slice in the original document based on the document's physical page break and logical paragraph identifier.

[0031] The vectorized embedding module is used to convert non-numerical data into low-dimensional dense vectors that can be understood by computers, while preserving the semantic information or feature relationships of the data.

[0032] Optionally, the knowledge and hint layer includes:

[0033] A dynamic suggestion builder is used to accept search results and connect to a vector database;

[0034] The three-dimensional knowledge base includes an external legal database, an internal institutional database, and a historical negative sample database. The external legal database is used to query external laws and regulations, the internal institutional database is used to query internal institutional regulations, and the historical negative sample database is used to query historical negative samples.

[0035] Optionally, the core inference layer includes:

[0036] The base model is used to deploy Decoder-only large models, serving as the foundation for semantic understanding and text generation.

[0037] The adapter module is used to attach lightweight weights for fine-tuning the corpus of the power grid and coal mining industries, so as to achieve standard alignment of the model with industry professional terms and unique systems.

[0038] The risk graph generator is used to calculate the conflict detection and risk propagation between clauses. Based on the risk graph algorithm, it identifies deep risks by recognizing logical loops and parameter contradictions in the clause dependency graph.

[0039] Optionally, the output and interaction layer includes:

[0040] An interpretable report generator is used to assemble risk assessment results, legal basis, and modification suggestions into a structured report, and call coordinate reverse mapping data to generate interactive annotations on the original document;

[0041] The feedback loop interface is used to collect human confirmation, ignoring, or correction behaviors of the review results, and transform them into fine-tuning samples to be sent back to the core inference layer in order to achieve dynamic optimization of the system review criteria.

[0042] One or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages:

[0043] This application effectively solves the "forgotten" problem caused by the context length limitation of the Transformer model, as well as cross-chapter logical conflicts (such as logical deadlocks in payment and delivery cycles) that cannot be detected by traditional linear scanning, by combining sliding window slicing technology and risk graphing algorithm; through coordinate reverse mapping technology, it realizes pixel-level highlighting of risk points on the original file, which greatly reduces the maintenance workload and the possibility of errors. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the intelligent document review method based on a large language model provided in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the structure of the intelligent document review system based on a large language model provided in an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] In the relevant descriptions of this embodiment, the terms "including," "containing," and "possessing" are all open terms and are generally understood to include but not be limited to; the term "at least one" is generally understood to mean one or more, where "multiple" refers to two or more; the term "at least one of the following" or similar expressions refer to any combination of these items, including any combination of single or multiple items, for example, "at least one of a, b, or c", or "at least one of a, b, and c", which can all mean: a, b, c, ab (i.e., a and b), ac, bc, or abc, where a, b, and c can be single or multiple; the symbol "A / B" is used to describe the selection relationship of associated objects, generally indicating an "or" relationship.

[0048] In the following description of the embodiments, the terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms "a" and "the" as used in the embodiments of this application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0049] Those skilled in the art should understand that, in the following description of the embodiments of this application, the sequence of numbers does not imply the order of execution. Some or all steps may be executed in parallel or sequentially. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0050] Those skilled in the art will understand that the numerical ranges in the embodiments of this application should be understood to specifically disclose each intermediate value between the upper and lower limits of the range. Any stated value or intermediate value within a stated range, as well as any other stated value or each smaller range between intermediate values ​​within a range, are also included within this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0051] Unless otherwise stated, the technical / scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. While this application describes only preferred methods and materials, any methods and materials similar or equivalent to those described herein may be used in the implementation or testing of this application. All references to this specification are incorporated by way of citation to disclose and describe the methods and / or materials associated with those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

[0052] This invention provides an intelligent document review method based on a large language model, which can be executed by a computer. (See reference...) Figure 1 The method may include the following steps:

[0053] S10: Input document, parse the document, and slice the parsed document.

[0054] In an exemplary embodiment, step S10 includes:

[0055] S101, Input PDF / WORD / OFD documents and remove watermark noise from the documents;

[0056] S102, perform OCR recognition on the document after removing watermark noise, and perform sliding window slicing to divide the data stream into semantic retrieval stream and graph construction stream.

[0057] OCR (Optical Character Recognition) is a technology that scans images or documents and uses computer vision algorithms to detect and recognize text content, symbols, and layout, converting them into editable and searchable text formats. It integrates image preprocessing, character segmentation, pattern recognition, and natural language processing technologies. It can not only "understand" text in images but also reconstruct the structure of complex layouts such as drawings and production reports. It is widely used in scenarios such as document digitization, license plate recognition, and invoice processing, and supports multilingual and special character recognition to improve information processing efficiency.

