Document review blind spot detection method and apparatus, electronic device, and storage medium

Abstract

The application relates to the technical field of document analysis, and provides a document review blind area detection method and device, electronic equipment and a storage medium, the method comprising the following steps: performing semantic unit division on a target document to be reviewed, and constructing a semantic topology structure based on logical dependency relationships between the semantic units obtained through the division; performing traversal mapping in a regulation knowledge graph based on the semantic topology structure, obtaining a specification node set that should be met by the target document, and generating a theoretical review path according to the specification node set; obtaining specification nodes actually triggered by the target document in a review process, forming an actual review path; calculating a difference set between the theoretical review path and the actual review path, and taking the difference set as a review blind area of the target document. Through difference analysis of the theoretical review path and the actual review path, implicit risks such as rule mismatching, missing of a review path or semantic mapping failure can be identified, so that effective evaluation of the completeness of document review coverage is realized.

Description

Technical Field

[0001] This invention relates to the field of document analysis technology, and in particular to a method, apparatus, electronic device, and storage medium for detecting blind spots in document review. Background Technology

[0002] With the improvement of digital management, official documents and other documents need to comply with multiple internal and external regulations and constraints during the circulation process. The rigor of the review process is directly related to the level of compliance risk prevention and control.

[0003] Current document review methods typically utilize semantic analysis models to understand document content, detect explicit violations or errors, and provide suggestions for modification to some extent. However, these methods often aim to "discover known problems," which means that even if the review result shows no errors, potential risks may still exist due to rule mismatches or missing review paths. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for detecting blind spots in document review, in order to address the deficiencies in the prior art.

[0005] This invention provides a method for detecting blind spots in document review, comprising the following steps:

[0006] The target document to be reviewed is divided into semantic units, and based on the logical dependencies between the semantic units obtained from the division, a semantic topology structure reflecting the internal logical relationships of the target document is constructed.

[0007] Based on the semantic topology, a traversal mapping is performed in the pre-constructed regulatory knowledge graph to obtain the set of normative nodes that the target document should satisfy, and a theoretical review path is generated based on the set of normative nodes.

[0008] Obtain the standard nodes actually triggered during the review process of the target document to form the actual review path;

[0009] Calculate the difference between the theoretical review path and the actual review path, and determine the difference as the review blind spot of the target document.

[0010] According to the document review blind spot detection method provided by the present invention, the step of constructing a semantic topology structure reflecting the internal logical relationships of the target document based on the logical dependencies between the semantic units obtained by segmentation includes:

[0011] Semantic analysis is performed on each semantic unit to obtain the semantic features of each semantic unit;

[0012] Based on the semantic features of each semantic unit, the logical dependencies between semantic units are identified;

[0013] Each semantic unit is treated as a node, and directed edges are established between the nodes based on the logical dependencies to construct the semantic topology.

[0014] According to a document review blind spot detection method provided by the present invention, the step of performing semantic analysis on each semantic unit to obtain the semantic features of each semantic unit includes:

[0015] Extract the text content features and structural attribute features of each semantic unit. The structural attribute features include the chapter level, paragraph position and preceding and succeeding relationships of the corresponding semantic unit in the target document.

[0016] The text content features and the structural attribute features are fused and encoded to generate semantic features for the corresponding semantic units.

[0017] According to the document review blind spot detection method provided by the present invention, the logical dependency relationship includes hierarchical relationship, reference relationship and causal relationship;

[0018] The steps for determining the directed edges include:

[0019] If the above hierarchical relationship exists between any two nodes, establish a hierarchical edge from the upper node to the lower node;

[0020] If the aforementioned reference relationship exists between any two nodes, establish an associated edge from the referencing node to the referenced node;

[0021] If the causal relationship exists between any two nodes, establish a logical edge from the preceding node to the following node.

[0022] According to a document review blind spot detection method provided by the present invention, the step of traversing and mapping a pre-constructed legal knowledge graph based on the semantic topology to obtain a set of normative nodes that the target document should satisfy, and generating a theoretical review path based on the set of normative nodes, includes:

[0023] Calculate the similarity between the semantic vector of each semantic unit in the semantic topology and the semantic vector of each normative node in the regulatory knowledge graph;

[0024] Based on the similarity, the normative nodes associated with each semantic unit are determined from the regulatory knowledge graph;

[0025] Based on the logical dependencies between the normative nodes in the aforementioned regulatory knowledge graph, the associated normative nodes are sorted to generate the theoretical review path.

[0026] According to a document review blind spot detection method provided by the present invention, the step of sorting the associated normative nodes according to the logical dependencies between normative nodes in the regulatory knowledge graph to generate the theoretical review path includes:

[0027] Identify the reference relationships and conflict / exclusion edges of the associated normative nodes in the legal knowledge graph;

[0028] Based on the reference relationship edges, the execution order of each associated specification node is determined;

[0029] Based on the conflict exclusion edges, conflicting standard nodes are removed from the sorted standard nodes to generate the theoretical review path.

[0030] According to the document review blind spot detection method provided by the present invention, in the process of traversing and mapping in a pre-constructed legal knowledge graph based on the semantic topology structure, the method further includes:

[0031] If any semantic unit in the semantic topology does not have a corresponding normative node in the regulatory knowledge graph, a missing mapping record is generated based on the semantic features of the semantic unit.

[0032] The semantic units corresponding to the missing mapping records are marked as missing mapping blind spots.

[0033] According to a document review blind spot detection method provided by the present invention, the step of obtaining the standard nodes actually triggered by the target document during the review process to form the actual review path includes:

[0034] Map the normative nodes actually triggered during the review process of the target document to the corresponding normative nodes in the legal knowledge graph;

[0035] Based on the logical dependencies between the normative nodes mapped in the aforementioned regulatory knowledge graph, the mapped normative nodes are organized to form the actual review path.

[0036] According to a document review blind spot detection method provided by the present invention, determining the difference set as the review blind spot of the target document includes:

[0037] If any specification node contained in the difference set exists in the theoretical review path but does not exist in the actual review path, then the specification node is determined to be a node missing blind spot.

[0038] If both the starting node and the ending node in the theoretical review path exist in the actual review path, and the intermediate node connecting the starting node and the ending node is missing in the actual review path, then the intermediate node is determined to be a path interruption blind spot.

[0039] According to the document review blind spot detection method provided by the present invention, after calculating the difference between the theoretical review path and the actual review path, the method further includes:

[0040] Based on the centrality of the normative nodes contained in the difference set in the regulatory knowledge graph, the risk weight of the audit blind spot is calculated, and the risk level of the audit blind spot is determined based on the risk weight.

[0041] Based on the clause content corresponding to the specification nodes contained in the difference set, modification suggestions are generated for the target document.

[0042] The present invention also provides a document review blind spot detection device, comprising the following modules:

[0043] The segmentation module is used to segment the target document to be reviewed into semantic units, and based on the logical dependencies between the segmented semantic units, to construct a semantic topology that reflects the internal logical relationships of the target document.

