Incremental normative document fair competition review method based on large language model

Abstract

The application discloses an incremental normative file fair competition review method based on a large language model, relates to the technical field of natural language processing, and comprises the following steps: obtaining a current version and a previous version of a to-be-reviewed normative file, constructing hierarchical text units, and performing version alignment to identify version change content and form version difference data; based on the version difference data, constructing a review semantic feature set by using a large language model to form review semantic change data; determining a semantic change level according to the review semantic change data, generating a semantic change event, updating a review knowledge graph, and forming a change knowledge node set; and performing associated influence analysis according to the change knowledge node set to form a candidate influence node set. The application realizes accurate influence analysis and incremental fair competition review in the version updating process of normative files.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an incremental method for fair competition review of normative documents based on a large language model. Background Technology

[0002] With the development of artificial intelligence technology, large language models, with their semantic understanding, knowledge reasoning and text analysis capabilities, are gradually being applied to the field of intelligent review of normative documents. Existing technologies use large language models to perform semantic analysis on normative documents, identify the subjects, behaviors, restrictions and clause relationships in the documents, and combine them with review rules and knowledge bases to achieve intelligent review and improve the information processing capabilities of fair competition review.

[0003] Existing normative document review technologies based on large language models typically require re-analysis of the entire document content in version update scenarios, making it difficult to fully utilize historical review conclusions and related information. When the content of a normative document changes, existing technologies struggle to accurately determine the scope of impact of the changes on historical review conclusions, leading to duplicate reviews. A method is needed that can determine the scope of impact based on the content changes in the version of the normative document and conduct incremental fair competition reviews for the affected areas. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an incremental method for fair competition review of normative documents based on a large language model to solve the problems of inaccurate determination of the review scope and duplicate review during the version update process of normative documents.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an incremental method for fair competition review of normative documents based on a large language model, which includes obtaining the current version and the previous version of the normative document to be reviewed, constructing hierarchical text units, and performing version alignment to identify version change content to form version difference data. Based on version difference data, a set of review semantic features is constructed using a large language model to form review semantic change data; Based on the semantic change data reviewed, the semantic change level is determined, semantic change events are generated, and the review knowledge graph is updated to form a set of change knowledge nodes; Based on the set of changed knowledge nodes, perform correlation impact analysis to form a set of candidate impact nodes. Then, match the set of candidate impact nodes with the historical review conclusion dependency slices to determine the minimum recalculation domain. Based on the minimum recalculation domain, obtain the corresponding review evidence, use the large language model to perform local incremental review, and generate an updated review conclusion. The updated review conclusions are combined with the unaffected historical review conclusions to obtain the current version of the review results.

[0007] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the construction of hierarchical text units includes: The text content of the current and previous versions of the normative documents to be reviewed is extracted, the titles, chapters, clauses, provisions and attachments in the text content are identified, and the structural relationships between the titles, chapters, clauses, provisions and attachments are analyzed to form structural content; The structure and content are analyzed to determine the hierarchical relationship between titles, chapters, clauses, items and attachments, and the corresponding text content is divided according to the hierarchical relationship to form hierarchical text fragments; Each hierarchical text fragment is assigned a unique text unit identifier according to its hierarchical relationship, and then organized according to the hierarchical relationship to form hierarchical text units.

[0008] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the version difference data includes: Extract text unit identifiers, hierarchical paths, text content, and referencing relationships from hierarchical text units to form text unit features; Based on the characteristics of text units, establish the correspondence between the hierarchical text units corresponding to the current version and the hierarchical text units corresponding to the previous version, thus forming a text unit correspondence relationship; Extract the change information between corresponding text units based on the correspondence between text units to form version change content; The version changes are associated with the corresponding text unit identifiers, hierarchical paths, and text content to form version difference data.

[0009] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the reviewed semantic change data includes: The text before and after the changes, text unit identifiers, hierarchical paths and reference relationships in the version difference data are organized to form version difference semantic input data. The version difference semantic input data is combined with the preset review semantic feature types, review semantic parsing instructions and review semantic output formats to construct a large language model review semantic parsing task; Input the big language model into the semantic parsing task of the big language model, and use the big language model to perform contextual semantic parsing on the text before and after the change, forming a set of censorship semantic features composed of subject features, behavioral features, constraint features and association features; Establish a correspondence between the review semantic features corresponding to the text before the change and the review semantic features corresponding to the text after the change, and form a review semantic feature correspondence relationship; Based on the correspondence of semantic features in the review, the changes in subject features, behavioral features, constraint features, and related features are extracted to form review semantic change data.

[0010] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the semantic change events include: Based on the review semantic change data, review semantic features with mutual constraints are established into semantic constraint chains according to subject, behavior, target, applicable conditions, exception conditions, definition dependency and reference relationship, thus forming review semantic association relationship; Based on the semantic association relationship, calculate the semantic change range of subject features, behavioral features, constraint features and association features at each constraint transmission position along the semantic constraint chain; The position where the scope of the semantic change is extended from the current review semantic feature to the associated review semantic feature is determined as the constraint transmission termination position, and the semantic change hierarchy is formed based on the semantic constraint chain depth of the constraint transmission termination position; Based on the semantic change level, review semantic change data with the same semantic change level and sharing semantic constraint chain are aggregated, and the change boundary is determined according to the continuity of the semantic constraint chain to form semantic change unit. Based on the semantic change unit, the corresponding change propagation path is extracted along the semantic constraint chain, and semantic change events are generated based on the change propagation path.

[0011] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the set of changed knowledge nodes includes: Based on semantic change events, determine the corresponding review semantic features, change direction, change scope and correlation, and form updated content for the knowledge graph; Based on the updated content of the knowledge graph, locate and examine the corresponding knowledge nodes and relationship edges in the knowledge graph to form a set of knowledge nodes to be updated; Based on the set of knowledge nodes to be updated and the updated content of the knowledge graph, the knowledge nodes and relation edges in the review knowledge graph are added, replaced, invalidated or associated and adjusted to form the updated review knowledge graph. Based on the updated review knowledge graph, knowledge nodes that are directly updated by semantic change events or affected by associations are extracted to form a set of changed knowledge nodes.

[0012] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the candidate influential node set includes: Based on the set of changed knowledge nodes, extract the relationship type, relationship direction and association level corresponding to each changed knowledge node to form the association features of changed knowledge nodes; Based on the characteristics of the knowledge nodes that have changed, the impact of the changes is transmitted step by step along the definition relationships, reference relationships, constraint relationships and dependency relationships in the knowledge graph under review, forming a propagation relationship of the impact of changes. Based on the propagation relationship of change impacts, related knowledge nodes that share the same change knowledge node and are located on the continuous change impact propagation path are grouped together to form an impact knowledge node set. In addition, based on the change impact propagation relationship, impact knowledge nodes that have direct or indirect review dependencies with the change knowledge node are retained to form a candidate impact node set.

[0013] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, wherein: determining the minimum recalculation domain includes, Extract the corresponding node identifier, node type, association path and scope of influence from the candidate impact node set, and associate them with the rule nodes, fact nodes, evidence nodes and association paths in the historical review conclusion dependency slice to form candidate impact node association data; Based on the candidate impact node association data, the dependency relationship between the candidate impact node and the historical review conclusion is analyzed along the dependency path in the historical review conclusion dependency slice, and the historical review conclusion dependency slice corresponding to the candidate impact node is determined to form the conclusion impact association relationship. Based on the impact of the conclusions on the relationships, historical review conclusions that have direct dependencies or indirect dependencies formed through the relationship path are collected to form a set of conclusions to be recalculated. Based on the set of conclusions to be recalculated, extract the text units, rule nodes, fact nodes, and evidence nodes from the corresponding historical review conclusion dependency slices, and combine the text units, rule nodes, fact nodes, and evidence nodes into the smallest recalculation domain; The scope of text, rules, and evidence that need to be re-examined is determined based on the minimum recalculation domain, and the scope of evidence to be obtained for the review is obtained.

[0014] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the updated review conclusion includes: Based on the minimum recalculation domain, extract the corresponding text units, rule nodes, fact nodes, and evidence nodes, determine the scope of review evidence corresponding to the minimum recalculation domain, and form the scope of review evidence; Based on the scope of the evidence to be reviewed, obtain the corresponding definitions, references, rules, and facts, and associate them to form an incremental set of evidence for review. Input the minimum recalculation domain, the incremental review evidence set, and the corresponding review semantic change data into the large language model to construct a local incremental review task. Using a large language model, we conduct correlation analysis on the minimum recalculation domain, the incremental review evidence set, and the review semantic change data in the local incremental review task to determine the correspondence between the updated review facts and review rules, and form an updated review fact set. The updated set of facts for review is used to generate the corresponding fair competition review results, and the fair competition review results are associated with the corresponding text units, rule nodes, fact nodes and evidence nodes to form the updated review conclusions.