[0058] In document parsing and slicing scenarios, "data flow" refers to the logical path of document data from its original input state through multiple levels of processing to its final output state. It defines how data is parsed, preprocessed, identified, organized, and distributed. The core function of data flow is to establish a "streaming awareness mechanism" that ensures that after complex preprocessing, the semantic features and spatial coordinate information of the document can be accurately extracted and orderly flowed to subsequent parallel computing units, providing underlying data support for cross-chapter logical verification.

[0059] In this embodiment, the data stream originates from the input PDF / WORD / OFD document. First, it undergoes image enhancement by "removing watermark noise" to eliminate interference from complex backgrounds for recognition. Then, "OCR recognition" converts pixel information into structured text containing coordinate metadata. Its core function is to parse a single raw file stream into structured stream data with spatial and semantic attributes, and efficiently allocate it to the semantic retrieval stream (for matching industry regulations) and the graph construction stream (for extracting logical entities from clauses). This ensures that when processing large industry regulations, the data flow is efficient, logically rigorous, and without information loss or context breakage.

[0060] S20 converts the sliced ​​text into vectors and constructs a term dependency graph.

[0061] In an exemplary embodiment, step S20 includes:

[0062] S201, slice the text T i Convert to an embedding vector and query the Tok-k regulation L in the vector database. k and system R k ;

[0063] S202, based on graph construction flow, employs a dual-driven model of model and rules to extract entities and relationships, and constructs a term dependency graph. .

[0064] Specifically, cosine similarity is used as the metric, and the formula is as follows:

[0065]

[0066] in, For the Embedding vector of the clauses to be reviewed, For the first in the regulations / institutions library A vector of data points. The system selects... The first k segments of the threshold are used as contextual basis.

[0067] The following section provides a detailed explanation of how to extract entities and relations:

[0068] Entity Extraction: Using Named Entity Recognition (NER) technology, the following are identified from the clauses: Subjects, such as "Mine Manager," "Safety Inspector," and "Power Grid Dispatcher"; Objects, such as "Coal Mining Machine," "110KV Busbar," and "Gas Monitor"; Parameter Values, such as "Voltage Level ≥ 110KV" and "Gas Concentration." ".

[0069] Relation Extraction: Utilizing dependency parsing and relation extraction (ER) techniques, logical conjunctions between entities are extracted: temporal relations, "first...then..." (e.g., cabinet doors can only be opened after a power outage for maintenance); causal / conditional relations, "if...then..." (e.g., power must be cut off immediately if gas concentration exceeds the limit); and inclusion / constraint relations, "belongs to... and should follow..." (e.g., this contract falls under the category of special operations and must comply with coal mine special operations regulations). Graph logic is then used, treating each independent clause as a node. The extracted logical connectors are used as directed edges. This forms a topological network with legally valid logic.

[0070] In this embodiment, the sliced ​​text T i This refers to the fragments obtained by splitting the original text according to preset rules. i This serves as the identifier for the segment, facilitating subsequent processing.

[0071] Embedding is a technique that transforms high-dimensional data such as text and images into low-dimensional dense vector representations, capturing deep semantic information through deep learning models.

[0072] Vector databases are systems specifically designed for storing, managing, and efficiently retrieving vector data. Their core principle is to utilize the Approximate Nearest Neighbor (ANN) algorithm to quickly locate semantically similar content within massive datasets. For sliced ​​text T... i In this context, it serves as the smallest semantic unit after logical decomposition of the original long document, through T i Identifiers enable association with the original physical coordinates, providing a processable feature carrier for machine computation.

[0073] To address cross-clause risks, a risk graph algorithm is used to construct a clause dependency graph. .

[0074] in, (Node) represents an independent clause in the contract (such as payment terms, default clauses, etc.). (edge) represents the dependency relationship or semantic association between clauses.

[0075] Risk graph algorithms are the core technology used to construct and analyze risk graphs. They utilize graph structures (nodes V and edges) This can be used to quantify the likelihood and impact of risks, helping to identify key risk points and potential threats.