[0044] The generation module is used to perform traversal mapping in a pre-constructed regulatory knowledge graph based on the semantic topology structure to obtain a set of normative nodes that the target document should satisfy, and to generate a theoretical review path based on the set of normative nodes.

[0045] The acquisition module is used to acquire the standard nodes actually triggered during the review process of the target document, forming the actual review path;

[0046] The detection module is used to calculate the difference between the theoretical review path and the actual review path, and to determine the difference as the review blind spot of the target document.

[0047] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the document review blind spot detection method as described above.

[0048] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the document review blind spot detection method as described above.

[0049] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the document review blind spot detection method as described above.

[0050] The document review blind spot detection method, device, electronic device, and storage medium provided by this invention divide the target document to be reviewed into semantic units and construct a semantic topology. This topology is then mapped onto a pre-constructed regulatory knowledge graph to generate a theoretical review path. This theoretical path is then compared with the actual review path triggered during the actual review process to calculate the difference set, thereby accurately locating review blind spots. Because this invention goes beyond simply discovering explicit errors, it identifies implicit risks such as rule mismatches, missing review paths, or semantic mapping failures through difference analysis between theoretical and actual review paths. This achieves an effective assessment of the completeness of document review coverage, avoiding compliance risks caused by incomplete review coverage. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the document review blind spot detection method provided by the present invention.

[0053] Figure 2 This is a schematic diagram of the overall architecture of the document review blind spot detection system provided by the present invention.

[0054] Figure 3 This is a flowchart illustrating another document review blind spot detection method provided by the present invention.

[0055] Figure 4 This is a schematic diagram of the document review blind spot detection device provided by the present invention.

[0056] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0058] Currently, document review mainly relies on manual review, rule-based automatic verification systems, intelligent review retrieval and generation systems based on natural language processing or large language models, and compliance review solutions based on knowledge graphs. Among these, manual review mainly relies on experience or fixed lists, lacking systematic modeling; rule-based automatic verification systems detect violations of specific clauses through keyword matching; intelligent review retrieval and generation systems focus on discovering compliance issues in the content itself; and compliance review solutions based on knowledge graphs focus on identifying explicit conflicts.

[0059] However, the aforementioned solutions generally focus on detecting violations or errors, paying only attention to the existence of problems while neglecting whether they have been adequately reviewed. They lack an assessment mechanism for the scope and completeness of the review, making it impossible to determine which content has not yet been reviewed by any rules. Secondly, they lack structured modeling of review paths and coverage relationships. Whether it's rule engines or model-based review, reviews are typically treated as independent judgments, failing to structurally represent the paths between official documents and internal / external regulations, making it difficult to detect breaks in the review chain. Thirdly, review blind spots cannot be explicitly expressed; clauses that do not trigger rules are often considered normal, leading to the overlooking of potentially high-risk areas. Finally, there is a lack of a unified perspective on collaborative review of multi-source regulations. Internal systems and external regulations are often reviewed separately, easily resulting in partial compliance while overall blind spots exist. In this context, even if the review results show no problems, potential compliance risks may still exist due to rule mismatches, missing paths, or failed mappings.

[0060] To address this issue, this invention provides a method for detecting blind spots in document review. The method aims to segment the target document into semantic units and construct a semantic topology. This topology is then mapped onto a pre-built regulatory knowledge graph to generate theoretical review paths. These paths are compared with the actual review paths triggered during the review process to calculate the difference set, thereby accurately locating review blind spots. This allows for the assessment of the completeness of document review coverage, automatically identifying potential compliance risks caused by missing rules, logical breaks, or mapping failures, and filling the gap in existing technologies for review completeness analysis. The method provided by this invention can be applied to detecting blind spots in official documents, conducting compliance reviews of contracts and agreements, and verifying the specifications of bidding documents; this invention does not specifically limit its application to these areas. For ease of explanation, the following embodiments are all illustrated using the application to detecting blind spots in official documents as an example.

[0061] It should be noted that all actions involving the acquisition of signal information or data in this invention are carried out in compliance with the relevant data protection laws and regulations of the country where the invention is located, and with the authorization granted by the owner of the relevant device.

[0062] in, Figure 1This is a flowchart illustrating the document review blind spot detection method provided by the present invention, as follows: Figure 1 As shown, the method includes steps 110, 120, 130 and 140.

[0063] Step 110: Divide the target document to be reviewed into semantic units, and based on the logical dependencies between the semantic units obtained from the division, construct a semantic topology that reflects the internal logical relationships of the target document.

[0064] Here, target documents can be understood as various text documents that require compliance review, such as official documents, internal company rules and regulations, contract texts, project application forms, etc. Target documents are used to represent the carriers of specific business information and regulatory requirements generated during the operation of the subject. For example, in administrative management scenarios, target documents are typically official documents such as notices, circulars, requests for instructions, and approvals; in business cooperation scenarios, target documents are typically procurement contracts, confidentiality agreements, etc.

[0065] As an optional implementation, the target document to be reviewed can be obtained through a document management system interface, OCR (Optical Character Recognition) scan results, or direct file upload. For example, a user can upload a PDF document titled "Notice on Carrying Out Annual A Work" to the official document review system, which can then be processed as the target document.

[0066] Because target documents typically have complex chapter structures and multiple logical nestings, traditional full-text search or paragraph matching alone is insufficient to accurately capture their deeper meaning. Considering that simple keyword matching or regular expressions struggle to identify synonyms, referential relationships, and implicit business logic constraints in natural language, this embodiment utilizes a pre-trained large language model to perform deep semantic parsing of the target document. Based on the parsing, according to the document's inherent logical structure, such as chapters, clauses, and paragraphs, the target document is broken down into semantic units with independent business meaning. A semantic unit can be understood as the smallest semantic fragment expressing a single fact, requirement, or norm. A semantic unit can be a complete clause, such as "Project budgets exceeding 1 million yuan must be approved by the General Manager's Office," or an independent conditional branch within a complex sentence, such as the conditional judgment fragment "If classified information is involved."

[0067] Furthermore, based on the semantic units obtained from the partitioning, the logical dependencies between them are identified. Logical dependencies refer to the associations between semantic units in business logic, including but not limited to hierarchical relationships, referential relationships, and causal relationships. Based on these identified logical dependencies, all semantic units are organized into a directed graph structure, i.e., a semantic topology. In this structure, nodes are semantic units, and edges are defined logical dependencies. This semantic topology reflects the internal logical network of the target document itself, revealing how the various clauses in the document support and constrain each other.

[0068] Step 120: Based on the semantic topology, traverse and map the pre-constructed regulatory knowledge graph to obtain the set of normative nodes that the target document should satisfy, and generate a theoretical review path based on the set of normative nodes.

[0069] Specifically, a regulatory knowledge graph refers to a pre-constructed structured knowledge base that includes internal rules and regulations as well as external laws and regulations. Nodes in this graph represent specific regulatory clauses, and edges represent relationships such as applicability, conflict, and citation between clauses. By constructing a regulatory knowledge graph, a unified structured representation of multi-source regulations is achieved.

