A method and apparatus for collaborative review of long documents using two images based on evidence slot gap constraints.

By combining document structure graphs and domain knowledge graphs, evidence gap vectors are generated, which solves the problems of missing evidence and insufficient conclusions in the review of long documents, and achieves more accurate and reliable review results.

CN123070124APending Publication Date: 2026-11-10CHINA UNIV OF PETROLEUM (BEIJING)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611266456.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In the review of long documents, existing technologies rely excessively on initial recall evidence for candidate conclusions. It is difficult to locate missing evidence, the direction for supplementary evidence is unclear, and the conclusion output lacks sufficient evidentiary constraints.

Method used

The long document dual-graph collaborative review method based on evidence slot gap constraints recalls initial evidence through document structure graph and domain knowledge graph, generates evidence gap vector, identifies evidence gaps and performs targeted supplementation, ensuring the integrity of evidence and the reliability of conclusions.

Benefits of technology

It improves the accuracy, interpretability, and reliability of long document review conclusions, and can accurately identify evidence gaps across chapters, tables, and rules, and supplement key evidence.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070124A_ABST
    Figure CN123070124A_ABST
Patent Text Reader

Abstract

This application relates to the field of natural language processing technology, and provides a collaborative review method and apparatus for long documents based on evidence slot gap constraints. The method includes: recalling initial evidence for the long document to be reviewed based on a pre-generated document structure graph and a domain knowledge graph; wherein anchor edges are provided between the document structure graph and the domain knowledge graph; generating an evidence gap vector for the long document to be reviewed based on the initial evidence and the review task; and reviewing the long document to be reviewed based on the initial evidence and the evidence gap vector. The method provided by this application can simultaneously characterize the structural relationships of long documents, domain rule constraints, and candidate conclusion evidence requirements. When facing complex review tasks involving multiple chapters, tables, and rules, it can accurately identify evidence gaps and provide targeted supplementation of key evidence, thereby improving the accuracy, interpretability, and reliability of long document review conclusions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to the field of natural language processing technology, specifically to a method and apparatus for collaborative review of long documents with dual graphs based on evidence slot gap constraints. Background Technology

[0002] Currently, in industries requiring high rigor, such as manufacturing, energy, construction, finance, and healthcare, the number of compliance documents, technical specifications, review reports, and business procedures is constantly increasing. These documents are typically lengthy, densely terminological, and their rules, parameter descriptions, applicable conditions, and exceptions may be scattered across different chapters, tables, or appendices. Existing methods for reviewing such long documents generally fall into the following categories: (1) Document retrieval and review method based on keywords or semantic similarity: This method retrieves text fragments related to the review issue from long documents through keyword matching, vector retrieval, or semantic similarity calculation, and generates a review conclusion based on the retrieved content. This method is simple to implement and has high retrieval efficiency, but it mainly relies on text similarity and it is difficult to ensure that key evidence such as rule version, scope of application, table basis, appendix parameters, and exception conditions are fully included.

[0003] (2) Compliance auditing method based on rule base or knowledge graph: This method constructs a rule base, entity relationship or industry knowledge graph, and matches the document content with rule clauses, entity parameters and constraints to determine whether the document meets the regulatory requirements. This method has a certain degree of interpretability, but it usually focuses on rule hit or knowledge association, lacks modeling of the necessary evidence requirements for specific candidate conclusions, and makes it difficult to determine whether the current evidence is sufficient to support the conclusion.

[0004] (3) Document review method based on large language model and multi-round retrieval: This method inputs the retrieved document fragments, rule content and review questions into the large language model, which generates review judgment, explanation of reasons and modification suggestions; when the evidence is insufficient, relevant content is supplemented by extended query or multi-round retrieval. This method has strong language understanding ability, but the process of supplementing evidence lacks clear constraints, is prone to introducing weakly relevant evidence, and is difficult to accurately locate the key evidence type missing in the candidate conclusion. Summary of the Invention

[0005] The present application provides a collaborative review method and apparatus for long documents based on evidence slot gap constraints, aiming to solve the problems in the prior art of long document review scenarios, such as excessive reliance on initial recall evidence for candidate conclusions, difficulty in locating missing evidence, unclear direction of supplementary evidence, and lack of sufficient evidence constraints for conclusion output.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a collaborative review method for long documents using two graphs based on evidence slot gap constraints, comprising: Initial evidence for the long document to be reviewed is retrieved based on the pre-generated document structure graph and domain knowledge graph of the long document to be reviewed; wherein, anchor edges are set between the document structure graph and the domain knowledge graph; Based on the initial evidence and the review task, generate the evidence gap vector of the long document to be reviewed; The long document to be examined is examined based on the initial evidence and the evidence gap vector.

