An AI review-based file reply generation system

Abstract

The application relates to the technical field of document reply generation, and discloses a file reply generation system based on AI auditing, which comprises the following modules: a document analysis module, which is used for obtaining an original document, performing fragmentation, carrying out semantic vector conversion, and constructing an evidence graph; a gist mapping module, which is used for analyzing review gists into computable rules, calculating matching confidence and evidence coverage, and outputting rule mapping results; a fact resolution module, which is used for extracting fact claims, determining and resolving contradictions, and updating evidence graph entity variables; a violation judgment module, which is used for substituting rules to judge violation gists and generating a severity score list; a rectification optimization module, which is used for selecting optimal rectification measures and forming a rectification scheme table; a reply generation module, which is used for calculating residual risks and generating a reply text with regulation anchor points and evidence anchor points; and an evidence archiving module, which is used for writing into a hash chain and archiving the reply document. The application realizes automatic document auditing, reply generation and whole-process tamper-proofing and trace-keeping.

Description

Technical Field

[0001] This invention belongs to the field of document approval generation technology, specifically relating to an AI-based document approval generation system. Background Technology

[0002] With the increasing complexity of various administrative management, corporate governance, and compliance review tasks, document review and approval play an increasingly important role in practical operations. Traditional document review processes rely primarily on manual word-by-word review, comparison, and annotation, which is not only time-consuming and labor-intensive but also easily influenced by the reviewer's experience and subjective judgment, leading to inconsistent conclusions or omissions. Furthermore, when the review involves multiple documents, various rules, and facts from different sources, manual methods often fail to promptly identify potential contradictions or logical conflicts across documents, reducing the reliability and authority of the review conclusions.

[0003] In recent years, the development of artificial intelligence, especially natural language processing technology, has made semantic understanding, key point extraction, and automated analysis of documents possible. While some existing document review systems have introduced intelligent retrieval or keyword matching functions, most remain at a superficial level of semantic recognition, lacking systematic processing for factual consistency, rule mapping, and the quantification of serious violations. Furthermore, in the generation of approval documents, common methods often involve filling in fixed templates, lacking quantitative evaluation of sentence structure standardization and interpretability, resulting in deficiencies in readability and traceability. In addition, traditional electronic document archiving methods struggle to guarantee the immutability and integrity of the evidence; when approval documents are needed for auditing, accountability, or traceability, their authenticity is often difficult to fully verify. Summary of the Invention

[0004] This invention provides an AI-based document approval generation system, which solves the technical problems of low document review efficiency, lack of objectivity in rule judgment, lack of standardization and traceability of approval documents, and difficulty in preventing tampering of evidence in related technologies.

[0005] This invention provides an AI-based document approval generation system, comprising:

[0006] The document parsing module is used to obtain the original document to be reviewed, divide it into segments by fixed window, perform semantic vector transformation, extract key entities and entity relationships, and construct an evidence graph.

[0007] The key point mapping module is used to obtain the review key points corresponding to the matters to be reviewed, parse the verification conditions described in natural language in the review key points into computable rules, calculate the matching confidence and evidence coverage of the review key points with the evidence map and document fragments, and output the rule mapping results.

[0008] The fact resolution module is used to extract factual claims from document fragments and evidence graphs, determine the contradictions of factual claims based on natural language, resolve conflicts, obtain a set of uncontradictory facts, and use this set to update the entity variables of the evidence graph.

[0009] The violation determination module is used to substitute a set of uncontradictory facts into computable rules, determine the key points of the violation based on the matching confidence and evidence coverage, and calculate a severity score list according to preset weights.

[0010] The rectification and optimization module is used to obtain rectification measures from the preset rectification measure library for each violation point, calculate the residual severity after taking rectification measures, construct an objective function including residual severity, cost and schedule, determine the unique rectification measure, and summarize to obtain a rectification plan table;

[0011] The approval generation module is used to calculate the overall residual risk value based on the rectification plan table and determine the approval conclusion; at the same time, based on the key points of the violation and the severity score list, it selects the sentence with the highest clarity score from a fixed candidate sentence library, sorts and filters the clauses according to the severity score, fills in the fixed sentence and attaches legal anchors and evidence anchors, and generates the approval text with double anchors.

[0012] The evidence archiving module is used to write the approval text and double anchor points into a hash chain to form audit records, export documents containing the approval text with double anchor points, and archive them for record keeping.

[0013] Furthermore, the construction of the evidence map includes:

[0014] Step 11: Represent the original document to be reviewed as a symbol sequence based on the preset sub-word decomposition rules, and segment it according to a fixed window length and a fixed step size to obtain document fragments;

[0015] Step 12: Input each document fragment into a semantic encoding model based on the Transformer structure to obtain the normalized semantic vector of the document fragment;

[0016] Step 13: Based on the normalized semantic vector of the document fragment, extract key entities using an entity recognition model and identify the semantic relationships between key entities using a relation extraction model to form an entity set and a relation set.