[0070] Specifically, based on the semantic topology constructed in step 110, the nodes corresponding to each semantic unit in the semantic topology are traversed, and their semantic features are used to search and match in the regulatory knowledge graph. The aim is to find all normative nodes that are highly relevant to the current semantic unit in terms of business meaning and match its applicable scope, i.e., the set of normative nodes that the target document should satisfy. For example, when a semantic unit in an official document involves "travel expense reimbursement," it will automatically associate the relevant nodes on invoices in external laws and regulations, as well as the nodes on approval authority in internal rules and regulations, in the regulatory knowledge graph.

[0071] After determining the set of all the normative nodes that should be satisfied, these normative nodes are sorted and organized according to their logical dependencies in the regulatory knowledge graph, thereby generating an ordered sequence of normative nodes, which is the theoretical review path. This path represents the sequence of all normative requirements that a complete review system should verify for the target document under ideal conditions.

[0072] Step 130: Obtain the standard nodes that are actually triggered during the review process of the target document to form the actual review path.

[0073] Specifically, the actual triggered rule node refers to the rule node that is actually called, checked, or referenced in the review process.

[0074] As an optional implementation, this information can be obtained by embedding log points in the auditing system. For example, the rule node IDs called by the rule engine or the regulatory nodes selected by the auditors on the interface can be recorded, and these collected records can be traced back to the corresponding regulatory nodes in the regulatory knowledge graph through unique ID mapping or semantic matching.

[0075] By combining all the successfully mapped specification nodes according to their triggering order or logical relationship, the actual review path is formed, which objectively reflects the true scope covered by the current review operation.

[0076] Step 140: Calculate the difference between the theoretical review path and the actual review path, and determine the difference as the review blind spot of the target document.

[0077] Considering that focusing solely on existing review results often overlooks gaps in review coverage, leading to the omission of potential risks, this embodiment identifies review blind spots by comparing theoretical and actual review paths.

[0078] Specifically, the difference between the theoretical review path and the actual review path is calculated. This difference includes not only missing nodes but also broken links. Missing nodes refer to specification nodes that exist in the theoretical review path, should be executed, but do not appear in the actual review path. Broken links refer to logical relationships connecting two nodes in the theoretical review path that are not reflected in the actual review path; that is, although the start and end nodes may both exist, the necessary flow or dependency between them is not triggered or verified.

[0079] Ultimately, this difference set is identified as the blind spot in the review process of the target document. For example, if the theoretical review path includes node A, but the actual path does not contain this node, then this node is a blind spot. Similarly, if the theoretical review path requires passing through the "risk control review" node from "application" to "approval," but the actual path jumps directly from "application" to "approval," this break in the link also constitutes a blind spot. Optionally, these blind spots can be output in a structured manner to alert users of potential risks that have not been covered by the review process.

[0080] The document review blind spot detection method provided in this embodiment divides the target document to be reviewed into semantic units and constructs a semantic topology. This topology is then mapped onto a pre-constructed regulatory knowledge graph to generate a theoretical review path. This theoretical path is then compared with the actual review path triggered during the actual review process to calculate the difference set, thereby accurately locating review blind spots. Because this embodiment goes beyond simply detecting explicit errors, it identifies implicit risks such as rule mismatches, missing review paths, or semantic mapping failures through difference analysis between theoretical and actual review paths. This achieves an effective assessment of the completeness of document review coverage, avoiding compliance risks caused by incomplete review coverage.

[0081] Given that target documents are typically presented in unstructured natural language, simple text segmentation is insufficient for computer systems to understand their complex business logic and the constraints between clauses, thus making them unsuitable for subsequent automated review path analysis. To address this issue, this embodiment further proposes a topology construction scheme based on deep semantic understanding.

[0082] Specifically, based on the logical dependencies between the semantic units obtained from the partitioning, the semantic topology structure reflecting the internal logical relationships of the target document is constructed, including:

[0083] Semantic analysis is performed on each semantic unit to obtain the semantic features of each semantic unit;

[0084] Based on the semantic features of each semantic unit, the logical dependencies between semantic units are identified;

[0085] Each semantic unit is treated as a node, and directed edges are established between nodes based on logical dependencies to construct a semantic topology.

[0086] Considering that computers cannot directly process the deep semantic information in natural language text, and that the business substance of clauses cannot be accurately distinguished based solely on surface vocabulary, it is necessary to convert the segmented text into a high-dimensional feature representation that can be computed and compared by computers, namely semantic features. Here, semantic features can be understood as a numerical representation that characterizes the deep business connotation of semantic units, which is usually expressed as a high-dimensional semantic vector.

[0087] As an alternative implementation, natural language processing techniques, particularly pre-trained language models, can be used to embed each segmented semantic unit. Specifically, the text content of the semantic unit is input into the pre-trained model, and the hidden layer states output by the model are extracted as the semantic features of that semantic unit. These semantic features not only contain the literal information of the semantic unit but also, through the model's contextual understanding capabilities, imply the deeper meaning of the unit within a specific business context.

[0088] After obtaining the semantic features of each semantic unit, considering that the clauses within the target document are not isolated but often have strict logical control relationships, such as hierarchical subordination, clause references, and precondition constraints, it is necessary to identify the logical dependencies between semantic units based on the extracted deep semantic features to accurately reconstruct these relationships.

[0089] Here, logical dependency refers to the logical relevance between semantic units. As an optional implementation, the vector similarity between the semantic features of any two semantic units can be calculated, or a trained relation classification model can be used, taking the semantic features of two semantic units as input, to predict whether a specific logical dependency exists between them. For example, if the semantic features of semantic unit A indicate that it is a "general rule," and the semantic features of semantic unit B indicate that it is a "detailed rule," and the two are adjacent in text position or have an inclusion relationship, then a hierarchical relationship is identified between them; if the semantic features of semantic unit C contain a reference to semantic unit D, then a reference relationship is identified between them; if semantic unit E describes a precondition, and semantic unit F describes the result after the condition is met, then a causal sequential relationship is identified between them.

[0090] After identifying the specific logical dependencies, in order to perform path analysis on the review logic of the target document from a global perspective, it is necessary to organize the scattered units and relationships into a structured graph.

[0091] As an optional implementation, each semantic unit can be instantiated as a node in the graph. For each pair of semantic units with logical dependencies, a directed edge is established between their corresponding nodes. The direction of this directed edge represents the direction of logical flow or dependency. For example, for hierarchical relationships, a hierarchical edge is established from the superior semantic unit node to the subordinate semantic unit node; for reference relationships, an association edge is established from the referrer semantic unit node to the referenced semantic unit node; for causal and sequential relationships, a logical edge is established from the preceding semantic unit node to the following semantic unit node. Through the above process, a directed graph containing all semantic units and their interrelationships is finally formed, which is the semantic topology reflecting the internal logical relationships of the target document.

[0092] This embodiment transforms unstructured target documents into semantic topology structures containing deep semantic features and logical dependencies, thereby achieving structured modeling of the document's internal logic and solving the problem of lacking a structured representation of review paths and coverage relationships in related technologies.