[0007] In some embodiments of this application, generating an evidence gap vector for the long document to be reviewed based on the initial evidence and the review task includes: Based on the initial evidence and the review task, candidate review conclusions are generated; An evidence gap vector is generated for the long document to be reviewed based on the candidate review conclusion; wherein the evidence gap vector is used to characterize the degree of lack of the candidate review conclusion in terms of structural evidence, substantive evidence, rule evidence, version evidence, scope of application evidence, exception condition evidence, tabular evidence, and rebuttal evidence.

[0008] In some embodiments of this application, generating the evidence gap vector of the long document to be examined based on the candidate examination conclusion includes: A necessary set of evidence slots is generated based on the candidate review conclusions; Map the initial evidence to the set of evidence slots, and determine the coverage status of each necessary evidence slot in the set of necessary evidence slots; The evidence gap vector of the long document to be reviewed is generated based on the coverage status.

[0009] In some embodiments of this application, examining the long document to be examined based on the initial evidence and the evidence gap vector includes: If the initial evidence is determined to be incomplete based on the evidence gap vector, a path type constraint is generated based on the type of the evidence gap vector and the uncovered necessary evidence slots. A supplementary verification search is performed on the anchored edge according to the path type constraint; Update the coverage status and the evidence gap vector based on the results of the supplementary evidence search; Determine the completeness of evidence, strength of supporting evidence, strength of rebuttal evidence, and residual gaps based on the updated evidence gap vector; The long document to be reviewed is examined based on the initial evidence, the updated coverage status, the integrity of the evidence, the strength of the supporting evidence, the strength of the counter-evidence, and the residual gaps.

[0010] In some embodiments of this application, the step of generating the document structure diagram includes: The long document to be reviewed is analyzed at a granular level to determine its document units; wherein, the document units include: chapters, paragraphs, clauses, tables, table units, and appendices; The document structure diagram is generated based on the document units; wherein, the document structure diagram is used to represent the hierarchical relationship, sequential relationship, referencing relationship, implicit association relationship, and table text association relationship of the long document to be reviewed.

[0011] In some embodiments of this application, the step of generating the domain knowledge graph includes: The entities, parameters, rules, rule versions, scope of application, exceptions, and logical constraints of the long document to be reviewed are determined based on the rule basis, applicable conditions, and semantic relationships of the review rules. The domain knowledge graph is generated based on the entity, the parameter, the rule, the rule version, the scope of application, the exception conditions, and the logical constraints.

[0012] Secondly, this application provides a collaborative review device for long documents with two diagrams based on evidence slot gap constraints, the device comprising: The initial evidence retrieval module is used to retrieve the initial evidence of the long document to be reviewed based on the pre-generated document structure diagram and domain knowledge graph of the long document to be reviewed; wherein, an anchor edge is set between the document structure diagram and the domain knowledge graph; The evidence gap vector generation module is used to generate an evidence gap vector for the long document to be reviewed based on the initial evidence and the review task. The long document review module is used to review the long document to be reviewed based on the initial evidence and the evidence gap vector.

[0013] In some embodiments of this application, the evidence gap vector generation module includes: The candidate review conclusion generation unit is used to generate candidate review conclusions based on the initial evidence and the review task. An evidence gap vector generation unit is used to generate an evidence gap vector for the long document to be reviewed based on the candidate review conclusion; wherein, the evidence gap vector is used to characterize the degree of lack of the candidate review conclusion in terms of structural evidence, substantive evidence, rule evidence, version evidence, scope of application evidence, exception condition evidence, tabular evidence, and rebuttal evidence.

[0014] In some embodiments of this application, the evidence gap vector generation unit includes: The necessary evidence slot set generation unit is used to generate a necessary evidence slot set based on the candidate review conclusion. The coverage status determination unit is used to map the initial evidence to the evidence slot set and determine the coverage status of each necessary evidence slot in the necessary evidence slot set. The evidence gap vector generation subunit is used to generate the evidence gap vector of the long document to be reviewed based on the coverage status.