[0017] Step 14: Based on the entity set and relation set, a directed graph structure is formed, and the normalized semantic vector of the document fragment is bound to the fragment source to construct an evidence graph containing entity nodes, relation edges and vector indexes.

[0018] Furthermore, the key entities include: personnel, job title, date, and numerical value;

[0019] The entity recognition model is a named entity recognition model based on conditional random fields;

[0020] The relation extraction model is constructed based on dependency syntax rules, and the semantic relations it identifies include: job title relations, superior-subordinate relations, time-based relations, and statistical numerical relations.

[0021] Furthermore, the key point mapping module specifically includes:

[0022] Step 21: Parse the natural language verification conditions in the review points into computable rules composed of variables, relational operators and thresholds, and extract the key evidence fields of the verification conditions to form an evidence field set; the variables are derived from key entities extracted from the evidence graph.

[0023] Step 22: Convert the review points into point vectors, calculate the cosine similarity with the normalized semantic vector of the document fragment, obtain the similarity value, and use the maximum similarity as the matching confidence of the review point.

[0024] Step 23: Based on the evidence field set, count the fields that appear in the evidence graph entity set or document fragment as covered fields, and calculate the ratio of the number of covered fields to the number of fields in the evidence field set to obtain the evidence coverage.

[0025] Step 24: Based on the comprehensive review of key points, computable rules, matching confidence, and evidence coverage, output the rule mapping results and store them in the form of quadruples.

[0026] Furthermore, factual claims are extracted from document fragments and evidence graphs. These claims are then subjected to contradiction assessment based on natural language, and conflicts are resolved to obtain a set of uncontradictory facts. This set is then used to update the entity variables of the evidence graph, including:

[0027] Step 31: Extract fact claims from document fragments and evidence graphs, and standardize the extraction results into a five-tuple format containing subject, predicate, value, time anchor and source fragment. The extraction is completed by a sequence labeling method based on a natural language processing model, and an extraction confidence score is assigned to each fact claim. When the extraction confidence score is lower than a preset confidence score threshold, the claim is removed.

[0028] Step 32: For factual claims with the same entity and predicate type and overlapping time anchors, backfill them into natural language sentences and input them into the pre-trained language model to obtain the semantic level of contradiction probability. Combine the numerical consistency verification results for weighted calculation to obtain the contradiction probability between factual claims and construct a contradiction graph.

[0029] Step 33: In the contradiction graph, for claim pairs with contradictory edges, the claim reliability is determined based on the claim reliability, which is obtained by weighting the extracted confidence and the matching confidence of the source fragment; for two contradictory fact claims, the one with higher claim reliability is retained, and fact claims with claim reliability lower than the preset reliability threshold are deleted using a greedy iterative method until all contradictory edges are eliminated, resulting in a set of contradictory facts;

[0030] Step 34: Aggregate the set of uncontradictory facts according to variable type. For numerical variables, use a weighted average with the claim reliability as the weight. For enumerated variables, use a weighted majority vote with the claim reliability as the weight. For time-related variables, use a weighted median within a preset tolerance range to determine the result. The updated entity variables are obtained and the updated results are written back to the evidence graph.

[0031] Furthermore, by substituting the set of uncontradictory facts into the computable rules, the key points of the violation are determined based on the matching confidence and evidence coverage, and a severity score list is calculated according to preset weights, including:

[0032] Step 41: Substitute the updated entity variables from step 34 into the preset computable rules to obtain a Boolean judgment result, where 1 indicates that the rule is satisfied and 0 indicates that the rule is not satisfied.

[0033] Step 42: For rules whose judgment results are not met, the matching confidence and evidence coverage corresponding to the rule are checked. When the matching confidence and evidence coverage are not lower than the preset threshold in the range of 0 to 1, the violation point corresponding to the rule is judged as a definite violation point; otherwise, it is judged as an uncertain violation point, and the uncertain violation point is excluded from the severity score calculation.

[0034] Step 43: For the identified violation points, the severity score is obtained by weighting the difference between 1 and the matching confidence level, the difference between 1 and the evidence coverage level, and the preset rule weights.

[0035] Step 44: Summarize all identified violations and their corresponding severity scores to form a set of binary pairs, and output the severity score list in descending order of severity score.

[0036] Furthermore, the rectification and optimization module specifically includes:

[0037] Step 51: For each violation point in the severity score list, retrieve candidate rectification measures from the preset rectification measure library. Each rectification measure in the rectification measure library includes three attributes: severity reduction ratio, implementation cost, and time. Calculate the product of the severity score and the severity reduction ratio as the residual severity after the implementation of each candidate rectification measure.

[0038] Step 52: Construct the objective function by weighted summation of residual severity, implementation cost, and schedule, and calculate the objective function value of the candidate remedial measures;

[0039] Step 53: Under the constraints that the implementation cost does not exceed the preset budget and the construction period does not exceed the preset time limit, select the rectification measure with the smallest objective function value from the candidate rectification measures and determine it as the unique rectification measure for the violation point. Then, summarize all the violation points and their corresponding unique rectification measures to form a rectification plan table.