[0093] In real-world document review scenarios, the business meaning of a clause is determined not only by its textual content but also by its position, hierarchy, and context within the document. Relying solely on textual content features may fail to distinguish between clauses with similar content but different levels of validity, or similar content but different logical positions, thus affecting the accuracy of semantic features. To more comprehensively and accurately represent each semantic unit, this embodiment proposes a semantic analysis scheme that integrates multi-dimensional features.

[0094] Specifically, the semantic analysis of each semantic unit described above yields the semantic features of each semantic unit, including:

[0095] Extract the text content features and structural attribute features of each semantic unit. The structural attribute features include the chapter level, paragraph position and preceding and succeeding relationships of the corresponding semantic unit in the target document.

[0096] Text content features and structural attribute features are fused and encoded to generate semantic features for corresponding semantic units.

[0097] Considering that the semantic expression of a document is multi-dimensional, containing both explicit textual information and implicit structural information, it is necessary to extract these two types of information separately in order to comprehensively capture the features of semantic units. Based on this, this embodiment extracts textual content features to represent explicit textual information and structural attribute features to represent implicit structural information.

[0098] Here, text content features refer to the lexical, syntactic, and semantic information inherent in the text itself within a semantic unit. Structural attribute features refer to the position and relational attributes of a semantic unit within the overall structure of the target document. Specifically, this includes the corresponding semantic unit's chapter level, paragraph position, and precedence / succession relationships within the target document. The chapter level characterizes whether the unit belongs to a chapter, section, or article; the paragraph position characterizes the absolute or relative position of the unit within the document; and the precedence / succession relationships characterize the unit's contextual adjacency within the text flow.

[0099] As an optional implementation, for text content features, pre-trained language models such as BERT can be used to encode the text of semantic units, and the output vector can be extracted as the text content feature vector. For structural attribute features, the metadata of each semantic unit can be obtained by parsing the layout analysis results of the document. For example, if a semantic unit is identified as "Article 3" under "Chapter 2 Financial Management", its chapter level feature is marked as "Level 2", and its preceding unit is recorded as "Article 2" and its following unit as "Article 4".

[0100] After extracting text content features and structural attribute features respectively, considering that single-dimensional features cannot fully reflect the business attributes and logical positioning of semantic units in complex documents, it is necessary to organically combine the two to form a unified, high-dimensional semantic representation, namely semantic features.

[0101] As an alternative implementation, feature concatenation or weighted fusion can be used for fusion encoding. For example, the text content feature vector and the structural attribute feature vector can be concatenated, or a fully connected neural network layer can be used to fuse and map the two, outputting a new high-dimensional vector, which is the semantic feature of the corresponding semantic unit.

[0102] This embodiment, by integrating text content features and structural attribute features, can generate more distinctive and expressive semantic features, effectively solving the problem that single text features are difficult to distinguish similar clauses in structured documents.

[0103] Given the diverse logical dependencies within the target document, and the varying degrees of guidance different aspects of constructing the review path, simply establishing a single type of connection edge would fail to accurately represent the document's complex hierarchical structure, legal references, and business flow logic, potentially affecting the accuracy of subsequent path generation. Therefore, this embodiment further subdivides the logical dependencies and establishes different types of directed edges accordingly.

[0104] Specifically, the logical dependencies mentioned above include hierarchical relationships, referential relationships, and causal relationships;

[0105] The steps for determining a directed edge include:

[0106] Given a hierarchical relationship between any two nodes, establish a hierarchical edge pointing from the higher node to the lower node.

[0107] If any two nodes have a reference relationship, establish an associated edge from the referencing node to the referenced node;

[0108] Given a causal relationship between any two nodes, establish a logical edge from the preceding node to the following node.

[0109] Given that documents typically organize content using a tree-like or hierarchical structure, this structure reflects the inclusion and being-included relationship of the specification. To recreate this architecture in the topology, this embodiment establishes a hierarchical edge from the parent node to the child node when any two nodes have a hierarchical relationship.

[0110] Here, hierarchical relationship refers to the relationship between two semantic units in document structure or business logic, where one is the whole and the other is the part, or the general and the specific. For example, "Chapter Two Financial Management" is the superior node of "Section One Reimbursement Process". As an optional implementation, when semantic unit A is identified as the superior node and semantic unit B as the subordinate node, a directed edge from node A to node B is established in the semantic topology graph and marked as a hierarchical edge. The establishment of hierarchical edges helps to ensure that the superior principle is reviewed first, and then the subordinate details are reviewed, when generating the review path.

[0111] Given that clauses often reference each other, these referencing relationships form a horizontal constraint network within the document. To capture these cross-level constraints, this embodiment establishes an association edge from the referencing node to the referenced node when any two nodes have a referencing relationship.

[0112] Here, a referencing relationship refers to the explicit or implicit reference of one semantic unit to another. As an optional implementation, when it is identified that the content of semantic unit C (the referrer) mentions semantic unit D (the referenced unit), a directed edge from node C to node D is created in the semantic topology graph and marked as an associated edge. The creation of associated edges helps to automatically jump to the referenced clause for joint verification during review, preventing the omission of associated constraints due to isolated review.

[0113] Given the strict business sequence or conditional triggering logic between certain terms, this relationship determines the legal path of business flow. To reflect this dynamic business flow logic, this embodiment establishes a logical edge from the preceding node to the following node when there is a causal relationship between any two nodes.

[0114] Here, causality refers to one semantic unit being a precondition or triggering cause for the occurrence of another semantic unit. As an optional embodiment, when semantic unit E (preceding node) is identified as a precondition for semantic unit F (subsequent node), a directed edge from node E to node F is established in the semantic topology graph and marked as a logical edge. The establishment of logical edges helps verify the compliance of the business operation sequence when generating the audit path, preventing reverse processes or missing preconditions.

[0115] This embodiment subdivides logical dependencies into three types: hierarchical, reference, and causal, and establishes hierarchical edges, association edges, and logical edges respectively. This enables the constructed semantic topology structure to not only reflect the static organizational structure of the document, but also its dynamic business logic and horizontal constraint relationships.

[0116] Considering that the semantic topology of the target document and the regulatory knowledge graph belong to two different information spaces—the semantic topology describes the target document to be reviewed, while the regulatory knowledge graph describes internal rules and regulations and external laws and regulations—an efficient and accurate mapping mechanism is needed to effectively correlate the two and identify the regulatory clauses that the target document must comply with. Based on this, this embodiment proposes a mapping and logical ranking scheme based on vector similarity.

[0117] Specifically, based on semantic topology, the above-mentioned process involves traversing and mapping within a pre-constructed regulatory knowledge graph to obtain a set of normative nodes that the target document should satisfy. A theoretical review path is then generated based on this set of normative nodes, including:

[0118] Calculate the similarity between the semantic vectors of each semantic unit in the semantic topology and the semantic vectors of each normative node in the regulatory knowledge graph;

[0119] Based on similarity, the normative nodes associated with each semantic unit are determined from the regulatory knowledge graph;

[0120] Based on the logical dependencies between the normative nodes in the regulatory knowledge graph, the associated normative nodes are sorted to generate a theoretical review path.