[0015] As a preferred embodiment of the incremental normative document fair competition review method based on a large language model described in this invention, the current version review results include: Extract the corresponding text units, rule nodes, fact nodes, and evidence nodes from the updated review conclusions and the unaffected historical review conclusions to form review conclusion-related data; The updated review conclusions are linked with the unaffected historical review conclusions according to the correspondence between text units, rule nodes, fact nodes, and evidence nodes to form a review conclusion fusion relationship; The changes in the updated review conclusions are combined with the unchanged content in the unaffected historical review conclusions to form the current version of the review conclusions set; Based on the current version of the review conclusion set, the corresponding text units, rule nodes, fact nodes, and evidence nodes are associated and saved to form the current version of the review result.

[0016] The beneficial effects of this invention are as follows: By acquiring the current and previous versions of normative documents and constructing hierarchical text units, version content change identification is achieved, forming version difference data; semantic parsing of the version difference data is performed using a large language model to extract subject features, behavioral features, constraint features, and association features, generating review semantic change data; the scope of semantic change is analyzed based on semantic constraint chains, the semantic change level is determined, and semantic change events are generated, updating the review knowledge graph to form a set of change knowledge nodes; the propagation relationship of change impact is analyzed based on the set of change knowledge nodes, and the minimum recalculation domain is determined by combining historical review conclusions with dependencies; corresponding review evidence is obtained based on the minimum recalculation domain, and local incremental review is carried out using a large language model to generate updated review conclusions; the updated review conclusions are fused with unaffected historical review conclusions to obtain the current version review results, achieving accurate impact analysis and incremental fair competition review during the normative document version update process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an incremental method for reviewing fair competition in normative documents based on a large language model. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Reference Figure 1This is one embodiment of the present invention, which provides a method for fair competition review of incremental normative documents based on a large language model, including the following steps: S1. Obtain the current and previous versions of the normative documents to be reviewed, construct hierarchical text units, and perform version alignment to identify version changes and generate version difference data.

[0023] S1.1 Extract the text content of the current and previous versions of the normative documents to be reviewed, identify the titles, chapters, clauses, provisions and attachments in the text content, and analyze the structural relationships between the titles, chapters, clauses, provisions and attachments to form structural content.

[0024] Furthermore, the current and previous versions of the normative documents under review are obtained electronically. The electronic text is then parsed to extract headings, chapter identifiers, clause numbers, item numbers, and attachments. The organizational relationships between different text contents are identified based on the text's order, numbering, and semantic connections. For normative documents containing chapter 1, article 1, and paragraph 1, the relationship between headings and chapters is determined by identifying the correspondence between chapter names and clause numbers. The hierarchical relationship between clauses and items is determined by identifying the association between clause numbers and item numbers. For attachments, the relationship between attachments and the main text is determined based on attachment names, citation locations, and related descriptions. For text content not using standard numbering formats, structural relationships are analyzed using text position relationships, contextual connections, and semantic integrity to identify the text's hierarchy, avoiding inaccurate extraction of text structure due to document format differences. By analyzing the structural relationships between headings, chapters, clauses, items, and attachments, continuous text content is converted into structured content with organizational relationships.

[0025] S1.2 Analyze the hierarchical relationship of the structural content, determine the hierarchical correspondence between titles, chapters, clauses, items and attachments, and divide the corresponding text content according to the hierarchical relationship to form hierarchical text fragments.

[0026] Furthermore, based on the structural relationships between titles, chapters, clauses, clauses, and appendices within the content structure, the inclusion relationships between different text contents are analyzed to determine the hierarchical position of different text contents. The text content within the corresponding scope is then divided according to the hierarchical position. For example, when a title corresponds to multiple chapter contents, the title is considered the upper-level structure, and the corresponding chapter contents are considered the lower-level structure; when a chapter corresponds to multiple clause contents, the chapter is considered the upper-level structure, and the corresponding clause contents are placed within the corresponding chapter scope; when a clause contains multiple clause contents, the clause contents are placed within the corresponding clause scope. In cases where clauses reference other chapters, appendices, or definitions, the association information between texts is preserved based on the reference relationship. This ensures that the divided hierarchical text fragments not only contain the corresponding text content but also retain the text's hierarchy and associated position. By dividing the content structure according to hierarchical relationships, each part of the text content corresponds to a clear hierarchical scope, forming hierarchically organized hierarchical text fragments.

[0027] S1.3 Assign a unique text unit identifier to the hierarchical text fragments according to the hierarchical relationship, and organize them according to the hierarchical relationship to form hierarchical text units.

[0028] Furthermore, based on the corresponding title, chapter, clause, item, and attachment positions of the hierarchical text fragments, a unique text unit identifier is generated for each hierarchical text fragment. The text unit identifier is then associated with and saved in relation to the corresponding hierarchical path, text content, and relationship. Even if the text content of different clauses in the same normative document is similar, different text unit identifiers are assigned according to their chapter positions and hierarchical relationships to ensure accurate matching between different versions based on the text unit identifiers. Based on the hierarchical correspondence between titles, chapters, clauses, items, and attachments, text units with inclusion relationships are organized to maintain the association between upper-level and lower-level text units, ensuring that each text unit can be located at a specific structural position in the normative document. By uniquely identifying and hierarchically organizing the hierarchical text fragments, hierarchical text units are formed.

[0029] S1.4 Extract text unit identifiers, hierarchical paths, text content, and reference relationships from hierarchical text units to form text unit features.

[0030] Furthermore, information is extracted from each text unit within the hierarchical text unit structure to obtain a text unit identifier representing the text unit's identity, a hierarchical path representing the text unit's structural position, text content representing the specific text content, and reference relationships representing the connections between text units. The text unit identifier, hierarchical path, text content, and reference relationships are then stored together to form text unit features. When a clause references definitions from other clauses, the reference relationships are extracted to record the association information between the current clause and the referenced clause. When an attachment has a corresponding explanatory relationship with the main text, the reference relationships are recorded to record the association information between the attachment and the corresponding main text content. For text units without explicit reference relationships, potential associations between text units are identified through textual descriptions, numbering correspondences, and structural positional relationships, enabling text unit features to simultaneously reflect text content information and text structural association information. Text unit features are formed by extracting text unit identifiers, hierarchical paths, text content, and reference relationships.

[0031] S1.5. Based on the characteristics of text units, establish the correspondence between the hierarchical text units corresponding to the current version and the hierarchical text units corresponding to the previous version, thus forming a text unit correspondence relationship.

[0032] Furthermore, based on the text unit characteristics of the current and previous versions, hierarchical text units in the two versions are matched. First, a direct correspondence is established between text units with the same identifier based on the text unit identifier. When the text unit identifier changes, association matching is performed based on hierarchical path, text content, and citation relationship to determine the corresponding text units between different versions. When a clause only adjusts its number but its text content and relationship with its chapter remain consistent, the corresponding text units in the two versions are determined through hierarchical path and text content. When a new clause is added to a chapter, causing changes in the numbering of subsequent clauses, the correspondence before and after the change is determined through citation relationship and hierarchical structure relationship. For text units whose text content is adjusted but still belong to the same structural position, a correspondence is established based on the consistency of hierarchical path and the degree of text content association, so that text units between versions can maintain continuous association. By establishing the correspondence between the hierarchical text units corresponding to the current version and the hierarchical text units corresponding to the previous version, a text unit correspondence relationship is formed.

[0033] S1.6 Extract the change information between corresponding text units based on the correspondence between text units to form version change content.

[0034] Furthermore, based on the correspondence between text units, the text content, hierarchical paths, and referencing relationships in the current version text unit and the previous version text unit are compared and analyzed. Information such as text additions, text deletions, text modifications, structural adjustments, and changes in relationships is extracted to form version change content. When a text unit in the current version adds new restrictive descriptions compared to the previous version, the added content is identified by comparing the corresponding text content; when a clause in the previous version is deleted in the current version, the deletion change is identified through the correspondence; when the chapter to which a text unit belongs is adjusted, structural adjustment information is identified through changes in hierarchical paths; when the referenced object changes, related change information is identified through changes in referencing relationships. The text scope corresponding to the change information is limited according to the correspondence between text units, so that the version change content only describes the information differences between the corresponding text units that have changed, avoiding the duplication of unchanged content as change information. By extracting the change information between corresponding text units, version change content is formed.

[0035] S1.7 Associate the version change content with the corresponding text unit identifier, hierarchical path and text content to form version difference data.

[0036] Furthermore, the text unit identifiers, hierarchical paths, and text content before and after the version change are associated with the content of the version change, so that each change can be located to a specific text unit position, and the correspondence between the text content before and after the change is preserved. When the content of a clause is modified, the modified version change content is associated with the text unit identifier and hierarchical path of the corresponding clause, so that the modification location can be accurately obtained later. When the reference relationship changes, the content of the reference relationship change is associated with the corresponding text unit identifier and reference path, so that the scope of the association can be identified later. By uniformly associating the version change content with the text unit features, the version difference data simultaneously contains change information, structural position information, and text association information. By associating the version change content with the corresponding text unit identifier, hierarchical path, and text content, version difference data is formed.