[0040] Furthermore, the approval generation module specifically includes:

[0041] Step 61: Calculate the weighted sum of the residual severity of each violation in the rectification plan table, calculate the overall residual risk value, compare it with the preset risk threshold, and determine the approval conclusion.

[0042] Step 62: For each violation point in the severity score list, retrieve the set of sentence patterns corresponding to the violation point type from the preset candidate sentence pattern library, and calculate the clarity score based on syntactic completeness, semantic conciseness and interpretability, and select the sentence pattern with the highest clarity score as the corresponding template.

[0043] Step 63: Sort the key points of the violation and their corresponding sentence structures in descending order of severity score in the severity score list, and fill them into the template in sequence to generate a sequence of approval clauses;

[0044] Step 64: Add legal anchors and evidence anchors to the approval clauses to generate a complete approval text containing the approval conclusion and the approval clauses with dual anchors; the legal anchors are legal article indexes, and the evidence anchors are entity indexes in the evidence graph.

[0045] Furthermore, clarity scores are calculated based on syntactic completeness, semantic conciseness, and interpretability, including:

[0046] Syntactic integrity is obtained by calculating the ratio of the number of correct dependency relations to the total number of dependency relations through dependency parsing;

[0047] Calculate the average word length and the proportion of redundant words, and then perform a weighted summation and standardization to obtain the semantic conciseness.

[0048] Interpretability is obtained by calculating the ratio of the number of core entities covering the key violations to the required number of entities;

[0049] Clarity score is obtained by weighted summation of syntactic completeness, semantic conciseness, and interpretability.

[0050] Furthermore, the approval text and the two anchor points are written into a hash chain to form an audit record. The document containing the approval text with the two anchor points is exported and archived, including:

[0051] Step 71: Concatenate the approval text with the legal anchor and evidence anchor in a structured manner into a data block, wherein the concatenation method is JSON format, and a generated timestamp is appended to the data block. Calculate the hash value of the data block using a one-way hash function.

[0052] Step 72: Write the hash value into the hash chain as the block header of the new block, and record the hash value and metadata of the previous block. The metadata includes: operator identifier, generation timestamp and system version number.

[0053] Step 73: Generate and archive the approval document based on the data block. The document content includes the approval text, legal anchors, and evidence anchors, and appends the hash value corresponding to the data block to the end of the document.

[0054] The beneficial effects of this invention are as follows: By segmenting the original document, semantically vectorizing it, and constructing an evidence graph, this invention achieves a structured representation of the document content, enabling accurate extraction of key entities and relationships, and providing high-quality data support for subsequent review; by parsing the review points described in natural language into computable rules and combining them with matching confidence and evidence coverage for judgment, it avoids the subjectivity and omissions inherent in traditional manual review; by introducing a fact claim conflict detection and resolution mechanism, it ensures the consistency and reliability of the fact set; simultaneously, by using severity scoring and optimization models to automatically select rectification measures, it achieves a balance between risk reduction, cost, and schedule, generating a scientifically reasonable rectification plan table; furthermore, by introducing clarity scoring and combining it with legal anchors and evidence anchors to output standardized approval documents, it improves the readability and traceability of the document. Finally, by writing and archiving hash chains, it achieves full-process traceability and tamper-proof evidence preservation of the approval document, significantly enhancing the authority and legal effect of the approval results. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of a document approval generation system based on AI review according to the present invention. Detailed Implementation

[0056] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0057] like Figure 1As shown, an AI-based document approval generation system includes:

[0058] The document parsing module is used to obtain the original document to be reviewed, divide it into segments by fixed window, perform semantic vector transformation, extract key entities and entity relationships, and construct an evidence graph.

[0059] The key point mapping module is used to obtain the review key points corresponding to the matters to be reviewed, parse the verification conditions described in natural language in the review key points into computable rules, calculate the matching confidence and evidence coverage of the review key points with the evidence map and document fragments, and output the rule mapping results.

[0060] The fact resolution module is used to extract factual claims from document fragments and evidence graphs, determine the contradictions of factual claims based on natural language, resolve conflicts, obtain a set of uncontradictory facts, and use this set to update the entity variables of the evidence graph.

[0061] The violation determination module is used to substitute a set of uncontradictory facts into computable rules, determine the key points of the violation based on the matching confidence and evidence coverage, and calculate a severity score list according to preset weights.

[0062] The rectification and optimization module is used to obtain rectification measures from the preset rectification measure library for each violation point, calculate the residual severity after taking rectification measures, construct an objective function including residual severity, cost and schedule, determine the unique rectification measure, and summarize to obtain a rectification plan table;

[0063] The approval generation module is used to calculate the overall residual risk value based on the rectification plan table and determine the approval conclusion; at the same time, based on the key points of the violation and the severity score list, it selects the sentence with the highest clarity score from a fixed candidate sentence library, sorts and filters the clauses according to the severity score, fills in the fixed sentence and attaches legal anchors and evidence anchors, and generates the approval text with double anchors.