[0121] Given the diversity of natural language expressions, the wording in the target document and the wording in the regulations may differ literally but be semantically similar. For example, the document may use the word "business trip" while the regulations use "official travel"; or the document may use the word "procurement" while the regulations use "material purchase". To achieve semantic alignment across texts, this embodiment uses semantic vectors for calculation.

[0122] Here, semantic vectors refer to the numerical vectors obtained by mapping text to a high-dimensional space through a pre-trained language model. As an optional implementation, a unified semantic vectorization model can first be used to map each normative node in the regulatory knowledge graph to the same high-dimensional semantic vector space as the semantic units in the semantic topology. Then, each semantic unit in the semantic topology is traversed, and the similarity between its semantic vector and the semantic vectors of each normative node in the regulatory knowledge graph is calculated. This similarity can be cosine similarity.

[0123] After calculating the similarity, it is necessary to filter out the truly relevant legal provisions based on the similarity level to eliminate irrelevant noise interference and ensure the accuracy of the mapping results.

[0124] As an optional implementation, a similarity threshold can be set. For each semantic unit, all canonical nodes whose semantic vector similarity to that unit is greater than or equal to the similarity threshold are selected. These selected canonical nodes are considered to have a substantial business relationship with the semantic unit and together form the set of canonical nodes that the target document should meet. For example, for the semantic unit "procure office computers", canonical nodes such as "fixed asset procurement management regulations" and "office equipment configuration standards" might be selected.

[0125] After determining the set of normative nodes that should be satisfied, and considering the inherent logical dependencies between these normative nodes within the legal framework, this embodiment organizes these normative nodes in an orderly manner to generate a theoretical review path that conforms to legal logic and business processes.

[0126] As an optional implementation, the logical dependencies between nodes in the aforementioned set of regulatory nodes are first obtained from the regulatory knowledge graph. These dependencies may include hierarchical dependencies, reference dependencies, etc. Then, all nodes in the set of regulatory nodes are sorted according to these logical dependencies. The sorted, ordered sequence of nodes constitutes the theoretical review path.

[0127] This embodiment achieves accurate mapping between documents and regulations through semantic vector similarity calculation, and generates an ordered theoretical review path by utilizing the inherent logical relationship of the regulatory graph. This effectively solves the problem of lack of a unified perspective when reviewing multi-source regulations collaboratively, and ensures the comprehensiveness of the review basis and the coherence of the review logic.

[0128] As an optional implementation, the regulatory knowledge graph includes internal rules and regulations nodes and external laws and regulations nodes; a theoretical review path is generated based on the set of regulatory nodes, including:

[0129] External legal paths are generated based on the external legal and regulatory nodes involved in the set of normative nodes, and internal institutional paths are generated based on the internal rules and regulations nodes involved in the set of normative nodes.

[0130] By integrating external regulatory pathways with internal institutional pathways, a theoretical review pathway is generated.

[0131] Here, external legal and regulatory nodes refer to nodes derived from external normative documents such as laws, regulations, and rules; internal rules and regulations nodes refer to nodes derived from internal documents such as management methods and implementation details issued by the enterprise.

[0132] As an optional implementation, the selected set of normative nodes is first categorized. For external laws and regulations nodes in the set, they are sorted according to their legal validity level to generate external regulatory paths; for internal rules and regulations nodes in the set, they are sorted according to the level of rule formulation or business process logic to generate internal rule paths.

[0133] After generating two separate paths, they need to be merged into a complete and executable audit logic chain. As an optional implementation, the external regulatory path can be used as the basic framework, and the internal institutional path can be used as a specific implementation and supplement to the external regulations, embedded into the corresponding external regulatory nodes to form a nested or parallel fusion path.

[0134] To make it easier to understand, let's take the "travel expense reimbursement approval" scenario as an example:

[0135] When a semantic unit in the target document is identified as relating to the "travel expense reimbursement approval" business, the associated regulatory node is activated in the regulatory knowledge graph. Regarding external regulatory paths, nodes concerning the authenticity of accounting documents and the compliance of invoices in external laws and regulations are automatically associated and activated.

[0136] Regarding internal system pathways, the system automatically links and activates the general principles node concerning reimbursement processes in internal rules and regulations, as well as specific clause nodes concerning approval authority for different job levels, accommodation and transportation standards and limits, and reimbursement time limits.

[0137] Finally, the above nodes are merged and sorted to generate a theoretical review path.

[0138] This embodiment achieves a unified perspective on multi-source regulations by distinguishing and integrating the two paths of external regulations and internal systems. It effectively avoids the risk of partial compliance but overall non-compliance caused by the separation of internal and external regulations in the review process, and ensures the completeness and systematic nature of the theoretical review path.

[0139] Given that the normative nodes in the legal knowledge graph are not simply arranged linearly, but rather involve complex dependencies and application conflicts—for example, the execution of some provisions presupposes the satisfaction of another, or some provisions are mutually exclusive under certain conditions—ignoring these logical relationships could lead to logical paradoxes or execution deadlocks in the generated theoretical review paths. To ensure that the generated paths are both logically sound and executable, this embodiment further proposes a refined sorting scheme based on reference relationships and conflict exclusion.

[0140] Specifically, based on the logical dependencies between the regulatory nodes in the regulatory knowledge graph, the associated regulatory nodes are sorted to generate a theoretical review path, including:

[0141] Identify the reference relationships and conflict / exclusion edges between the associated regulatory nodes in the regulatory knowledge graph;

[0142] Based on the reference relationship edges, determine the execution order of each associated specification node;

[0143] Based on conflict-rejection edges, conflicting standard nodes are removed from the sorted standard nodes to generate a theoretical review path.

[0144] After determining the relevant set of specification nodes, it is necessary to clarify the relationship network between them. Here, a reference relationship edge indicates that a specification node is based on or points to another specification node in terms of content, indicating that the latter is the basis or supplement to the former; a conflict exclusion edge indicates that two specification nodes are mutually exclusive in terms of applicable conditions or processing methods, that is, they cannot be applied at the same time in the same business scenario.

[0145] As an optional implementation, the connection edge attributes between any two normative nodes within the aforementioned set of normative nodes are queried in the regulatory knowledge graph. If an edge is marked as a reference or basis, it is identified as a reference relationship edge; if an edge is marked as a conflict, exception, or exclusion, it is identified as a conflict exclusion edge.

[0146] After identifying the referencing edges, considering that the referenced clauses usually contain more fundamental definitions or higher-level principles, they should be reviewed before the clauses that reference them. Therefore, this embodiment determines the execution order of the associated specification nodes based on the referencing edges.

[0147] As an optional implementation, reference edges can be treated as directed dependency edges. By traversing all reference edges, a local dependency graph can be constructed. In this dependency graph, specification nodes that do not depend on other nodes are prioritized in the execution sequence. For example, if clause A references clause B, clause B should be placed before clause A in the review path to ensure that the review of A is supported by the conclusion of clause B. This method determines the execution order of all nodes with reference dependencies.