[0037] S2. Based on version difference data, construct a set of review semantic features using a large language model to form review semantic change data.

[0038] S2.1 Organize the text before the change, the text after the change, the text unit identifier, the hierarchical path and the reference relationship in the version difference data to form version difference semantic input data.

[0039] Furthermore, the changes recorded in the version difference data are extracted. The text representing the pre-modification state, the text representing the post-modification state, the text unit identifiers for locating text positions, the hierarchical paths for representing text structural relationships, and the reference relationships for representing text associations are uniformly organized and arranged according to the information structure required for semantic parsing by the large language model. This forms the version difference semantic input data. When the restrictions in a clause are adjusted, both the pre-modification and post-modification texts are retained, and associated with the corresponding text unit identifiers and hierarchical paths, enabling the large language model to identify the location of the changed content. When the reference relationship of a text unit changes, the reference relationships before and after the change are simultaneously organized, allowing the large language model to determine the contextual information corresponding to the text change based on the association relationships. By retaining the correspondence between the pre-modification and post-modification texts, the version difference semantic input data can fully express the semantic environment before and after the text change. Through semantic input organization of the version difference data, version difference semantic input data is formed.

[0040] S2.2 Combine the version difference semantic input data with the preset review semantic feature types, review semantic parsing instructions and review semantic output formats to construct a large language model review semantic parsing task.

[0041] Furthermore, based on the semantic change types that need to be identified during the fair competition review process, review semantic feature types are set. Subject features, behavioral features, constraint features, and relational features are designated as the semantic feature categories that the large language model needs to parse. Version difference semantic input data is associated with the review semantic feature types. Review semantic parsing instructions are set according to the large language model's processing method, limiting the parsing target, parsing scope, and output content of the large language model. This enables the large language model to analyze the semantic differences between the text before and after the change, generating corresponding results according to a preset review semantic output format. When both the text before and after the change contain the same subject but the behavioral description has changed, the review semantic parsing instructions guide the large language model to focus on identifying changes in behavioral features. When changes in citation relationships lead to changes in related content, the review semantic output format requires the large language model to output information on changes in related features. By combining version difference semantic input data, review semantic feature types, review semantic parsing instructions, and review semantic output formats, a large language model review semantic parsing task is constructed.

[0042] S2.3 Input the large language model review semantic parsing task into the large language model, and use the large language model to perform contextual semantic parsing on the text before and after the change, forming a review semantic feature set composed of subject features, behavioral features, constraint features and association features.

[0043] Furthermore, the large language model performs contextual analysis on the text content based on the pre-change and post-change texts, hierarchical paths, and citation relationships in the semantic parsing task. This identifies object information, behavioral descriptions, constraints, and related content within the text, forming subject features, behavioral features, constraint features, and related features. For example, when the applicable object in the text changes, the large language model identifies changes in subject features based on contextual semantics; when the behavior is adjusted, it identifies changes in behavioral features; when constraints are added or deleted, it identifies changes in constraint features; and when citations of other content change, it identifies changes in related features. By combining the contextual relationships between the pre-change and post-change texts, the large language model avoids judging semantic changes solely based on local word changes, ensuring that the review semantic features reflect the overall meaning change of the text. Through contextual semantic parsing of the pre-change and post-change texts using the large language model, a set of review semantic features consisting of subject features, behavioral features, constraint features, and related features is formed.

[0044] S2.4 Establish a correspondence between the review semantic features corresponding to the text before the change and the review semantic features corresponding to the text after the change, and form a review semantic feature correspondence relationship.

[0045] Furthermore, based on the subject features, behavioral features, constraint features, and related features corresponding to the text before the change in the review semantic feature set, and the subject features, behavioral features, constraint features, and related features corresponding to the text after the change, correspondences are established to determine the corresponding positions of different semantic features during the version change process. When the subject features in the text before the change remain consistent with those in the text after the change, a subject feature maintenance relationship is established; when the behavioral features in the text before the change are adjusted, a behavioral feature change relationship is established; when constraint features in the text before the change are added or deleted, a constraint feature change relationship is established; and when the referenced object is adjusted, a related feature change relationship is established. By mapping the review semantic features before and after the change, different types of semantic changes can be distinguished according to feature categories, providing a foundation for subsequent content extraction. By establishing correspondences between the review semantic features corresponding to the text before and after the change, a review semantic feature correspondence relationship is formed.

[0046] S2.5. Extract the changes in subject features, behavioral features, constraint features, and related features based on the correspondence of review semantic features to form review semantic change data.

[0047] Furthermore, based on the correspondence of semantic features in the review process, the differences between the subject features, behavioral features, constraint features, and related features before and after the change are extracted. The changed content corresponding to different semantic features is determined, and different types of changed content are associated and saved to form review semantic change data. For example, when a new applicable subject is added to the changed text, the changed subject feature content is extracted based on the correspondence of subject features; when the changed text adjusts the behavioral description, the changed behavioral feature content is extracted based on the correspondence of behavioral features; when the changed text modifies the limiting conditions, the changed constraint feature content is extracted based on the correspondence of constraint features; when the changed text adjusts the cited content, the changed related feature content is extracted based on the correspondence of related features. By extracting changed content according to semantic feature categories, the review semantic change data can simultaneously reflect the different semantic dimensions involved in the text change. Review semantic change data is formed by extracting the changed content of subject features, behavioral features, constraint features, and related features based on the correspondence of review semantic features.

[0048] S3. Determine the semantic change level based on the semantic change data, generate semantic change events, update the review knowledge graph, and form a set of change knowledge nodes.

[0049] S3.1. Based on the review semantic change data, establish a semantic constraint chain for review semantic features with mutual constraints according to subject, behavior, target, applicable conditions, exception conditions, definition dependency and reference relationship, and form a review semantic association relationship.

[0050] Furthermore, correlation analysis is conducted on the subject characteristics, behavioral characteristics, constraint characteristics, and related characteristics in the review semantic change data to identify the constraint relationships between different review semantic features. Semantic constraint chains are established based on the interaction between subject and behavior, the association between behavior and its object, the restriction between behavior and applicable conditions, the exclusion between applicable conditions and exceptions, the dependency between defined content and corresponding features, and textual citation relationships. For example, when a text change involves an adjustment of the subject's scope, a constraint relationship from subject characteristics to behavioral characteristics is established based on the correspondence between subject characteristics and behavioral characteristics; when a behavioral content is restricted by applicable conditions, a corresponding constraint relationship is established based on the restriction between behavioral characteristics and constraint characteristics; when a clause cites other defined content, a cross-textual association relationship is established based on definition dependency and citation relationships. Review semantic features with continuous constraint relationships are connected, forming directional semantic associations between different review semantic features. By establishing semantic constraint chains based on review semantic change data, review semantic associations are formed.

[0051] S3.2. Based on the semantic association relationship, calculate the semantic change range of the subject feature, behavior feature, constraint feature and association feature at each constraint transmission position along the semantic constraint chain.

[0052] Furthermore, based on the review semantic feature types corresponding to different constraint transmission positions in the semantic constraint chain, the influence of changes in subject features, behavioral features, constraint features, and related features is analyzed. Combining the influence weights corresponding to review semantic feature types, the semantic change intensity of review semantic features at the current constraint transmission position, and the association transmission coefficient between review semantic features, the semantic change carrying range corresponding to different constraint transmission positions is determined. The semantic change carrying range represents the associated range that can be covered when the current semantic change propagates along the semantic constraint chain to the associated review semantic features. When the subject feature changes, the semantic change corresponding to the subject feature can be transmitted to the behavioral feature along the constraint relationship between the subject and the behavior; when the constraint feature changes, the semantic change corresponding to the constraint feature can be transmitted to the constrained behavioral feature and the object of action along the restriction relationship; when the related feature changes, the semantic change corresponding to the related feature can be transmitted to the review semantic feature corresponding to the referenced content along the citation relationship. By analyzing the semantic change carrying range for each constraint transmission position, the degree of propagation of semantic change in the semantic constraint chain is determined. By calculating the semantic change carrying range based on the review semantic association relationship, the semantic change carrying range corresponding to different constraint transmission positions is obtained.

[0053] The semantic change carrying scope expression is: ; in, For the first The semantic change range corresponding to each constraint transmission position. For the first The influence weights corresponding to the semantic features of the class review For the first Class review semantic features in the first The intensity of semantic change at each constraint transmission location, For the first Class review semantic features in the first The correlation transmission coefficient of each constraint transmission location, This is the constraint transit position number in the semantic constraint chain. To review the semantic feature type number.