[0064] The evidence archiving module is used to write the approval text and double anchor points into a hash chain to form audit records, export documents containing the approval text with double anchor points, and archive them for record keeping.

[0065] In one embodiment of the present invention, the construction of the evidence map includes:

[0066] Step 11: The original document to be reviewed is represented as a symbol sequence based on a preset sub-word decomposition rule, and then segmented according to a fixed window length and a fixed step size to obtain document fragments. This method avoids the problem of low processing efficiency of the semantic encoding model due to excessively long documents, while ensuring that the text content maintains contextual continuity after slicing. For example, when the review object is a request for a leadership reshuffle, it can be split into fragments such as a personnel list, appointment descriptions, and a schedule, each fragment capable of independently carrying a relatively complete semantic unit.

[0067] Step 12: Input each document fragment into a semantic encoding model based on the Transformer structure to obtain a normalized semantic vector of the document fragment. This semantic vector can capture the semantic features of the text fragment, so that subsequent processing can be performed based on numerical vectors for calculation and comparison, rather than relying on manual sentence-by-sentence review.

[0068] Step 13: Based on the normalized semantic vector of the document fragment, key entities are extracted using an entity recognition model, and semantic relationships between key entities are identified using a relation extraction model, forming an entity set and a relation set. The key entities include: personnel, positions, dates, and numerical values. The entity recognition model is a named entity recognition model based on conditional random fields, used to label the key entities contained in the text fragment from the normalized semantic vector. The relation extraction model is constructed based on dependency syntax rules and is used to identify semantic relationships between different entities. The identified semantic relationships include: job title relationships, superior-subordinate relationships, time-related relationships, and statistical numerical relationships. In this way, the semantic information in the original document is structured into clear entities and relationships, avoiding misunderstandings caused by diverse expression methods during manual review.

[0069] Step 14 involves constructing a directed graph structure based on the entity set and relation set, and binding the normalized semantic vectors of document fragments to their sources to build an evidence graph containing entity nodes, relation edges, and vector indices. This evidence graph not only visually displays the semantic logic of the relationships existing in the original document but also provides unified data support for subsequent review point matching and factual conflict determination. For example, when the system needs to verify whether personnel appointments comply with regulations, it can directly retrieve relevant personnel nodes and their corresponding appointment relationships from the evidence graph without re-parsing the original text. This approach greatly improves the efficiency and accuracy of document review and reduces the risk of omissions or ambiguities during manual review.

[0070] In one embodiment of the present invention, the key point mapping module specifically includes:

[0071] Step 21: Parse the natural language verification conditions in the review criteria into computable rules composed of variables, relational operators, and thresholds, and extract the key evidence fields of the verification conditions to form an evidence field set; the variables are derived from key entities extracted from the evidence graph; the relational operators include: =, ≠, >, <, ≥, ≤; for example, when the review criteria is that the candidate's age must not be lower than 35 years old, the system will parse the data to obtain the following: variable: age, relational operator: ≥, threshold: 35, and extract the key evidence field as the candidate's age.

[0072] Step 22: Convert the review points into point vectors, calculate the cosine similarity with the normalized semantic vector of the document fragment, obtain the similarity value, and use the maximum similarity as the matching confidence of the review point.

[0073] Step 23: Based on the evidence field set, count the fields that appear in the evidence graph entity set or document fragment as covered fields, and calculate the ratio of the number of covered fields to the number of fields in the evidence field set to obtain the evidence coverage. The evidence coverage can determine the degree of semantic matching and whether the review point is supported by sufficient entity evidence.

[0074] Step 24: Based on the comprehensive review points, computable rules, matching confidence level, and evidence coverage, output the rule mapping results and store them in the form of quadruples for easy access by subsequent modules. This enables unified management and automated judgment of the correspondence between different review points and document evidence.

[0075] In one embodiment of the present invention, factual claims are extracted based on document fragments and evidence graphs, contradictions are determined based on natural language, and conflicts are resolved to obtain a set of uncontradictory facts. This set is then used to update the entity variables of the evidence graph, including:

[0076] Step 31 involves extracting factual claims from document fragments and evidence graphs, standardizing the extraction results into a five-tuple format containing subject, predicate, value, time anchor, and source fragment. This five-tuple structure expresses the basic elements of facts in a unified data structure, making subsequent conflict determination and aggregation calculations more convenient. The extraction is accomplished using a sequence labeling method based on a natural language processing model, i.e., word-by-word labeling of the input text to identify the subject, predicate, and value in the facts, and extracting the corresponding time anchor and document source information from the context. An extraction confidence score is assigned to each factual claim, and claims with an extraction confidence score lower than a preset confidence threshold are discarded. The extraction confidence score measures the probability that the claim is correctly identified.