[0076] Furthermore, the existence of a "deadlock cycle" in the clause dependency graph is calculated using cycle detection and risk propagation formulas, and the cumulative risk score is calculated. The specific formula is as follows:

[0077]

[0078] in, The set of all nodes in the dependency graph, each node It represents a separate clause in a contract or procedure (such as payment terms or maintenance safety terms). This represents the risk value identified by the Large Language Model (LLM) for a single clause; This is a preset risk weighting coefficient for each clause, used to adjust the proportion of a single violation in the overall score. , The set of edges connecting the clauses in the dependency graph represents the dependency, sequence, or constraint relationships between the clauses. Indicates the probability of conflict between terms (such as a logical conflict where "payment time" is earlier than "delivery and acceptance time"). This is the weighting coefficient for the risk of logical conflict; This is an indicator function; if a logical loop closure is detected, the indicator function is 1, otherwise it is 0. The penalty weight representing the risk of deadlock is usually set to an extremely large value (such as infinity) to ensure that once a logical loop is detected, the system will directly trigger the highest level of risk alert.

[0079] S30: Input the risk characteristics of the graph and make risk judgments based on the regulations of the power grid and coal mining industries.

[0080] The aforementioned power grid and coal mining industry regulations cover industry access standards and internal red line regulations for enterprises, and implement differentiated analysis for the power grid and coal mining sectors.

[0081] In an exemplary embodiment, step S30 includes:

[0082] S301, Input Risk characteristics of the map;

[0083] S302, Target Clauses Based on Regulatory Analysis of Power Grid and Coal Mining Industries;

[0084] Specifically, this calls a large model with LoRA (Low-Rank Adaptation) adapter weights for the power grid and coal mining industries. This model injects a low-rank matrix into the Attention layer of the Transformer architecture. To achieve a deep semantic understanding of industry-specific systems, the forward propagation calculation formula is as follows;

[0085]

[0086] in, This represents the pre-trained weight matrix of the large model, which remains frozen during fine-tuning to preserve the model's general semantic processing capabilities. This represents the weight update amount, implemented through low-rank decomposition; The projection matrix is ​​typically initialized using a Gaussian distribution and is responsible for reducing the dimensionality of the high-dimensional input. This represents the projection matrix, initially set to 0, which is responsible for converting the feature dimensions back to the original dimensions. Denotes the rank of a low-rank matrix, satisfying With only a few parameter updates, precise alignment of industry terminology and regulations can be achieved. Represents the feature vector of the input sliced ​​text; This represents the final output vector containing industry-specific semantic features.

[0087] S303, based on the analysis results, generates suggestions through a step-by-step approach of risk positioning, legal source tracing, and text reshaping.

[0088] Next, the processes of risk identification, legal origination, and text reshaping will be elaborated:

[0089] Risk identification uses a logical reasoning engine to link the target terms with the system queried in step S201. A comparison is made. If the parameters of the target clause (such as power restoration time) are inferior to the institutional standards, or if necessary clauses (such as mine water exploration and drainage agreements) are missing, it is judged as a "compliance risk".

[0090] Legal source tracing refers to rule matching and source tracing, automatically linking to specific institutional clauses. For example, determining that the clause violates Article X of the "Coal Mine Safety Regulations" which stipulates that "personnel entering the mine must carry self-rescue devices with them."

[0091] Text reshaping, or rewriting suggestion generation, leverages the generation capabilities of a Decoder-only large model to generate suggestions using a standard format of "problem description + risk level + modified text + reference basis".

[0092] For example, the description of the property boundary point in the power grid contract is vague, creating a gap in maintenance responsibility. It is recommended to revise it to use the location of the bushing connection at the substation as the boundary. Refer to Article X of the "Electricity Supply Business Rules".

[0093] S40 generates an initial report based on the risk assessment results. The generated content is checked for compliance and standards through filters. If it does not comply with the rules, it is forced to be rewritten.

[0094] Specifically, the filter is an automated monitoring system integrated behind the core inference layer, and its core logic includes:

[0095] The dual verification mechanism involves the filter performing a secondary scan on the suggestions generated by the large model using a pre-set rule base (containing non-compliant terms in the coal mining / power grid industry and negative samples of legal logic).

[0096] Compliance alignment involves calculating the semantic distance between the generated text and industry standard text to identify sensitive information, illegal elements, or low-quality content, ensuring that the output complies with regulatory platforms or legal requirements.

[0097] Feedback rewrite flow: If the generated content fails the filter verification, the system will trigger a forced rewrite instruction, and the reason for the failure will be fed back to the core inference layer for correction until it meets industry standards.