[0054] =1: Indicates the main characteristic. =2: Indicates behavioral characteristics. =3: Indicates a constraint feature. =4: Indicates a related feature.

[0055] S3.3. The position where the scope of the semantic change is extended from the current review semantic feature to the associated review semantic feature is determined as the constraint transmission termination position, and a semantic change hierarchy is formed based on the semantic constraint chain depth of the constraint transmission termination position.

[0056] Furthermore, based on the scope of semantic change, the propagation is extended along the semantic constraint chain. When the semantic change can affect the associated review semantic features, the current extension position is determined as the effective propagation position, and subsequent associated positions are analyzed. When the semantic change can no longer affect new associated review semantic features, the current propagation position is determined as the constraint transmission termination position. When the change of the subject feature only affects the corresponding behavioral feature and constraint feature, the constraint transmission termination position is determined based on the propagation scope of the semantic constraint chain. When the change of the citation relationship affects the associated definition content, the propagation scope of the semantic change continues to expand based on the citation relationship until there are no new associated review semantic features. The semantic change level is determined based on the depth of the semantic constraint chain traversed from the current review semantic feature to the constraint transmission termination position, so that semantic changes of different scopes can be distinguished according to the propagation scope. By determining the semantic constraint chain depth based on the constraint transmission termination position, a semantic change level is formed.

[0057] S3.4. Based on the semantic change level, aggregate the review semantic change data that have the same semantic change level and share the semantic constraint chain, and determine the change boundary based on the continuity of the semantic constraint chain to form a semantic change unit.

[0058] Furthermore, the semantic change levels corresponding to different review semantic change data are compared. Review semantic change data with the same semantic change level and propagating along the same semantic constraint chain are aggregated, so that semantic change content with a relationship forms a unified change range. For example, when multiple behavioral feature changes are caused by changes in the same subject feature and propagate along the same semantic constraint chain, multiple behavioral feature changes are aggregated to form corresponding semantic change units; when there is no continuous semantic constraint chain relationship between different changes, different changes are divided into different semantic change units, and the change boundary is determined according to the continuity of the semantic constraint chain. When the semantic constraint chain remains continuous, related changes are assigned to the same semantic change unit; when the semantic constraint chain is broken, the break point is determined as the boundary between different semantic change units, and semantic change units are formed by aggregating according to the semantic change level and the continuity of the semantic constraint chain.

[0059] S3.5. Based on the semantic change unit, extract the corresponding change propagation path along the semantic constraint chain, and generate semantic change events based on the change propagation path.

[0060] Furthermore, the changes in the review semantic features within the semantic change unit are analyzed. The propagation relationship from the initial review semantic feature to the associated review semantic features is extracted along the semantic constraint chain, forming a change propagation path. When a change in the subject feature leads to a change in the behavior feature, the propagation path from the subject feature to the behavior feature is extracted; when a change in the behavior feature affects the constraint feature, a continuous propagation path from the behavior feature to the constraint feature is extracted; when a change in the citation relationship leads to a change in the associated content, the cross-text propagation path corresponding to the citation relationship is extracted. The change propagation path is then associated with the changed content within the semantic change unit to generate a semantic change event that represents the starting point of the change, the scope of propagation, and the associated object. This is achieved by extracting the change propagation path from the semantic change unit and generating semantic change events based on the change propagation path.

[0061] S3.6. Based on the semantic change events, determine the corresponding review semantic features, change direction, change scope and correlation, and form knowledge graph update content.

[0062] Furthermore, the starting position, propagation path, and ending position of semantic change events are analyzed to determine the subject characteristics, behavioral characteristics, constraint characteristics, and related characteristics corresponding to the semantic change events. The direction of change is determined based on the difference between the pre-change and post-change states. The direction of change includes at least addition, deletion, content adjustment, and related change directions. For cases where new subjects, behaviors, or constraints are added to the post-change text, the direction of change is determined as addition; for cases where the post-change text no longer contains subjects, behaviors, or constraints from the pre-change text, the direction of change is determined as deletion; for cases where the semantic features are retained but the semantic content changes, the direction of change is determined as content adjustment; and for cases where the definition dependencies or reference relationships between semantic features change, the direction of change is determined as related change.

[0063] The scope of change is determined based on the semantic change unit and propagation path corresponding to the semantic change event. The scope of change is limited to the review semantic features at the start of the change, the related review semantic features along the propagation path, and the review semantic features corresponding to the end position of the change. Subject relationships, behavioral relationships, constraint relationships, definition dependencies, and reference relationships in the semantic constraint chain are extracted as association relationships. When the scope of application of the subject features is extended and affects the behavioral and constraint features along the semantic constraint chain, the subject features, behavioral features, and constraint features are determined as the corresponding review semantic features. The extension of the scope of application is determined as the direction of change. The review semantic features covered by the continuous propagation path of change are determined as the scope of change. The interaction relationship between the subject and behavior and the restriction relationship between behavior and constraint conditions are determined as association relationships. By organizing the review semantic features, direction of change, scope of change, and association relationships accordingly, the updated content of the knowledge graph is formed.

[0064] S3.7. Based on the updated content of the knowledge graph, locate and examine the corresponding knowledge nodes and relationship edges in the knowledge graph to form a set of knowledge nodes to be updated.

[0065] Furthermore, based on the review semantic features in the updated knowledge graph content, knowledge nodes corresponding to semantic content, node type, text unit identifier, and hierarchical path are retrieved in the review knowledge graph, and relation edges connecting the corresponding knowledge nodes are retrieved based on the association relationships in the updated knowledge graph content. For subject features recorded in the updated knowledge graph content, subject type knowledge nodes are located in the review knowledge graph. For behavioral features recorded in the updated knowledge graph content, behavioral type knowledge nodes are located in the review knowledge graph. For constraint features recorded in the updated knowledge graph content, knowledge nodes corresponding to applicable conditions, exception conditions, or restrictions are located in the review knowledge graph. For association features recorded in the updated knowledge graph content, knowledge nodes and relational edges corresponding to the defined dependencies or references are located in the review knowledge graph. First, the search scope of knowledge nodes is limited based on text unit identifiers and hierarchical paths. Then, the target knowledge nodes are determined based on the semantic content and node type of the reviewed semantic features. This avoids incorrect correspondences between the same or similar text content in different structural positions. When a review semantic feature in the updated knowledge graph content does not have a corresponding knowledge node in the review knowledge graph, the newly added position corresponding to the review semantic feature is recorded. When a corresponding knowledge node exists in the review knowledge graph, the corresponding knowledge node and the relational edges connecting it are included in the set of knowledge nodes to be updated. By locating the knowledge nodes and relational edges corresponding to the updated knowledge graph content, a set of knowledge nodes to be updated is formed.

[0066] S3.8. Based on the set of knowledge nodes to be updated and the updated content of the knowledge graph, add, replace, invalidate or adjust the knowledge nodes and relation edges in the review knowledge graph to form the updated review knowledge graph. Furthermore, based on the change direction recorded in the updated knowledge graph content, the update method for knowledge nodes and relation edges in the set of knowledge nodes to be updated is determined. For cases where the change direction is addition and there is no corresponding knowledge node in the reviewed knowledge graph, a new knowledge node is created based on the reviewed semantic features, and new relation edges are created based on the association relationships. For cases where the change direction is content adjustment and there is a corresponding knowledge node in the reviewed knowledge graph, the semantic content corresponding to the changed reviewed semantic features is replaced with the corresponding knowledge node, while retaining the correspondence between the knowledge node and the text unit identifier and hierarchical path. For cases where the change direction is deletion, the corresponding knowledge node or relation edge is marked as invalid, so that the corresponding knowledge node or relation edge no longer participates in the current version's association analysis. For cases where the change direction is an association change direction, the update method is determined based on the changed definition dependencies. Alternatively, the relationship edges between knowledge nodes can be adjusted by modifying the reference relationships. For example, if a behavioral feature in the original text references a defined content, and the text changes to reference a different defined content, the knowledge node corresponding to the behavioral feature is retained, the reference relationship between the behavioral feature's knowledge node and the original defined content's knowledge node is removed, and a new reference relationship is established between the behavioral feature's knowledge node and the changed defined content's knowledge node. When the same semantic change event simultaneously includes changes to the review semantic feature content and changes to the relationship relationships, the knowledge nodes are first replaced based on the changed review semantic features, and then the relationship edges are adjusted based on the changed relationship relationships to ensure consistency between the knowledge node content and the relationship status between knowledge nodes. By adding, replacing, invalidating, or adjusting the relationship edges of knowledge nodes and relationship edges in the review knowledge graph, an updated review knowledge graph is formed.

[0067] S3.9 Extract knowledge nodes that are directly updated by semantic change events or affected by association relationships based on the updated review knowledge graph to form a set of changed knowledge nodes.