[0077] Step 32 involves determining contradictions in factual claims that share the same entity, predicate type, and overlapping time anchors. Specifically, these claims are backfilled into natural language sentences and input into a pre-trained language model. The model's natural language reasoning capabilities are used to calculate the semantic contradiction probability between claims. Simultaneously, consistency checks are performed on factual claims involving numerical values. For example, if two claims both involve term limits, the numerical differences are checked to ensure they do not exceed acceptable deviations. Finally, the semantic contradiction probability and numerical consistency results are weighted and fused to obtain the comprehensive contradiction probability between claims. Based on this, the system constructs a contradiction graph, where nodes represent factual claims, edges represent contradictory relationships, and edge weights represent the claim contradiction probability.

[0078] Step 33: In the contradiction graph, for claim pairs with contradictory edges, the claim reliability is determined based on the claim reliability. The claim reliability is obtained by weighting the extracted confidence and the matching confidence of the source fragment. For two contradictory fact claims, the one with higher reliability is retained, and fact claims with reliability lower than a preset reliability threshold are deleted using a greedy iterative method until all contradictory edges are eliminated, resulting in a set of uncontradictory facts. Through this step, the system can automatically filter out facts with low confidence or mutual contradictions, ensuring the internal consistency and reliability of the fact set.

[0079] Step 34: Aggregate the set of uncontradictory facts according to variable type. For numerical variables, use a weighted average with the claim reliability as the weight. For enumerated variables, use a weighted majority vote with the claim reliability as the weight. For time-related variables, use a weighted median within a preset tolerance range to determine the result. The updated entity variables are obtained and the updated results are written back to the evidence graph to ensure that the entity status in the evidence graph remains consistent and up-to-date.

[0080] Through the above steps, this invention can automatically extract factual claims, detect and resolve conflicts when multiple source documents contain differing or even conflicting statements, and write consistent factual information back into the evidence graph. This not only improves the accuracy and consistency of factual data but also provides a more reliable data foundation for subsequent determination of violations and generation of approval content.

[0081] In one embodiment of the present invention, a set of uncontradictory facts is substituted into a computable rule, and the key points of violation are determined based on the matching confidence and evidence coverage. A severity score list is then calculated according to preset weights, including:

[0082] Step 41: Substitute the updated entity variables from step 34 into the preset computable rules to obtain a Boolean judgment result, where 1 indicates that the rule is satisfied and 0 indicates that the rule is not satisfied. For example, for the rule: candidate age ≥ 35, if the age of a candidate in the set of uncontradictory facts is 42, the judgment result is 1, indicating that the condition is satisfied; if the age is 30, the judgment result is 0, indicating that the condition is not satisfied.

[0083] Step 42: For rules whose judgment results are not met, the matching confidence and evidence coverage corresponding to the rule are checked. When the matching confidence and evidence coverage are not lower than the preset threshold in the range of 0 to 1, the violation point corresponding to the rule is judged as a definite violation point; otherwise, it is judged as an uncertain violation point, and the uncertain violation point is excluded from the severity score calculation.

[0084] Step 43: For the identified violation points, the severity score is obtained by weighting the difference between 1 and the matching confidence level, the difference between 1 and the evidence coverage level, and the preset rule weight. The rule weight indicates the importance of the identified violation point in the overall review task, and the severity score can comprehensively measure the severity of the violation point.

[0085] Step 44: Summarize all identified violations and their corresponding severity scores to form a set of binary pairs, and output the severity score list in descending order of severity score.

[0086] Through the above steps, this invention achieves full-process automation from fact set to identification of key violations and quantification of severity, which not only ensures the objectivity of the review results, but also improves the system's ability to process complex documents and control risks.

[0087] In one embodiment of the present invention, the rectification and optimization module specifically includes:

[0088] Step 51: For each violation in the severity score list, retrieve candidate rectification measures from the pre-set rectification measure library. Each rectification measure in the library includes three attributes: severity reduction ratio, implementation cost, and time. The severity reduction ratio is used to indicate the degree to which the rectification measure alleviates the severity of the violation, with a value ranging from 0 to 1. The larger the value, the more significant the rectification effect. Calculate the product of the severity score and the severity reduction ratio as the residual severity after the implementation of each candidate rectification measure, which is used to quantify the actual effect of different measures in reducing risk.

[0089] Step 52: Construct the objective function by weighted summation of residual severity, implementation cost, and schedule, and calculate the objective function value of the candidate remedial measures;

[0090] Step 53: Under the constraints of implementation cost not exceeding the preset budget and construction period not exceeding the preset time limit, select the rectification measure with the smallest objective function value from the candidate rectification measures, and determine it as the unique rectification measure for the violation point. Then, summarize all violation points and their corresponding unique rectification measures to form a rectification plan table. Each record in the rectification plan table includes the violation point identifier, the selected rectification measure, the residual severity, implementation cost, and construction period.