[0148] After determining the initial execution order, considering that mutually exclusive clauses may be mixed in the collection, it is necessary to make choices based on the specific attributes of the target document; otherwise, the theoretical review path will be contradictory.

[0149] As an optional implementation, it is possible to check whether there are node pairs connected by conflicting edges in the sorted node sequence. If so, based on the target document's metadata or manually preset priority rules, one node matching the applicable conditions is retained, and the other conflicting node is removed. For example, if the target document involves a project amount of 500,000 yuan, and the graph contains two mutually exclusive nodes: "Public bidding applies to projects over 1 million yuan" and "Invitation to tender applies to projects under 1 million yuan," then the "Invitation to tender" node is retained based on the amount attribute, and the "Public bidding" node is removed.

[0150] After the above sorting and elimination operations, the final ordered sequence of nodes is generated as a logically coherent and conflict-free theoretical review path.

[0151] This embodiment achieves refined arrangement of theoretical review paths by deeply analyzing the reference dependencies and conflict exclusion relationships between standard nodes, avoiding logical deadlocks and contradictions in the paths, and ensuring that the generated review standards not only conform to the inherent logic of the regulatory system, but also accurately adapt to the specific business scenarios of the target documents.

[0152] Considering that in the actual document review process, in addition to routine compliance checks, a special situation may arise where the document content involves entirely new business areas or special scenarios, and the existing regulatory knowledge graph has not yet included the corresponding regulatory nodes, or although relevant regulations exist, they have not been mapped due to significant semantic differences. If this situation is ignored, it will lead to gaps in compliance supervision coverage, posing a significant risk. To capture these special blind spots, this embodiment further proposes a blind spot labeling scheme based on mapping anomalies.

[0153] Specifically, the process of traversing and mapping within a pre-constructed regulatory knowledge graph based on semantic topology also includes:

[0154] If any semantic unit in the semantic topology does not have a corresponding normative node in the regulatory knowledge graph, a missing mapping record is generated based on the semantic features of the semantic unit.

[0155] Mark the semantic units corresponding to the missing mapping records as missing mapping blind spots.

[0156] Considering that when the traversal algorithm cannot find a sufficiently similar regulatory node to a certain semantic unit in the regulatory knowledge graph, it means that the content described by that semantic unit may be outside the coverage of the current regulatory knowledge graph. In this case, ignoring it directly could lead to potential compliance risks going undetected due to a lack of review basis. Therefore, this embodiment first generates a mapping missing record based on the semantic features of any semantic unit. This mapping missing record can be understood as a log entry recording semantic units and their feature information that are not covered by the regulatory knowledge graph.

[0157] As an optional implementation, if any semantic unit in the semantic topology fails to match any normative node with a similarity greater than the similarity threshold in the regulatory knowledge graph, it is determined that the semantic unit has a mapping anomaly. In this case, the semantic features of the semantic unit are extracted and combined with its unique unit ID to generate a mapping missing record.

[0158] After generating the mapping missing record, in order to explicitly indicate this risk in the final audit report, it is necessary to classify and mark it. Specifically, in this embodiment, the semantic unit corresponding to the mapping missing record is marked as a mapping missing blind spot.

[0159] Among these, mapping gaps reveal two potential problems: first, entirely new business situations have emerged that are not covered by existing internal or external regulations; second, external regulations have been updated, but the regulatory knowledge graph has not been synchronized in a timely manner, making it impossible to establish a mapping. In such cases, these mapping gaps can be prominently displayed to prompt relevant personnel to conduct manual review or initiate the process of supplementing regulations.

[0160] Considering that the theoretical audit path is merely a standard expected to be implemented, the key to identifying blind spots lies in accurately obtaining the actual audit status. However, current audit systems typically generate discrete, unstructured operational traces, which often lack a direct correspondence with the normative nodes in the regulatory knowledge graph. To align and compare actual operations with theoretical standards, this embodiment proposes a path construction scheme based on mapping backtracking.

[0161] Specifically, the above-mentioned acquisition of the specification nodes actually triggered during the review process of the target document forms the actual review path, including:

[0162] Map the normative nodes actually triggered during the review process of the target document to the corresponding normative nodes in the legal knowledge graph;

[0163] Based on the logical dependencies between the normative nodes mapped in the regulatory knowledge graph, the mapped normative nodes are organized to form the actual review path.

[0164] Considering the diversity of actual review environments, review records may manifest as rule IDs, text descriptions of inspection items, or even manually entered remarks. This information needs to be uniformly mapped to the coordinate system of the regulatory knowledge graph to achieve data dimension consistency and measurement standard alignment. Therefore, this embodiment maps the regulatory nodes actually triggered during the review process of the target document to the corresponding regulatory nodes in the regulatory knowledge graph.

[0165] Here, the actual triggered regulatory nodes can be understood as the specific compliance basis that is substantially invoked or referenced during the review and execution phase. As an optional implementation, data collection probes can be pre-embedded in the review system. When the rule engine hits a rule, it queries the corresponding legal clause for that rule ID through a pre-defined mapping table, thereby locating the specific regulatory node in the legal knowledge graph. For manual review operations, natural language processing technology is used to semantically match the selected text and map it to the corresponding regulatory node in the graph. In this way, all scattered review traces are aggregated into a set of activated associated nodes on the legal knowledge graph, i.e., the mapped regulatory nodes.

[0166] After completing the node mapping, considering that the mapped standard nodes are still scattered and disordered, in order to truly restore the full logical flow of the review process, this embodiment organizes the mapped standard nodes according to the logical dependencies between the mapped standard nodes in the regulatory knowledge graph to form the actual review path.

[0167] As an optional implementation, the logical dependencies between these mapped normative nodes in the regulatory knowledge graph can be read. Then, based on these logical dependencies, the mapped normative nodes are connected and ordered to form the actual review path. This actual review path not only includes the points being reviewed, but also implicitly contains the review order and logical structure.

[0168] This embodiment solves the problem that traditional audit records are difficult to compare with compliance standards in a structured way by uniformly mapping heterogeneous actual audit records to a regulatory knowledge graph and reconstructing them into a structured actual audit path based on the graph logic.

[0169] Considering that the calculated difference set is merely a dataset, without in-depth classification and analysis, users will find it difficult to intuitively understand the specific causes of risks. Different types of blind spots correspond to different risk levels and remediation strategies. To provide more instructive blind spot diagnostic results, this embodiment further proposes a blind spot classification and determination scheme based on difference features.

[0170] Specifically, the above-mentioned identification of the difference set as a blind spot for reviewing the target document includes:

[0171] If any specification node contained in the difference set exists in the theoretical review path but not in the actual review path, then any specification node is identified as a node missing blind spot.

[0172] If both the starting and ending nodes in the theoretical review path exist in the actual review path, and the intermediate node connecting the starting and ending nodes is missing in the actual review path, then the intermediate node is identified as a path interruption blind spot.