[0068] Furthermore, based on the updated knowledge graph content corresponding to the semantic change event, knowledge nodes that have been added, replaced, invalidated, or have undergone association adjustments are extracted from the updated review knowledge graph. These nodes are considered directly updated knowledge nodes due to the semantic change event. Starting from the directly updated knowledge nodes, knowledge nodes that are associated with the directly updated knowledge nodes are retrieved along the subject relationships, behavioral relationships, constraint relationships, definition dependencies, and reference relationships in the updated review knowledge graph. The impact of the associated knowledge nodes is determined based on the scope of change corresponding to the semantic change event. When an associated knowledge node is located within the change propagation path corresponding to the semantic change event, and there is a continuous association between the associated knowledge node and the directly updated knowledge node, the associated knowledge node is identified. The set of knowledge nodes affected by association relationships is defined as follows: when an associated knowledge node is not located within the change propagation path, or when the association relationship terminates at the change boundary, the associated knowledge node is not included in the set of changed knowledge nodes. When the knowledge node corresponding to the constraint feature is replaced, and the knowledge node corresponding to the behavioral feature that has a direct restriction relationship with the knowledge node corresponding to the constraint feature is located within the change propagation path, the knowledge node corresponding to the behavioral feature is included in the set of changed knowledge nodes. Knowledge nodes that only have a non-continuous reference relationship with the knowledge node corresponding to the constraint feature and are not located within the change propagation path are not included in the set of changed knowledge nodes. The set of changed knowledge nodes is formed by extracting knowledge nodes that are directly updated by semantic change events or affected by association relationships.

[0069] S4. Perform correlation impact analysis based on the set of changed knowledge nodes to form a set of candidate impact nodes. Match the set of candidate impact nodes with the historical review conclusion dependency slices to determine the minimum recalculation domain.

[0070] S4.1 Based on the set of changed knowledge nodes, extract the relationship type, relationship direction and association level corresponding to each changed knowledge node to form the association features of changed knowledge nodes.

[0071] Furthermore, the association information of each changed knowledge node in the set of changed knowledge nodes is extracted to obtain the relationship type, relationship direction, and association level of the changed knowledge node in the knowledge graph under review. The relationship type represents the definition relationship, reference relationship, constraint relationship, and dependency relationship between the changed knowledge node and other knowledge nodes; the relationship direction represents the connection direction between the changed knowledge node and related knowledge nodes; and the association level represents the position range of the changed knowledge node when expanding outward along the knowledge relationship. When there is a constraint relationship between the subject feature and the behavior feature corresponding to the changed knowledge node, the constraint relationship between the knowledge node corresponding to the subject feature and the knowledge node corresponding to the behavior feature is extracted, and the relationship direction from the knowledge node corresponding to the subject feature to the knowledge node corresponding to the behavior feature is determined. When the definition content corresponding to the changed knowledge node is referenced by other text units, the definition relationship and reference relationship are extracted, and the corresponding reference direction is determined. Different relationship types, relationship directions, and association levels corresponding to the same changed knowledge node are associated and saved, so that the changed knowledge node can express the connection state with other knowledge nodes through association features. By extracting the relationship type, relationship direction, and association level corresponding to the changed knowledge node, the association features of the changed knowledge node are formed.

[0072] S4.2 Based on the characteristics of the association between the knowledge nodes of change, the impact of change is transmitted step by step along the definition relationship, reference relationship, constraint relationship and dependency relationship in the knowledge graph of review, forming the change impact propagation relationship.

[0073] Furthermore, based on the association characteristics of changed knowledge nodes, the association paths of changed knowledge nodes in the review knowledge graph are determined, and the process is expanded step by step along definition relationships, reference relationships, constraint relationships, and dependency relationships to analyze the scope of influence of changed knowledge nodes on associated knowledge nodes. When the definition relationship corresponding to a changed knowledge node changes, associated knowledge nodes referencing the definition content are found along the definition relationship, and the impact of the change is transmitted; when the reference relationship corresponding to a changed knowledge node changes, the knowledge nodes corresponding to the referenced content are found along the reference relationship, and the direction of the association impact is determined; when the constraint relationship corresponding to a changed knowledge node changes, the knowledge nodes corresponding to the behavioral characteristics or objects affected by the constraints are found along the constraint relationship; when the change... When the dependencies corresponding to a knowledge node change, the knowledge node whose dependency has changed is searched along the dependency relationship for judgment or association analysis. When the knowledge node corresponding to a certain constraint feature changes, the behavioral feature knowledge node corresponding to the constraint feature is first identified, and the fact node or evidence node that has a dependency relationship with the behavioral feature is identified. The continuous association path is recorded as the change impact propagation relationship. The propagation path between the changed knowledge node and the associated knowledge node is recorded according to the relationship type and relationship direction, so that the change impact propagation relationship can represent the source of change, the direction of propagation, and the scope of association. The change impact is formed by passing the change impact level by level along the definition relationship, reference relationship, constraint relationship and dependency relationship in the knowledge graph.

[0074] S4.3. Based on the propagation relationship of change impact, related knowledge nodes that share the same change knowledge node and are located on the continuous change impact propagation path are grouped together to form an impact knowledge node set. In addition, impact knowledge nodes that have direct or indirect review dependence on the change knowledge node are retained in combination with the change impact propagation relationship to form a candidate impact node set.

[0075] Furthermore, based on the propagation relationship of change impacts, path analysis is performed on the associated knowledge nodes corresponding to different change knowledge nodes. Associated knowledge nodes that share the same change knowledge node source and are on a continuous change impact propagation path are grouped together, forming an impact knowledge node set of knowledge nodes with the same source and continuous propagation relationship. When the same change knowledge node simultaneously affects multiple behavioral feature knowledge nodes through constraint relationships, the multiple behavioral feature knowledge nodes are grouped into the corresponding impact knowledge node set. When multiple change knowledge nodes affect the same fact node through reference relationships, the corresponding fact node is included in the associated grouping scope. Based on the impact knowledge node set, the existence of direct or indirect review dependency relationships between associated knowledge nodes and change knowledge nodes is determined according to the change impact propagation relationship. When there is a direct relationship between the related knowledge node and the changed knowledge node, the related knowledge node is retained in the candidate impact node set. When there is no direct relationship between the related knowledge node and the changed knowledge node, but the changed knowledge node can be traced along the continuous change impact propagation path, the related knowledge node is identified as an indirect review dependency node and retained in the candidate impact node set. When the related knowledge node only has a non-continuous relationship or no review dependency relationship, it is not included in the candidate impact node set. The candidate impact node set is formed by combining the change impact propagation relationship to screen the impact knowledge nodes that have a direct or indirect review dependency relationship with the changed knowledge node.

[0076] S4.4 Extract the corresponding node identifier, node type, association path and scope of influence from the candidate impact node in the candidate impact node set, and associate them with the rule nodes, fact nodes, evidence nodes and association paths in the historical review conclusion dependency slice to form candidate impact node association data; Furthermore, information is extracted from each candidate impact node in the candidate impact node set to obtain node identifiers representing the identity of knowledge nodes, node types representing the categories of knowledge nodes, association paths representing the connections between knowledge nodes, and impact range representing the scope of change propagation. This extracted information is then matched and associated with rule nodes, fact nodes, evidence nodes, and association paths in the historical review conclusion dependency slice. When a candidate impact node corresponds to a behavioral characteristic, the correspondence between the candidate impact node and rule nodes or fact nodes is determined based on the node type, and the connection method by which the candidate impact node affects the historical review conclusion is determined through the association path. When the citation relationship corresponding to a candidate impact node changes, evidence nodes and association paths in the historical review conclusion dependency slice are searched based on the association path, and the association relationship between the candidate impact node and evidence nodes is determined. For candidate impact nodes with consistent node identifiers and rule nodes, fact nodes, or evidence nodes in the historical review conclusion dependency slice, a direct association relationship is established. For candidate impact nodes with different node identifiers but continuous association paths and nodes in the historical review conclusion dependency slice, an indirect association relationship is established through the association path, thus forming a correspondence between the candidate impact node and the historical review conclusion dependency slice. Candidate impact nodes are associated with rule nodes, fact nodes, evidence nodes, and related paths in historical review conclusion dependency slices to form candidate impact node association data.