[0091] Through the above steps, this invention not only enables automated optimization and selection among multiple candidate rectification measures, but also balances risk reduction, resource investment, and time constraints, ensuring that the output rectification plan is both scientifically sound and feasible. For example, in practical applications, the system can prioritize short-term, low-cost rectification measures based on budget constraints, while selecting long-term, efficient measures when budget constraints are not present, thereby making the generated plan more compliant with regulatory requirements and actual operating conditions.

[0092] In one embodiment of the present invention, the approval generation module specifically includes:

[0093] Step 61: Calculate the weighted sum of the residual severity of each violation in the rectification plan table to determine the overall residual risk value. Compare this value with a preset risk threshold to determine the approval conclusion. Residual severity refers to the degree of risk that a violation still poses after rectification measures are implemented. During the weighted summation, each violation is assigned a weight based on its importance to ensure that the influence of different categories of violations is reflected in the overall risk calculation. Approval conclusions include, but are not limited to: approved, further rectification required, or not approved.

[0094] Step 62: For each violation point in the severity scoring list, retrieve the set of sentence patterns corresponding to the violation point type from the preset candidate sentence pattern library, and calculate the clarity score based on syntactic completeness, semantic conciseness and interpretability, and select the sentence pattern with the highest clarity score as the corresponding template; wherein, the candidate sentence pattern library is a fixed setting, containing a variety of compliance approval expression templates that have been manually reviewed to ensure that the language is standardized and formal;

[0095] Clarity is scored by weighted summation of syntactic completeness, semantic conciseness, and interpretability. Syntactic completeness is calculated by using dependency parsing to determine the ratio of correct dependency relations to the total number of dependencies, reflecting the rigor of the sentence structure. Semantic conciseness is calculated by weighted summation and standardization of average word length and redundant word ratio, reflecting the conciseness of the sentence expression. Interpretability is calculated by weighted summation of the number of core entities covering the violation to the required number of entities, reflecting whether the sentence can fully explain the violation.

[0096] Step 63: Sort the key points of the violation and their corresponding sentence structures in descending order of severity score in the severity score list, and fill the template in sequence to generate a sequence of approval clauses; through this sorting method, the violation clauses with higher severity will be presented first in the approval text, so that the reviewers or users can quickly focus on the key risk points and improve the organization and highlighting of key points in the approval document.

[0097] Step 64: Add legal anchors and evidentiary anchors to the approval clauses to generate a complete approval text containing the approval conclusion and the approval clauses with dual anchors; the legal anchors are indexes of legal provisions, clarifying the legal or policy basis corresponding to the clauses; the evidentiary anchors are entity indexes in the evidence graph, indicating the factual source on which the clauses are based; through the binding of dual anchors, each clause in the approval text can achieve traceability and verifiability in both legal and evidentiary dimensions.

[0098] Through the above steps, this invention can further generate quantitative risk conclusions and structured approval texts after automatically generating rectification plans, thus realizing a complete closed loop from data-driven risk assessment to standardized document generation.

[0099] In one embodiment of the present invention, the approval text and dual anchor points are written into a hash chain to form an audit record, and a document containing the approval text with dual anchor points is exported and archived for record keeping, including:

[0100] Step 71: The approval text, legal anchors, and evidence anchors are concatenated into a data block in a structured manner. The concatenation method is JSON format, and a generated timestamp is appended to the data block. A one-way hash function is used to calculate the hash value of the data block. This hash value can serve as a digital fingerprint of the data block, ensuring that even if any character in the document is modified, the hash value will change significantly, thereby achieving the technical effect of preventing tampering.

[0101] Step 72: The hash value is written into the hash chain as the block header of the new block, and the hash value and metadata of the previous block are recorded. The metadata includes: operator identifier, generation timestamp, and system version number. Through mutual references between the hash values ​​of the new block and the previous block, the system forms a chained storage structure, that is, each block points to the hash value of the previous block, thereby constructing an immutable linked list similar to a blockchain. In this way, even if the data in a block is maliciously modified, the tampering can be detected immediately because its hash value no longer matches the reference in subsequent blocks, thus ensuring the integrity and credibility of the audit records.

[0102] Step 73: Generate and archive the approval document based on the data block. The document content includes the approval text, legal anchors, and evidence anchors. The hash value corresponding to the data block is appended to the end of the document to ensure consistency between the document content and the hash chain storage result. In this way, when users review the approval document, they can not only see the conclusive content and the source information of relevant laws and evidence, but also determine whether the document has been tampered with by verifying whether the hash value attached to the document matches the hash chain storage record. After the document is archived, the system will create an index in the archive repository, supporting fast retrieval and tracing based on the document's unique identifier or hash value.

[0103] Through the above steps, this invention enables traceability and tamper-proof evidence preservation throughout the entire process while generating the approval document. The approval document serves as both an official output file for compliance review and provides non-repudiation and tamper-proof technical safeguards through hash chain storage and archive indexing, thereby enhancing the credibility and legal validity of the approval results.