[0098] S50, based on coordinate reverse mapping technology, maps risk points back to the coordinates of the document to generate an annotation layer, realizing the accurate regression of risk points from the "digital semantic space" to the "physical page space".

[0099] Specifically, the coordinate reverse mapping technology involves the system calling the absolute coordinate metadata of the text blocks recorded in the segmentation-page stage, and using a mapping algorithm to convert the risk point index identified by the core inference layer into pixel-level coordinate information (i.e., page number x, y axis coordinates) in the original PDF / WORD / OFD document.

[0100] Furthermore, the annotation layer is an independent level within a document or image used for adding, marking, or other supplementary information, without affecting the integrity of the original content. At this level, the system generates visual highlights or pop-up annotations based on calculated coordinates, allowing users to directly view the corresponding compliance risk points and suggested modifications within the original document interface.

[0101] This embodiment effectively solves the problem of ambiguous risk positioning in traditional text review by using pixel-level linkage, and realizes pixel-level highlighting of risk points on the original document, which greatly reduces the workload of manual maintenance and secondary comparison.

[0102] In the above embodiments of this application, by combining sliding window slicing technology with risk graph algorithm, the "forgetting" problem caused by the context length limitation of Transformer model and cross-chapter logical conflicts (such as logical deadlocks in payment and delivery cycles) that traditional linear scanning cannot detect are effectively solved.

[0103] In addition, by using coordinate reverse mapping technology, risk points are highlighted at the pixel level on the original file, which greatly reduces the maintenance workload and the possibility of errors.

[0104] Based on the above embodiments, this application also provides an intelligent document review system based on a large language model. Figure 2 This is a schematic diagram of the structure of an intelligent document review system based on a large language model, as described in an embodiment of the present invention. Figure 2 As shown, this intelligent document review system based on a large language model may include:

[0105] The input layer is used to distinguish and extract content related to the main text, tables, headers and footers, and seals.

[0106] The data processing layer is used to process the extracted relevant content;

[0107] The knowledge and hint layer is used to receive search results, connect to the vector database, and query external regulations, internal systems, and historical negative samples.

[0108] The core inference layer is used to deploy large Decoder-only models and to mount industry-specific fine-tuning weights.

[0109] The output and interaction layer is used to generate visual annotations at the corresponding coordinate positions in the original document. It is also used to output an interpretable review report containing risk classification, legal basis and customized rewriting suggestions based on coordinate reverse mapping technology, and provides a human proofreading feedback interface to realize dynamic correction and incremental learning of the knowledge base.

[0110] In an exemplary embodiment, the data processing layer includes:

[0111] The segmentation-pagination engine is used to divide the parsed unstructured text into slices that maintain semantic integrity. It is also used to record the page index, spatial coordinates and hierarchical topology of each slice in the original document based on the document's physical page break and logical paragraph identifier.

[0112] The vectorized embedding module is used to convert non-numerical data into low-dimensional dense vectors that can be understood by computers, while preserving the semantic information or feature relationships of the data.

[0113] In an exemplary embodiment, the knowledge and prompting layer includes:

[0114] A dynamic suggestion builder is used to accept search results and connect to a vector database;

[0115] The three-dimensional knowledge base includes an external legal database, an internal institutional database, and a historical negative sample database. The external legal database is used to query external laws and regulations, the internal institutional database is used to query internal institutional regulations, and the historical negative sample database is used to query historical negative samples.

[0116] In an exemplary embodiment, the core inference layer includes:

[0117] The base model is used to deploy large Decoder-only models as the foundation for semantic understanding and text generation.

[0118] The adapter module is used to attach lightweight weights for fine-tuning the corpus of the power grid and coal mining industries, so as to achieve standard alignment of the model with industry professional terms and unique systems.

[0119] The risk graph generator is used to calculate the conflict detection and risk propagation between clauses. Based on the risk graph algorithm, it identifies deep risks by recognizing logical loops and parameter contradictions in the clause dependency graph.

[0120] Optionally, the output and interaction layer includes:

[0121] An interpretable report generator is used to assemble risk assessment results, legal basis, and modification suggestions into a structured report, and call coordinate reverse mapping data to generate interactive annotations on the original document;

[0122] The feedback loop interface is used to collect human confirmation, ignoring, or correction behaviors of the review results, and transform them into fine-tuning samples to be sent back to the core inference layer in order to achieve dynamic optimization of the system review criteria.