[0077] S4.5. Based on the candidate impact node association data, analyze the dependency relationship between the candidate impact node and the historical review conclusion along the dependency path in the historical review conclusion dependency slice, and determine the historical review conclusion dependency slice corresponding to the candidate impact node to form the conclusion impact association relationship. Furthermore, based on the node identifiers, node types, and association paths recorded in the candidate impact node association data, rule nodes, fact nodes, and evidence nodes that have dependencies on the candidate impact nodes are searched in the historical review conclusion dependency slices. Association analysis is then performed along the dependency paths in the historical review conclusion dependency slices. When a candidate impact node directly participates in the historical review conclusion generation process, it is identified as a direct impact node, and its corresponding historical review conclusion dependency slice is located. When a candidate impact node indirectly affects the historical review conclusion through rule nodes, fact nodes, or evidence nodes, the connection between the candidate impact node and the historical review conclusion is traced along the dependency path, and the corresponding historical review conclusion dependency slice is determined. When the rule node corresponding to a candidate impact node changes, it is analyzed whether the rule node is cited in the historical review conclusion. When the fact node corresponding to a candidate impact node changes, it is analyzed whether the fact node serves as the basis for judging the historical review conclusion. When the evidence node corresponding to a candidate impact node changes, it is analyzed whether the evidence node participates in the generation of the historical review conclusion. By recording the direct or indirect dependencies between candidate impact nodes and historical review conclusion dependency slices, a conclusion impact association relationship that can represent the influence relationship between the changed node and the historical review conclusion is formed. By analyzing the dependencies between candidate impact nodes and historical review conclusions, the correlation between the impact of conclusions is formed.

[0078] S4.6. Based on the impact of the conclusions on the relationship, the historical review conclusions that have direct dependencies or indirect dependencies formed through the relationship path are collected to form a set of conclusions to be recalculated. Furthermore, based on the influence relationships of the conclusions, historical review conclusion dependency slices affected by candidate impact nodes are screened and grouped according to dependency type. For candidate impact nodes that directly correspond to rule nodes, fact nodes, or evidence nodes in historical review conclusion dependency slices, the corresponding historical review conclusion dependency slices are directly included in the set of conclusions to be recalculated. For candidate impact nodes that are not directly connected to historical review conclusion dependency slices but can be transmitted to them through continuous association paths, the corresponding historical review conclusion dependency slices are included in the set of conclusions to be recalculated as indirect impact objects. For example, if a rule node changes, historical review conclusion dependency slices directly referencing the rule node are included in the set of conclusions to be recalculated; if a fact node changes, and the dependency relationship between the fact node and the rule node affects the corresponding historical review conclusion dependency slice, the corresponding historical review conclusion dependency slice is included in the set of conclusions to be recalculated. Cases without direct dependencies or where the association path cannot be continuously transmitted to the historical review conclusion dependency slice are not included in the set of conclusions to be recalculated. The set of conclusions to be recalculated is formed by grouping the affected historical review conclusion dependency slices.

[0079] S4.7 Based on the set of conclusions to be recalculated, extract the text units, rule nodes, fact nodes, and evidence nodes from the corresponding historical review conclusion dependency slices, and combine the text units, rule nodes, fact nodes, and evidence nodes into the smallest recalculation domain; Furthermore, based on the set of conclusions to be recalculated, the corresponding historical review conclusion dependency slices are located. The text units, rule nodes, fact nodes, and evidence nodes that participate in the review conclusion generation process in each historical review conclusion dependency slice are obtained. The association range between nodes is determined based on the dependency path between each node. Taking the change knowledge node corresponding to the candidate influencing node as the starting node, the dependency path in the historical review conclusion dependency slice is traced backward to determine the shortest association range that affects the current historical review conclusion generation process. The text units, rule nodes, fact nodes, and evidence nodes that can affect the generation of review conclusions are identified as valid recalculation nodes.

[0080] Specifically, when a text unit corresponding to a clause changes, and the changed knowledge node affects a rule node through rule dependencies, the association path between the changed text unit and the rule node is first determined. Then, the fact nodes and evidence nodes involved in generating the review conclusion are obtained along the fact dependency path and evidence dependency path corresponding to the rule node. The changed text unit, rule node, fact node, and evidence node are jointly determined as the smallest recalculation domain. When the changed knowledge node only affects the citation relationship of a certain evidence node, the dependency path corresponding to the evidence node is traced upwards only to extract the text unit, rule node, and fact node that have a direct dependency relationship with the evidence node, without obtaining other nodes that do not have a dependency relationship.

[0081] Based on the dependencies in the historical review conclusions' dependency slices, the extracted text units, rule nodes, fact nodes, and evidence nodes are filtered. Nodes are retained in the minimum recalculation domain when they are on the change propagation path corresponding to candidate impact nodes, or when they participate in the generation of historical review conclusions through continuous dependency paths. Nodes are excluded from the minimum recalculation domain when there is no continuous dependency path between them and candidate impact nodes, or when they do not participate in the corresponding historical review conclusion generation process. This ensures that the minimum recalculation domain contains only the necessary text units, rule nodes, fact nodes, and evidence nodes that are affected by changes and participate in the review conclusion generation process. The minimum recalculation domain is formed by extracting the text units, rule nodes, fact nodes, and evidence nodes corresponding to the set of conclusions to be recalculated.

[0082] S4.8 Determine the scope of text, rules, and evidence that need to be re-examined based on the minimum recalculation domain, and obtain the scope of evidence to be obtained for the review.

[0083] Furthermore, based on the relationships between text units, rule nodes, fact nodes, and evidence nodes within the minimum recalculation domain, the content requiring re-examination is categorized and determined. The text scope corresponding to the re-examination is determined based on text units; the rule scope requiring re-matching or analysis is determined based on rule nodes; and the evidence scope requiring re-acquisition or verification is determined based on evidence nodes. The relationships between rule nodes, fact nodes, and evidence nodes serve as the basis for scope determination. When the minimum recalculation domain contains only text units and rule nodes corresponding to a specific changed clause, the corresponding text content and rule content are determined as the scope of re-examination. When the minimum recalculation domain contains fact nodes and evidence nodes, the corresponding fact-related content and evidence content are determined as the scope of evidence acquisition for review. By associating the text scope, rule scope, and evidence scope, the scope of evidence acquisition for review corresponding to the changed content in the current version is determined, ensuring that subsequent review processes only acquire evidence content related to the changed content. By determining the text scope, rule scope, and evidence scope requiring re-examination based on the minimum recalculation domain, the scope of evidence acquisition for review is obtained.

[0084] S5. Obtain the corresponding review evidence based on the minimum recalculation domain, perform local incremental review using the large language model, and generate the updated review conclusion.

[0085] S5.1 Extract the corresponding text units, rule nodes, fact nodes, and evidence nodes based on the minimum recalculation domain, determine the scope of review evidence corresponding to the minimum recalculation domain, and form the scope of review evidence.

[0086] Furthermore, based on the text units, rule nodes, fact nodes, and evidence nodes contained in the minimum recalculation domain, the relevant content involved in the current review conclusion generation process is extracted, and the scope of review evidence that needs to be re-acquired and analyzed is determined according to the dependencies between each node. Specifically, text units are used to determine the scope of normative document changes that need to be re-analyzed, rule nodes are used to determine the scope of rule content that needs to be re-matched, fact nodes are used to determine the scope of factual associations that need to be re-confirmed, and evidence nodes are used to determine the scope of evidence content that needs to be re-acquired or verified. When the minimum recalculation domain contains changed text units and corresponding rule nodes, the relevant review evidence is obtained based on the fact nodes and evidence nodes corresponding to the rule nodes. When the minimum recalculation domain only contains changes in evidence nodes, the text units, rule nodes, and fact nodes related to the evidence nodes are obtained based on the dependencies corresponding to the evidence nodes. Based on the relationships between text units, rule nodes, fact nodes, and evidence nodes, valid content participating in the current version review process is screened, and the scope of definition content, cited content, rule content, and fact content that can support subsequent review judgments is determined as the scope of review evidence. By extracting the corresponding text units, rule nodes, fact nodes, and evidence nodes based on the minimum recalculation domain, the scope of review evidence corresponding to the minimum recalculation domain is determined, thus forming the scope of review evidence.

[0087] S5.2 Obtain the corresponding definitions, references, rules, and facts based on the scope of the evidence to be reviewed, and associate them to form an incremental set of evidence for review.

[0088] Furthermore, based on the text units, rule nodes, fact nodes, and evidence nodes within the scope of the review evidence, the definition content, cited content, rule content, and factual content related to the review process are obtained respectively. These are then linked according to the correspondence between text units and rule nodes, the judgment relationship between rule nodes and fact nodes, and the supporting relationship between fact nodes and evidence nodes. When a rule node corresponds to a review rule, the corresponding rule content is obtained based on the rule node, and the factual content supporting the rule judgment is obtained based on the fact node. When a text unit cites other content, the corresponding cited content is obtained based on the citation relationship, and the cited content is linked to the text unit and rule node. When an evidence node changes, the corresponding evidence content is obtained based on the evidence node, and a supporting relationship is established with the fact node. The definition content, cited content, rule content, and factual content are organized according to the review dependency relationship, enabling different types of review evidence to correspond to specific text units, rule nodes, fact nodes, and evidence nodes, forming an incremental review evidence set with related relationships. By obtaining and linking the corresponding definition content, cited content, rule content, and factual content according to the scope of the review evidence, an incremental review evidence set is formed.