[0104] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0105] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.

Claims

1. A document approval generation system based on AI review, characterized in that, include: The document parsing module is used to obtain the original document to be reviewed, divide it into segments by fixed window, perform semantic vector transformation, extract key entities and entity relationships, and construct an evidence graph. The key point mapping module is used to obtain the review key points corresponding to the items to be reviewed, parse the verification conditions described in natural language in the review key points into computable rules, calculate the matching confidence of the review key points based on the similarity between the review key points and the document fragments, calculate the evidence coverage based on the number of fields between the verification conditions and the evidence graph or between the verification conditions and the document fragments, and output the rule mapping results. The fact resolution module extracts factual claims from document fragments and evidence graphs, performs contradiction assessments on these claims using natural language, resolves conflicts, and obtains a set of uncontradictory facts. This set is then used to update the entity variables of the evidence graph, including: Step 31: Extract factual claims from document fragments and evidence graphs, assign extraction confidence to each factual claim, eliminate factual claims with extraction confidence below the preset confidence threshold, and standardize the extraction results into a five-tuple format containing subject, predicate, value, time anchor, and source fragment; the extraction confidence is used to measure the probability that the claim is correctly identified. Step 32: For factual claims with the same entity and predicate type and overlapping time anchors, backfill them into natural language sentences and input them into the pre-trained language model to obtain the semantic level of contradiction probability. Combine the numerical consistency verification results for weighted calculation to obtain the contradiction probability between factual claims and construct a contradiction graph. Step 33: In the contradiction graph, for claim pairs with contradictory edges, the claim reliability is determined based on the claim reliability. The claim reliability is obtained by weighting the extracted confidence and the matching confidence of the source fragment. For two contradictory fact claims, the one with higher claim reliability is retained, and fact claims with claim reliability lower than the preset reliability threshold are deleted using a greedy iterative method until all contradictory edges are eliminated, resulting in a set of contradictory facts. The violation determination module is used to substitute a set of uncontradictory facts into a computable rule, determine the key points of violation based on the matching confidence and evidence coverage corresponding to the rule, and calculate a severity score list according to preset weights. The rectification and optimization module is used to retrieve rectification measures from a pre-set rectification measure library for each violation, including three attributes: severity reduction ratio, implementation cost, and schedule. By calculating the product of the severity score and the severity reduction ratio, the residual severity of each violation is obtained after the rectification measures are implemented. An objective function containing residual severity, cost, and schedule is constructed to determine the unique rectification measure and summarize the results into a rectification plan table. The severity reduction ratio is used to indicate the degree to which the rectification measure alleviates the severity of the violation. The approval generation module is used to calculate the weighted sum of the residual severity of each violation point in the rectification plan table to obtain the overall residual risk value and determine the approval conclusion. At the same time, based on the violation points and severity score list, it selects the sentence with the highest clarity score from the sentence set corresponding to the violation point type from a fixed candidate sentence library, sorts and filters the clauses according to the severity score, fills the fixed sentence with legal anchors and evidence anchors, and generates the approval text with double anchors. Clarity score is calculated based on syntactic completeness, semantic conciseness, and interpretability. Syntactic completeness is obtained by calculating the ratio of the number of correct dependency relations to the total number of dependency relations through dependency parsing. Semantic conciseness is obtained by calculating the average word length and the proportion of redundant words, and then performing weighted summation and standardization on the two. Interpretability is obtained by calculating the ratio of the number of core entities covering the violation points to the number of entities required.

2. The document approval generation system based on AI review according to claim 1, characterized in that, The construction of the evidence map includes: Step 11: Represent the original document to be reviewed as a symbol sequence based on the preset sub-word decomposition rules, and segment it according to a fixed window length and a fixed step size to obtain document fragments; Step 12: Input each document fragment into a semantic encoding model based on the Transformer structure to obtain the normalized semantic vector of the document fragment; Step 13: Based on the normalized semantic vector of the document fragment, extract key entities using an entity recognition model and identify the semantic relationships between key entities using a relation extraction model to form an entity set and a relation set. Step 14: Based on the entity set and relation set, a directed graph structure is formed, and the normalized semantic vector of the document fragment is bound to the fragment source to construct an evidence graph containing entity nodes, relation edges and vector indexes.

3. The document approval generation system based on AI review according to claim 2, characterized in that, The key entities include: personnel, job title, date, and numerical value; The entity recognition model is a named entity recognition model based on conditional random fields; The relation extraction model is constructed based on dependency syntax rules, and the semantic relations it identifies include: job title relations, superior-subordinate relations, time-based relations, and statistical numerical relations.