[0015] In some embodiments of this application, the long document review module includes: The path type constraint generation unit is used to generate path type constraints based on the type of the evidence gap vector and the uncovered necessary evidence slots when the initial evidence is determined to be incomplete according to the evidence gap vector. A supplementary evidence search unit is used to perform a supplementary evidence search on the anchored edge according to the path type constraint. A coverage status update unit is used to update the coverage status and the evidence gap vector according to the result of the supplementary evidence search. The evidence integrity determination unit is used to determine the evidence integrity, supporting evidence strength, rebuttal evidence strength, and residual gaps based on the updated evidence gap vector. The long document review unit is used to review the long document to be reviewed based on the initial evidence, the updated coverage status, the integrity of the evidence, the strength of the supporting evidence, the strength of the counter-evidence, and the residual gap.

[0016] In some embodiments of this application, a collaborative review device for long documents with two graphs based on evidence slot gap constraints further includes: Document structure diagram generation module, used to generate the document structure diagram; the document structure diagram generation module includes: The document granularity parsing unit is used to perform granular parsing on the long document to be reviewed and determine the document units of the long document to be reviewed; wherein, the document units include: chapters, paragraphs, clauses, tables, table units and appendices; A document structure diagram generation unit is used to generate the document structure diagram based on the document unit; wherein, the document structure diagram is used to represent the hierarchical relationship, sequential relationship, referencing relationship, implicit association relationship, and table text association relationship of the long document to be reviewed.

[0017] In some embodiments of this application, a collaborative review device for long documents with two graphs based on evidence slot gap constraints further includes: A domain knowledge graph generation module is used to generate the domain knowledge graph; the domain knowledge graph generation module includes: The entity determination unit is used to determine the entities, parameters, rules, rule versions, scope of application, exceptions, and logical constraints of the long document to be reviewed based on the rule basis, applicable conditions, and semantic relationships of the review rules. The domain knowledge graph generation unit is used to generate the domain knowledge graph based on the entity, the parameter, the rule, the rule version, the scope of application, the exception conditions, and the logical constraints.

[0018] Thirdly, this application provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of a long document dual-graph collaborative review method based on evidence slot gap constraints.

[0019] Fourthly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a long document dual-graph collaborative review method based on evidence slot gap constraints.

[0020] Fifthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a long document dual-graph collaborative review method based on evidence slot gap constraints.

[0021] As described above, this application provides a method and apparatus for collaborative review of long documents based on evidence slot gap constraints. First, initial evidence for the long document to be reviewed is recalled based on a pre-generated document structure graph and domain knowledge graph of the long document to be reviewed. Anchor edges are set between the document structure graph and the domain knowledge graph. Next, an evidence gap vector of the long document to be reviewed is generated based on the initial evidence and the review task. Finally, the long document to be reviewed is reviewed based on the initial evidence and the evidence gap vector.

[0022] In summary, the method provided in this application can simultaneously characterize the structural relationships of long documents, domain rule constraints, and the evidence requirements for candidate conclusions. When faced with complex review tasks that span chapters, tables, and rules, it can accurately identify evidence gaps and provide targeted supplementation of key evidence, thereby improving the accuracy, interpretability, and reliability of long document review conclusions. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating a collaborative review method for long documents based on evidence slot gap constraints, provided in an embodiment of this application; Figure 2A second flowchart illustrating a collaborative review method for long documents based on evidence slot gap constraints, provided in an embodiment of this application; Figure 3 A flowchart illustrating step 400 provided in an embodiment of this application; Figure 4 A schematic diagram of a third process for a collaborative review method of two graphs in long documents based on evidence slot gap constraints, provided for an embodiment of this application; Figure 5 A flowchart illustrating step 500 provided in an embodiment of this application; Figure 6 A flowchart illustrating step 200 provided in an embodiment of this application; Figure 7 A flowchart illustrating step 202 provided in an embodiment of this application; Figure 8 A flowchart illustrating step 300 provided in an embodiment of this application; Figure 9 A flowchart illustrating a collaborative review method for long documents based on evidence slot gap constraints, provided as a specific application example of this application; Figure 10 A technical roadmap for a collaborative review method for long documents based on evidence slot gap constraints, provided for a specific application example of this application; Figure 11 This is a schematic diagram of a long document dual-image collaborative review device based on evidence slot gap constraints provided in this application embodiment; Figure 12 This is a schematic diagram of the evidence gap vector generation module 20 provided in the embodiments of this application; Figure 13 This is a schematic diagram of the structure of the evidence gap vector generation unit 20b provided in the embodiments of this application; Figure 14 This is a schematic diagram of the structure of the long document review module 30 provided in this embodiment of the application; Figure 15 This is a schematic diagram of a second structure of a long document dual-image collaborative review device based on evidence slot notch constraints provided in the embodiments of this application; Figure 16 This is a schematic diagram of the document structure diagram generation module 40 provided in the embodiments of this application; Figure 17 This is a schematic diagram of a third structure of a long document dual-image collaborative review device based on evidence slot notch constraints provided in this application embodiment; Figure 18 This is a schematic diagram of the domain knowledge graph generation module 50 provided in the embodiments of this application; Figure 19 This is a schematic diagram of the structure of the electronic device in the embodiments of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product or device.