[0089] S5.3 Input the minimum recalculation domain, the incremental review evidence set, and the corresponding review semantic change data into the large language model to construct a local incremental review task.

[0090] Furthermore, the text units, rule nodes, fact nodes, and evidence nodes in the minimum recalculation domain are used as the input for the review scope, and the definition content, cited content, rule content, and fact content in the incremental review evidence set are used as the input for the review basis. Combined with the changes in subject features, behavioral features, constraint features, and relational features in the review semantic change data, the review task content of the large language model is organized to form a local incremental review task. When the review semantic change data indicates a change in subject features, the changed subject features are associated with the corresponding rule content and fact content, enabling the large language model to perform review analysis for changes in subject scope. When the review semantic change data indicates a change in citation relationships, the changed citation relationships are associated with the corresponding citation content and evidence content, enabling the large language model to perform review analysis for changes in related content. By combining the review scope defined by the minimum recalculation domain, the review basis provided by the incremental review evidence set, and the change direction provided by the review semantic change data, the large language model performs review inference only for the affected area. By inputting the minimum recalculation domain, the incremental review evidence set, and the corresponding review semantic change data into the large language model, a local incremental review task is constructed.

[0091] S5.4. Using a large language model, perform correlation analysis on the minimum recalculation domain, incremental review evidence set, and review semantic change data in the local incremental review task to determine the correspondence between the updated review facts and review rules, and form an updated review fact set.

[0092] Furthermore, the large language model analyzes the correspondence between the changed text content and review rules based on the changes in text unit content, rule content, factual content, evidence content, and review semantic change data in the local incremental review task. It also determines the matching relationship between the changed review facts and review rules. When behavioral features in a text unit change, the large language model analyzes whether the corresponding rule node still applies to the changed behavioral content based on the behavioral feature changes, and determines the updated review facts based on the factual and evidence nodes. When constraint features change, the large language model analyzes whether the restrictions in the rule content have changed based on the constraint feature changes, and redetermines the correspondence between the factual and rule content. The text units, rule nodes, factual nodes, and evidence nodes after the correlation analysis are saved accordingly, ensuring that the updated review facts reflect the review basis after the current version of the text changes. By using the large language model to perform correlation analysis on the minimum recalculation domain, incremental review evidence set, and review semantic change data in the local incremental review task, the updated review facts and review rules correspondence is determined, forming an updated review fact set.

[0093] S5.5 Generate the corresponding fair competition review result based on the updated review fact set, and associate the fair competition review result with the corresponding text unit, rule node, fact node and evidence node to form the updated review conclusion.

[0094] Furthermore, based on the correspondence between text units, rule nodes, fact nodes, and evidence nodes in the updated review fact set, the changes in the current version are reviewed and judged, and corresponding fair competition review results are generated. These results are then associated with the text units, rule nodes, fact nodes, and evidence nodes that generated them, ensuring the review results can be traced back to the corresponding text content, judgment basis, and supporting evidence. When a change in a text unit leads to a change in the correspondence of rule nodes, the fair competition review result is associated and saved with the changed text unit, the updated rule node, and the corresponding fact and evidence nodes. When the changes do not alter existing review facts, the updated fair competition review result continues to be associated with the remaining valid text units, rule nodes, fact nodes, and evidence nodes. The fair competition review results generated by the partial incremental review are saved as the updated review conclusion, ensuring the updated review conclusion corresponds to the changes in the current version. By generating corresponding fair competition review results based on the updated review fact set and associating them with the corresponding text units, rule nodes, fact nodes, and evidence nodes, an updated review conclusion is formed.

[0095] S6. Merge the updated review conclusions with the unaffected historical review conclusions to obtain the current version of the review results.

[0096] S6.1 Extract the corresponding text units, rule nodes, fact nodes, and evidence nodes from the updated review conclusion and the unaffected historical review conclusion to form review conclusion related data.

[0097] Furthermore, the updated review conclusions and the associated information corresponding to unaffected historical review conclusions are obtained. Text units, rule nodes, fact nodes, and evidence nodes involved in the conclusion generation process are extracted from each review conclusion. Text units represent the location of the normative document content corresponding to the review conclusion; rule nodes represent the rule basis for the review conclusion; fact nodes represent the factual judgment content corresponding to the review conclusion; and evidence nodes represent the evidence information supporting the generation of the review conclusion. The text units, rule nodes, fact nodes, and evidence nodes in the updated review conclusions, as well as those in the unaffected historical review conclusions, are linked and recorded. When the updated review conclusion originates from a changed text unit, the rule nodes, fact nodes, and evidence nodes corresponding to the changed text unit are extracted; when the text unit corresponding to the unaffected historical review conclusion remains unchanged, the text units, rule nodes, fact nodes, and evidence nodes corresponding to the historical review conclusion are retained. By extracting the text units, rule nodes, fact nodes, and evidence nodes corresponding to the updated review conclusions and the unaffected historical review conclusions respectively, review conclusion association data is formed.

[0098] S6.2. The updated review conclusions are associated with the unaffected historical review conclusions according to the correspondence of text units, rule nodes, fact nodes, and evidence nodes to form a review conclusion fusion relationship.

[0099] Furthermore, based on the correspondence between text units, rule nodes, fact nodes, and evidence nodes in the review conclusion association data, the updated review conclusion and the unaffected historical review conclusion are matched and associated. The correspondence between text units determines the corresponding positions of text content in the current version and text content in the historical version; the correspondence between rule nodes determines the association between updated rule content and historical rule content; the correspondence between fact nodes determines the correspondence between the factual judgment basis before and after the update; and the correspondence between evidence nodes determines the correspondence between the evidence content supporting the review conclusion. When a text unit changes and its corresponding rule node is updated, a correspondence is established between the changed text unit and the updated rule node. When a historical review conclusion is unaffected, the corresponding content of the historical review conclusion is retained in the current version's review conclusion association scope based on the retention relationships between text units, rule nodes, fact nodes, and evidence nodes. A connection relationship is established between the updated review conclusion and the unaffected historical review conclusion through the correspondences between text units, rule nodes, fact nodes, and evidence nodes, enabling the changed content and the retained content to be integrated according to the correspondence. By associating text units, rule nodes, fact nodes, and evidence nodes, a fusion relationship for review conclusions is formed.

[0100] S6.3 Combine the changes in the updated review conclusions with the unchanged content in the historical review conclusions that were not affected to form the current version of the review conclusions set.

[0101] Furthermore, based on the fusion relationship of the review conclusions, the review conclusion content that needs to be replaced and the historical review conclusion content that needs to be retained are determined. The changed content in the updated review conclusion is replaced in the corresponding position, and the content that remains unchanged in the historical review conclusion is retained, thus forming a complete set of review conclusions corresponding to the current version. When the review conclusion corresponding to a certain text unit needs to be regenerated due to changes in the rule node, the updated review conclusion replaces the corresponding part in the original historical review conclusion. When a certain text unit has not changed and the corresponding rule node, fact node, and evidence node remain valid, the corresponding content of the historical review conclusion is retained. The replacement boundary between the updated review conclusion and the historical review conclusion is determined based on the fusion relationship of the review conclusions, so that only the changed review conclusion content is updated, while the unaffected historical review conclusion content remains continuous. By combining the changed content in the updated review conclusion with the content that remains unchanged in the unaffected historical review conclusion, the current version of the review conclusion set is formed.

[0102] S6.4. Based on the current version's review conclusion set, associate and save the corresponding text units, rule nodes, fact nodes, and evidence nodes to form the current version's review result.

[0103] Furthermore, for each review conclusion in the current version's review conclusion set, the corresponding text units, rule nodes, fact nodes, and evidence nodes are extracted. The relationships between these nodes are saved based on the review conclusion fusion relationship, ensuring that the current version's review results correspond to specific text content, rule basis, factual judgments, and evidence information. For current version review results generated from updated review conclusions, the relationships between changed text units and updated rule nodes, fact nodes, and evidence nodes are saved. For current version review results formed from unaffected historical review conclusions, the relationships between the original text units, rule nodes, fact nodes, and evidence nodes are saved. By saving the text units, rule nodes, fact nodes, and evidence nodes corresponding to the current version's review conclusion set through association, the current version's review results possess corresponding and traceable relationships. By saving the text units, rule nodes, fact nodes, and evidence nodes corresponding to the current version's review conclusion set through association, the current version's review results are formed.

[0104] In summary, this invention identifies version content changes and generates version difference data by acquiring the current and previous versions of normative documents and constructing hierarchical text units. It then uses a large language model to perform semantic analysis on the version difference data, extracting subject features, behavioral features, constraint features, and association features to generate review semantic change data. Based on semantic constraint chains, it analyzes the scope of semantic changes, determines the semantic change hierarchy, generates semantic change events, and updates the review knowledge graph to form a set of change knowledge nodes. Based on the set of change knowledge nodes, it analyzes the propagation relationship of change impacts and determines the minimum recalculation domain by combining historical review conclusions with dependency slices. Based on the minimum recalculation domain, it obtains corresponding review evidence, conducts local incremental review using a large language model, and generates updated review conclusions. Finally, it merges the updated review conclusions with unaffected historical review conclusions to obtain the current version review result, achieving precise impact analysis and incremental fair competition review during the version update process of normative documents.