4. The document approval generation system based on AI review according to claim 1, characterized in that, The key point mapping module specifically includes: Step 21: Parse the verification conditions described in natural language in the review points into computable rules composed of variables, relational operators and thresholds, and extract the key evidence fields of the verification conditions to form an evidence field set; the variables are derived from key entities extracted from the evidence graph. Step 22: Convert the review points into point vectors, calculate the cosine similarity with the normalized semantic vector of the document fragment, obtain the similarity value, and use the maximum similarity as the matching confidence of the review point. Step 23: Based on the evidence field set, count the fields that appear in the evidence graph entity set or document fragment as covered fields, and calculate the ratio of the number of covered fields to the number of fields in the evidence field set to obtain the evidence coverage. Step 24: Based on the comprehensive review of key points, computable rules, matching confidence, and evidence coverage, output the rule mapping results and store them in the form of quadruples.

5. The document approval generation system based on AI review according to claim 1, characterized in that, The fact resolution module updates the entity variables of the evidence graph with the set of contradictory facts, including: Step 34: Aggregate the set of uncontradictory facts according to variable type. For numerical variables, use a weighted average with the claim reliability as the weight. For enumerated variables, use a weighted majority vote with the claim reliability as the weight. For time-related variables, use a weighted median within a preset tolerance range to determine the result. The updated entity variables are obtained and the updated results are written back to the evidence graph.

6. The document approval generation system based on AI review according to claim 1, characterized in that, Substituting the set of uncontradictory facts into computable rules, the key points of violation are determined based on matching confidence and evidence coverage, and a severity score list is calculated according to preset weights, including: Step 41: Substitute the updated entity variables from step 34 into the preset computable rules to obtain a Boolean judgment result, where 1 indicates that the rule is satisfied and 0 indicates that the rule is not satisfied. Step 42: For rules whose judgment results are not met, the matching confidence and evidence coverage corresponding to the rule are checked. When the matching confidence and evidence coverage are not lower than the preset threshold in the range of 0 to 1, the violation point corresponding to the rule is judged as a definite violation point; otherwise, it is judged as an uncertain violation point, and the uncertain violation point is excluded from the severity score calculation. Step 43: For the identified violation points, the severity score is obtained by weighting the difference between 1 and the matching confidence level, the difference between 1 and the evidence coverage level, and the preset rule weight; the preset rule weight indicates the importance of the identified violation point in the overall review task. Step 44: Summarize all identified violations and their corresponding severity scores to form a set of binary pairs, and output the severity score list in descending order of severity score.

7. The document approval generation system based on AI review according to claim 1, characterized in that, The rectification and optimization module specifically includes: Step 51: For each violation point in the severity scoring list, retrieve candidate rectification measures from the preset rectification measure library. Each rectification measure in the rectification measure library includes three attributes: severity reduction ratio, implementation cost, and time. Calculate the residual severity after implementing each candidate rectification measure. Step 52: Construct the objective function by weighted summation of residual severity, implementation cost, and schedule, and calculate the objective function value of the candidate remedial measures; Step 53: Under the constraints that the implementation cost does not exceed the preset budget and the construction period does not exceed the preset time limit, select the rectification measure with the smallest objective function value from the candidate rectification measures and determine it as the unique rectification measure for the violation point. Then, summarize all the violation points and their corresponding unique rectification measures to form a rectification plan table.

8. The document approval generation system based on AI review according to claim 1, characterized in that, The approval generation module specifically includes: Step 61: Calculate the weighted sum of the residual severity of each violation in the rectification plan table, calculate the overall residual risk value, compare it with the preset risk threshold, and determine the approval conclusion. Step 62: For each violation point in the severity score list, retrieve the set of sentence patterns corresponding to the violation point type from the preset candidate sentence pattern library, and calculate the clarity score based on syntactic completeness, semantic conciseness and interpretability, and select the sentence pattern with the highest clarity score as the corresponding template. Step 63: Sort the key points of the violation and their corresponding sentence structures in descending order of severity score in the severity score list, and fill them into the template in sequence to generate a sequence of approval clauses; Step 64: Add legal anchors and evidence anchors to the approval clauses to generate a complete approval text containing the approval conclusion and the double-anchor approval clauses; the legal anchors are legal article indexes, and the evidence anchors are entity indexes in the evidence map.

9. The document approval generation system based on AI review according to claim 1, characterized in that, Clarity scores are obtained by weighted summation of syntactic completeness, semantic conciseness, and interpretability.

10. The document approval generation system based on AI review according to claim 1, characterized in that, It also includes an evidence archiving module, used to write the approval text and dual anchor points into a hash chain to form audit records, export documents containing the approval text with dual anchor points, and archive them for record keeping, including: Step 71: Concatenate the approval text with the legal anchor and evidence anchor in a structured manner into a data block, wherein the concatenation method is JSON format, and a generated timestamp is appended to the data block. Calculate the hash value of the data block using a one-way hash function. Step 72: Write the hash value into the hash chain as the block header of the new block, and record the hash value and metadata of the previous block. The metadata includes: operator identifier, generation timestamp and system version number. Step 73: Generate and archive the approval document based on the data block. The document content includes the approval text, legal anchors, and evidence anchors, and appends the hash value corresponding to the data block to the end of the document.

Images