[0173] Specifically, if any normative node contained in the difference set exists in the theoretical audit path but not in the actual audit path, it indicates that the compliance requirement represented by that normative node was not fulfilled during the actual audit process. In this case, the normative node is considered a node-missing blind spot. Here, a node-missing blind spot refers to normative nodes that are theoretically required to be executed but leave no trace in actual operation.

[0174] As an optional implementation, for any specification node in the difference set, it is checked whether it belongs to the node set of the theoretical review path, and simultaneously checked whether it does not belong to the node set of the actual review path. If this condition is met, the node is determined to be a node-missing blind spot. For example, if the theoretical path requires verification of "whether there is a legal counsel's signature," but this item is not present in the actual path, then the "legal counsel's signature" node is a node-missing blind spot.

[0175] Besides the lack of point-to-point checks, considering that the review process is often a workflow, if only the start and end nodes are checked while key intermediate steps are skipped, although the process appears closed-loop, the logical chain of the process is actually broken. To identify this hidden risk of workflow breakage, this embodiment also determines whether the start and end nodes in the theoretical review path both exist in the actual review path, and whether the intermediate nodes connecting the start and end nodes are missing in the actual review path.

[0176] If both the starting and ending nodes in the theoretical audit path exist in the actual audit path, and the intermediate node connecting the starting and ending nodes is missing in the actual audit path, it indicates that the audit process lacks necessary logical coherence and there is a risk of skipping audits or missing key control points. In this case, the intermediate node is regarded as a path interruption blind spot.

[0177] Here, an intermediate node refers to a standardized node located between the start and end nodes in the theoretical review path, and having a direct or indirect logical connection with both the start and end nodes. A path-disruption blind spot refers to a situation where, in a complete logical chain, although both ends are covered, key nodes that connect and transition in the middle are omitted. As an optional implementation, any pair of start and end nodes can be selected in the theoretical review path, and it can be checked whether they also exist in the actual review path. If both exist, the intermediate nodes connecting these two nodes in the theoretical path are further examined. If these intermediate nodes exist in the aforementioned calculated difference set, then these intermediate nodes are determined to constitute a path-disruption blind spot.

[0178] Considering that not all blind spots are equally dangerous—some may involve core legal red lines, while others may only be minor procedural flaws—simply providing a straightforward list of blind spots often makes it difficult for users to grasp the key points and efficiently manage risks. To help users quickly identify high-risk situations and provide actionable rectification guidelines, this embodiment further proposes a risk quantification and intelligent suggestion generation scheme based on graph attributes.

[0179] Specifically, after calculating the difference between the theoretical review path and the actual review path, the above also includes:

[0180] Based on the centrality of the normative nodes contained in the difference set in the regulatory knowledge graph, the risk weight of the review blind spot is calculated, and the risk level of the review blind spot is determined based on the risk weight.

[0181] Based on the clause content corresponding to the specification nodes contained in the difference set, modification suggestions are generated for the target document.

[0182] Given that different nodes in a regulatory knowledge graph have varying degrees of connectivity, a node with higher centrality usually means that it is referenced or relied upon by more provisions, and its position in the regulatory system is more central. Once it is missing, the chain reaction and compliance risks will be greater.

[0183] Here, centrality can be understood as a metric for the importance of a node in the network, specifically measured by degree centrality or betweenness centrality. As an optional implementation, the centrality of each regulatory node within the difference set can be obtained by first querying the regulatory knowledge graph. Then, using a pre-defined risk calculation model, the centrality is weighted and summed with other auxiliary factors (such as the node's effectiveness level) to calculate the risk weight of the blind spot. Finally, based on a pre-defined grading threshold, the risk weight is mapped to different risk levels, such as high-risk, medium-risk, and low-risk.

[0184] After clarifying the risk level, and considering that the large language model has powerful text generation capabilities, it can be used to generate targeted modification suggestions by combining the clause content corresponding to the normative nodes contained in the difference set.

[0185] Here, the modification suggestions are operational texts used to guide users on how to supplement missing content or correct errors in the process. As an optional implementation, the specific clause content of each specification node in the difference set can be extracted and combined with the context of the target document to construct prompt words for input into the large language model. Based on this information, the model generates guided completion suggestions.

[0186] Figure 2 This is a schematic diagram of the overall architecture of the document review blind spot detection system provided by the present invention, as shown below. Figure 2 As shown, the architecture mainly consists of a data input layer, a processing and computation core layer, and a result output layer. The specific processing flow is as follows:

[0187] First, at the data input layer, the target document to be reviewed is acquired, and external legal and regulatory databases and internal rules and regulations databases are used as sources of standards. Next, at the core processing and computation layer, natural language processing technology is used to decompose and vectorize the target document into semantic units, transforming it into semantic features that computers can understand. Simultaneously, the standard clauses in the external legal and regulatory databases and internal rules and regulations databases are structured and vectorized, converting unstructured legal text into structured data nodes, ensuring that both exist in a unified semantic space.

[0188] Then, in the dual-driven knowledge core layer, based on the decomposed semantic units and their logical dependencies, a semantic topology structure reflecting the internal logical network of the target document is constructed; at the same time, based on the structured normative clauses and their logical dependencies, a legal knowledge graph is constructed.

[0189] Next, the blind spot detection engine is activated. This engine consists of three key components: First, it performs traversal mapping in the legal knowledge graph based on semantic topology, performs theoretical review path derivation, and generates theoretical review paths; second, it obtains the rules or clauses triggered by the target document during the actual review process, performs actual coverage path capture, and obtains the actual review path; third, it performs topological differential analysis to calculate the difference between the theoretical review path and the actual review path, thereby accurately locating the review blind spot.

[0190] Finally, at the results output layer, based on the results of topology differential analysis, a coverage quantification score is output to assess the completeness of the audit, and blind spot risk classification prompts are provided according to the node attributes in the differential set. At the same time, modification suggestions are generated by combining the specific clause content of the nodes in the differential set, thus forming a complete intelligent closed loop from blind spot detection to auxiliary correction.

[0191] Figure 3This is a flowchart illustrating another document review blind spot detection method provided by the present invention, as shown below. Figure 3 As shown, the process includes: first, obtaining the semantic units of the target document to be reviewed through semantic analysis, and locating the coordinates of each semantic unit in the constructed semantic topology.

[0192] Then, based on the coordinates of the located semantic units, the corresponding normative nodes are determined in the pre-constructed regulatory knowledge graph. By traversing the mapping, the set of normative nodes that the target document should satisfy is found. On this basis, the normative nodes in the set of normative nodes are sorted and organized to generate a theoretical review path. This path includes all internal normative and external legal and regulatory nodes that the target document should be bound by.

[0193] Next, by collecting logs or backtracking rules, actual audit records are mapped to form the actual audit path. Then, topological difference operations are performed to calculate the difference between the theoretical path and the actual path.