[0105] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fair competition review method for incremental normative documents based on a large language model, characterized by: This includes obtaining the current and previous versions of the normative documents to be reviewed, constructing hierarchical text units, and performing version alignment to identify version changes and generate version difference data. Based on version difference data, a set of review semantic features is constructed using a large language model to form review semantic change data; Based on the semantic change data reviewed, the semantic change level is determined, semantic change events are generated, and the review knowledge graph is updated to form a set of change knowledge nodes; Based on the set of changed knowledge nodes, perform correlation impact analysis to form a set of candidate impact nodes. Then, match the set of candidate impact nodes with the historical review conclusion dependency slices to determine the minimum recalculation domain. Based on the minimum recalculation domain, obtain the corresponding review evidence, use the large language model to perform local incremental review, and generate an updated review conclusion. The updated review conclusions are combined with the unaffected historical review conclusions to obtain the current version of the review results.

2. The incremental normative document fair competition review method based on a large language model as described in claim 1, characterized in that: The construction of hierarchical text units includes, The text content of the current and previous versions of the normative documents to be reviewed is extracted, the titles, chapters, clauses, provisions and attachments in the text content are identified, and the structural relationships between the titles, chapters, clauses, provisions and attachments are analyzed to form structural content; The structure and content are analyzed to determine the hierarchical relationship between titles, chapters, clauses, items and attachments, and the corresponding text content is divided according to the hierarchical relationship to form hierarchical text fragments; Each hierarchical text fragment is assigned a unique text unit identifier according to its hierarchical relationship, and then organized according to the hierarchical relationship to form hierarchical text units.

3. The incremental normative document fair competition review method based on a large language model as described in claim 2, characterized in that: The version difference data includes, Extract text unit identifiers, hierarchical paths, text content, and referencing relationships from hierarchical text units to form text unit features; Based on the characteristics of text units, establish the correspondence between the hierarchical text units corresponding to the current version and the hierarchical text units corresponding to the previous version, thus forming a text unit correspondence relationship; Extract the change information between corresponding text units based on the correspondence between text units to form version change content; The version changes are associated with the corresponding text unit identifiers, hierarchical paths, and text content to form version difference data.

4. The incremental normative document fair competition review method based on a large language model as described in claim 3, characterized in that: The reviewed semantic change data includes, The text before and after the changes, text unit identifiers, hierarchical paths and reference relationships in the version difference data are organized to form version difference semantic input data. The version difference semantic input data is combined with the preset review semantic feature types, review semantic parsing instructions and review semantic output formats to construct a large language model review semantic parsing task; Input the big language model into the semantic parsing task of reviewing the big language model, and use the big language model to perform contextual semantic parsing on the text before and after the change, forming a set of review semantic features composed of subject features, behavioral features, constraint features and association features; Establish a correspondence between the review semantic features corresponding to the text before the change and the review semantic features corresponding to the text after the change, and form a review semantic feature correspondence relationship; Based on the correspondence of semantic features in the review, the changes in subject features, behavioral features, constraint features, and related features are extracted to form review semantic change data.

5. The incremental normative document fair competition review method based on a large language model as described in claim 4, characterized in that: The semantic change events include, Based on the review semantic change data, review semantic features with mutual constraints are established into semantic constraint chains according to subject, behavior, target, applicable conditions, exception conditions, definition dependency and reference relationship, thus forming review semantic association relationship; Based on the semantic association relationship, calculate the semantic change range of subject features, behavioral features, constraint features and association features at each constraint transmission position along the semantic constraint chain; The position where the scope of the semantic change is extended from the current review semantic feature to the associated review semantic feature is determined as the constraint transmission termination position, and the semantic change hierarchy is formed based on the semantic constraint chain depth of the constraint transmission termination position; Based on the semantic change level, review semantic change data with the same semantic change level and sharing semantic constraint chain are aggregated, and the change boundary is determined according to the continuity of the semantic constraint chain to form semantic change unit. Based on the semantic change unit, the corresponding change propagation path is extracted along the semantic constraint chain, and semantic change events are generated based on the change propagation path.

6. The incremental normative document fair competition review method based on a large language model as described in claim 5, characterized in that: The set of changed knowledge nodes includes, Based on semantic change events, determine the corresponding review semantic features, change direction, change scope and correlation, and form updated content for the knowledge graph; Based on the updated content of the knowledge graph, locate and examine the corresponding knowledge nodes and relationship edges in the knowledge graph to form a set of knowledge nodes to be updated; Based on the set of knowledge nodes to be updated and the updated content of the knowledge graph, the knowledge nodes and relation edges in the review knowledge graph are added, replaced, invalidated or associated and adjusted to form the updated review knowledge graph. Based on the updated review knowledge graph, knowledge nodes that are directly updated by semantic change events or affected by associations are extracted to form a set of changed knowledge nodes.

7. The incremental normative document fair competition review method based on a large language model as described in claim 6, characterized in that: The set of candidate influential nodes includes, Based on the set of changed knowledge nodes, extract the relationship type, relationship direction and association level corresponding to each changed knowledge node to form the association features of changed knowledge nodes; Based on the characteristics of the knowledge nodes that have changed, the impact of the changes is transmitted step by step along the definition relationships, reference relationships, constraint relationships and dependency relationships in the knowledge graph under review, forming a propagation relationship of the impact of changes. Based on the propagation relationship of change impacts, related knowledge nodes that share the same change knowledge node and are located on the continuous change impact propagation path are grouped together to form an impact knowledge node set. In addition, based on the change impact propagation relationship, impact knowledge nodes that have direct or indirect review dependencies with the change knowledge node are retained to form a candidate impact node set.

8. The incremental normative document fair competition review method based on a large language model as described in claim 7, characterized in that: The determination of the minimum recalculation domain includes... Extract the corresponding node identifier, node type, association path and scope of influence from the candidate impact node set, and associate them with the rule nodes, fact nodes, evidence nodes and association paths in the historical review conclusion dependency slice to form candidate impact node association data; Based on the candidate impact node association data, the dependency relationship between the candidate impact node and the historical review conclusion is analyzed along the dependency path in the historical review conclusion dependency slice, and the historical review conclusion dependency slice corresponding to the candidate impact node is determined to form the conclusion impact association relationship. Based on the impact of the conclusions on the relationships, historical review conclusions that have direct dependencies or indirect dependencies formed through the association path are collected to form a set of conclusions to be recalculated. Based on the set of conclusions to be recalculated, extract the text units, rule nodes, fact nodes, and evidence nodes from the corresponding historical review conclusion dependency slices, and combine the text units, rule nodes, fact nodes, and evidence nodes into the smallest recalculation domain; The scope of text, rules, and evidence that need to be re-examined is determined based on the minimum recalculation domain, and the scope of evidence to be obtained for the review is obtained.

9. The incremental normative document fair competition review method based on a large language model as described in claim 8, characterized in that: The updated review conclusions include, Based on the minimum recalculation domain, extract the corresponding text units, rule nodes, fact nodes, and evidence nodes, determine the scope of review evidence corresponding to the minimum recalculation domain, and form the scope of review evidence; Based on the scope of the evidence to be reviewed, obtain the corresponding definitions, references, rules, and facts, and associate them to form an incremental set of evidence for review. Input the minimum recalculation domain, the incremental review evidence set, and the corresponding review semantic change data into the large language model to construct a local incremental review task; Using a large language model, we conduct correlation analysis on the minimum recalculation domain, the incremental review evidence set, and the review semantic change data in the local incremental review task to determine the correspondence between the updated review facts and review rules, and form an updated review fact set. The updated set of facts for review is used to generate the corresponding fair competition review results, and the fair competition review results are associated with the corresponding text units, rule nodes, fact nodes and evidence nodes to form the updated review conclusions.

10. The incremental normative document fair competition review method based on a large language model as described in claim 9, characterized in that: The current version review results include, Extract the corresponding text units, rule nodes, fact nodes, and evidence nodes from the updated review conclusions and the unaffected historical review conclusions to form review conclusion-related data; The updated review conclusions are linked with the unaffected historical review conclusions according to the correspondence between text units, rule nodes, fact nodes, and evidence nodes to form a review conclusion fusion relationship; The changes in the updated review conclusions are combined with the unchanged content in the unaffected historical review conclusions to form the current version of the review conclusions set; Based on the current version of the review conclusion set, the corresponding text units, rule nodes, fact nodes, and evidence nodes are associated and saved to form the current version of the review result.

Images