[0123] The above embodiments, by introducing a three-dimensional knowledge base, ensure that the generated reports strictly follow the structure of "conclusion + legal basis + modification suggestions," thus solving the AI ​​"illusion" problem.

[0124] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0125] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0126] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0127] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0128] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0129] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0130] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0131] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0135] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

[0137] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. An intelligent document review method based on a large language model, characterized in that, include: S10, Input a document, parse the document, and slice the parsed document; S20, convert the sliced ​​text into vectors and construct the term dependency graph; S30, Input risk characteristics of the graph and make risk assessment based on the regulations of the power grid and coal mining industries; S40: Based on the risk assessment results, an initial report is generated. The generated content is checked for compliance and standards through filters. If it does not comply with the rules, it is forced to be rewritten. S50, based on coordinate reverse mapping technology, maps risk points back to the coordinates of the document to generate an annotation layer.

2. The method according to claim 1, characterized in that, Step S10 includes: S101, Input a PDF / WORD / OFD document and remove watermark noise from the document; S102, perform OCR recognition on the document after removing watermark noise, and perform sliding window slicing to divide the data stream into semantic retrieval stream and graph construction stream.

3. The method according to claim 2, characterized in that, Step S20 includes: S201, slice the text Convert to embedding vectors and query Tok-k regulations in a vector database. and system ; S202, based on graph construction flow, employs a dual-driven model of model and rules to extract entities and relationships, and constructs a term dependency graph. .

4. The method according to claim 3, characterized in that, Step S30 includes: S301, Input Risk characteristics of the map; S302, Target Clauses Based on Regulatory Analysis of Power Grid and Coal Mining Industries; S303, based on the analysis results, generates suggestions through a step-by-step approach of risk positioning, legal source tracing, and text reshaping.

5. An intelligent document review system based on a large language model, characterized in that, include: The input layer is used to distinguish and extract content related to the main text, tables, headers and footers, and seals. The data processing layer is used to process the extracted relevant content; The knowledge and hint layer is used to receive search results, connect to the vector database, and query external regulations, internal systems, and historical negative samples. The core inference layer is used to deploy large Decoder-only models and to attach industry-specific fine-tuning weights. The output and interaction layer is used to generate visual annotations at the corresponding coordinate positions in the original document. It is also used to output an interpretable review report containing risk classification, legal basis and customized rewriting suggestions based on coordinate reverse mapping technology, and provides a human proofreading feedback interface to realize dynamic correction and incremental learning of the knowledge base.

6. The system according to claim 5, characterized in that, The data processing layer includes: The segmentation-pagination engine is used to divide the parsed unstructured text into slices that maintain semantic integrity. It is also used to record the page index, spatial coordinates and hierarchical topology of each slice in the original document based on the document's physical page break and logical paragraph identifier. The vectorized embedding module is used to convert non-numerical data into low-dimensional dense vectors that can be understood by computers, while preserving the semantic information or feature relationships of the data.

7. The system according to claim 5, characterized in that, The knowledge and hint layer includes: A dynamic suggestion builder is used to accept search results and connect to a vector database; The three-dimensional knowledge base includes an external legal database, an internal institutional database, and a historical negative sample database. The external legal database is used to query external laws and regulations, the internal institutional database is used to query internal institutional regulations, and the historical negative sample database is used to query historical negative samples.

8. The system according to claim 5, characterized in that, The core inference layer includes: The base model is used to deploy Decoder-only large models, serving as the foundation for semantic understanding and text generation. The adapter module is used to attach lightweight weights for fine-tuning the corpus of the power grid and coal mining industries, so as to achieve standard alignment of the model with industry professional terms and unique systems. The risk graph generator is used to calculate the conflict detection and risk propagation between clauses. Based on the risk graph algorithm, it identifies deep risks by recognizing logical loops and parameter contradictions in the clause dependency graph.

9. The system according to claim 5, characterized in that, The output and interaction layer includes: An interpretable report generator is used to assemble risk assessment results, legal basis, and modification suggestions into a structured report, and call coordinate reverse mapping data to generate interactive annotations on the original document; The feedback loop interface is used to collect human confirmation, ignoring, or correction behavior of the review results, and transform it into fine-tuning samples to be sent back to the core inference layer in order to achieve dynamic optimization of the system review criteria.

Images