[0194] Finally, it is determined whether the difference set is empty. If it is, it is considered to have full coverage without blind spots, and the process ends; otherwise, it proceeds to blind spot classification and risk assessment. Specifically, if the difference set indicates a missing external regulatory node, it is classified as a node-missing blind spot; if it indicates a path interruption, it is classified as a path-interrupted blind spot; if it indicates the absence of any regulatory mapping, it is classified as a mapping-missing blind spot. These identified blind spots are then summarized to generate a structured blind spot report and completion suggestions, providing users with specific rectification guidance, thus concluding the process.

[0195] The document review blind spot detection device provided by the present invention is described below. The document review blind spot detection device described below can be referred to in correspondence with the document review blind spot detection method described above.

[0196] Based on any of the above embodiments Figure 4 This is a schematic diagram of the document review blind spot detection device provided by the present invention, as shown below. Figure 4 As shown, the device includes:

[0197] The segmentation module 410 is used to segment the target document to be reviewed into semantic units, and to construct a semantic topology structure that reflects the internal logical relationship of the target document based on the logical dependency relationship between the segmented semantic units.

[0198] The generation module 420 is used to traverse and map the pre-built regulatory knowledge graph based on the semantic topology to obtain the set of normative nodes that the target document should satisfy, and generate a theoretical review path based on the set of normative nodes.

[0199] The acquisition module 430 is used to acquire the standard nodes actually triggered during the review process of the target document, forming the actual review path;

[0200] The detection module 440 is used to calculate the difference between the theoretical review path and the actual review path, and to determine the difference as the review blind spot of the target document.

[0201] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540. The processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions from the memory 530 to execute a document review blind spot detection method.

[0202] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part 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 the present invention. 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.

[0203] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the document review blind spot detection method provided by the above methods.

[0204] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the document review blind spot detection method provided by the above methods.

[0205] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0206] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0207] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications 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 the present invention.

Claims

1. A method for detecting blind spots in document review, characterized in that, include: The target document to be reviewed is divided into semantic units, and based on the logical dependencies between the semantic units obtained from the division, a semantic topology structure reflecting the internal logical relationships of the target document is constructed. Calculate the similarity between the semantic vector of each semantic unit in the semantic topology and the semantic vector of each normative node in the regulatory knowledge graph; determine the normative nodes associated with each semantic unit from the regulatory knowledge graph based on the similarity; sort the associated normative nodes according to the logical dependencies between the normative nodes in the regulatory knowledge graph to generate a theoretical review path. Map the normative nodes actually triggered during the review process of the target document to the corresponding normative nodes in the legal knowledge graph; Based on the logical dependencies between the normative nodes mapped in the aforementioned regulatory knowledge graph, the mapped normative nodes are organized to form the actual review path; Calculate the difference between the theoretical review path and the actual review path, and determine the difference as the review blind spot of the target document.

2. The document review blind spot detection method according to claim 1, characterized in that, The construction of a semantic topology structure reflecting the internal logical relationships of the target document, based on the logical dependencies between the semantic units obtained from the partitioning, includes: Semantic analysis is performed on each semantic unit to obtain the semantic features of each semantic unit; Based on the semantic features of each semantic unit, the logical dependencies between semantic units are identified; Each semantic unit is treated as a node, and directed edges are established between the nodes based on the logical dependencies to construct the semantic topology.

3. The document review blind spot detection method according to claim 2, characterized in that, The semantic analysis of each semantic unit to obtain the semantic features of each semantic unit includes: Extract the text content features and structural attribute features of each semantic unit. The structural attribute features include the chapter level, paragraph position and preceding and succeeding relationships of the corresponding semantic unit in the target document. The text content features and the structural attribute features are fused and encoded to generate semantic features for the corresponding semantic units.

4. The document review blind spot detection method according to claim 2, characterized in that, The logical dependencies include hierarchical relationships, referential relationships, and causal relationships; The steps for determining the directed edges include: If the above hierarchical relationship exists between any two nodes, establish a hierarchical edge from the upper node to the lower node; If the aforementioned reference relationship exists between any two nodes, establish an associated edge from the referencing node to the referenced node; If the causal relationship exists between any two nodes, establish a logical edge from the previous node to the next node.

5. The document review blind spot detection method according to any one of claims 1 to 4, characterized in that, The step of sorting the associated regulatory nodes according to the logical dependencies between the regulatory nodes in the regulatory knowledge graph to generate the theoretical review path includes: Identify the reference relationships and conflict / exclusion edges of the associated normative nodes in the legal knowledge graph; Based on the reference relationship edges, the execution order of each associated specification node is determined; Based on the conflict exclusion edge, conflicting standard nodes are removed from the sorted standard nodes to generate the theoretical review path.

6. The document review blind spot detection method according to any one of claims 1 to 4, characterized in that, The process of traversing and mapping within a pre-constructed regulatory knowledge graph based on the aforementioned semantic topology also includes: If any semantic unit in the semantic topology does not have a corresponding normative node in the regulatory knowledge graph, a missing mapping record is generated based on the semantic features of the semantic unit. The semantic units corresponding to the missing mapping records are marked as missing mapping blind spots.

7. The document review blind spot detection method according to any one of claims 1 to 4, characterized in that, The step of identifying the difference set as the blind spot for review of the target document includes: If any specification node contained in the difference set exists in the theoretical review path but not in the actual review path, then the specified specification node is determined to be a node missing blind spot. If both the starting node and the ending node in the theoretical review path exist in the actual review path, and the intermediate node connecting the starting node and the ending node is missing in the actual review path, then the intermediate node is determined to be a path interruption blind spot.

8. The document review blind spot detection method according to any one of claims 1 to 4, characterized in that, After calculating the difference between the theoretical review path and the actual review path, the method further includes: Based on the centrality of the normative nodes contained in the difference set in the regulatory knowledge graph, the risk weight of the audit blind spot is calculated, and the risk level of the audit blind spot is determined based on the risk weight. Based on the clause content corresponding to the specification nodes contained in the difference set, modification suggestions are generated for the target document.

9. A document review blind spot detection device, characterized in that, include: The segmentation module is used to segment the target document to be reviewed into semantic units, and based on the logical dependencies between the segmented semantic units, to construct a semantic topology that reflects the internal logical relationships of the target document. The generation module is used to calculate the similarity between the semantic vector of each semantic unit in the semantic topology and the semantic vector of each normative node in the regulatory knowledge graph; based on the similarity, determine the normative nodes associated with each semantic unit from the regulatory knowledge graph; and sort the associated normative nodes according to the logical dependencies between the normative nodes in the regulatory knowledge graph to generate a theoretical review path. The acquisition module is used to map the normative nodes actually triggered by the target document during the review process to the corresponding normative nodes in the legal knowledge graph; Based on the logical dependencies between the normative nodes mapped in the aforementioned regulatory knowledge graph, the mapped normative nodes are organized to form the actual review path; The detection module is used to calculate the difference between the theoretical review path and the actual review path, and to determine the difference as the review blind spot of the target document.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the document review blind spot detection method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the document review blind spot detection method as described in any one of claims 1 to 8.

Images