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] See Figure 1 This application provides a collaborative review method for long documents based on evidence slot gap constraints, comprising the following: Step 100: Retrieve initial evidence of the long document to be reviewed based on the pre-generated document structure graph and domain knowledge graph of the long document to be reviewed; wherein, anchor edges are set between the document structure graph and the domain knowledge graph; Step 200: Generate the evidence gap vector of the long document to be reviewed based on the initial evidence and the review task; Step 300: Examine the long document to be examined based on the initial evidence and the evidence gap vector.

[0030] As can be seen from the above description, the embodiments of this application provide a collaborative review method for long documents based on evidence slot gap constraints. This method can simultaneously characterize the structural relationships of long documents, domain rule constraints, and evidence requirements for candidate conclusions. When faced with complex review tasks involving multiple chapters, tables, and rules, it can accurately identify evidence gaps and provide targeted supplementation of key evidence, thereby improving the accuracy, interpretability, and reliability of the review conclusions of long documents.

[0031] It should be noted that the initial evidence set in step 100 may still have issues such as missing structural evidence, unconfirmed rule versions, missing table bases, or unsearched exception conditions.

[0032] In addition, anchor edges are established between the document structure graph and the domain knowledge graph. These anchor edges are used to establish a correspondence between the evidence locations in the document and entities, parameters, rules, versions, scope of application, or exceptions in the domain knowledge.

[0033] The evidence gap vector in step 200 represents the degree of absence of the corresponding candidate review conclusion in dimensions such as structural evidence, substantive evidence, rule evidence, version evidence, scope of application evidence, exception condition evidence, tabular evidence, and rebuttal evidence. If a key evidence slot is not covered, the candidate review conclusion cannot be output as a definitive conclusion.

[0034] Step 300, when implemented, specifically involves: gating the candidate review conclusions based on the key evidence slot coverage status, evidence completeness, strength of supporting evidence, strength of rebuttal evidence, and remaining gaps, and outputting the final review result. The final review result includes the issue location, conclusion status, review judgment, rule basis, supporting evidence path, rebuttal evidence path, evidence gap information, and modification suggestions. The conclusion status includes a definitive conclusion, insufficient evidence, conflicting evidence, and the need for manual review.

[0035] In some embodiments of this application, see Figure 2 A collaborative review method for long documents based on evidence slot gap constraints, comprising two graphs, further comprising: Step 400: Generate the document structure diagram; then, see... Figure 3 Step 400 includes: Step 401: Perform multi-granularity parsing on the long document to be reviewed to determine the document units of the long document to be reviewed; wherein, the document units include: chapters, paragraphs, clauses, tables, table units, and appendices; Specifically, long documents awaiting review Parsed as a collection of document units that retain the original organizational structure: (1) in, u i Indicates the first iEach document unit.

[0036] Each document unit should include at least the following attributes: unit number, unit type, hierarchy information, title, body text, location, parent unit, page number, table coordinates, and semantic representation. This can be represented as: (2) in, id i Number the document unit. For document unit type, For hierarchy depth, h i For title or table heading, x i For the original text, p i This is the location of the full text. parent i Number the parent document unit. m i For metadata such as page numbers, serial numbers, and table coordinates, e i It is a semantic vector representation.

[0037] This analysis preserves the main text judgments, table values, appendix parameters, and rule application conditions as different types of evidence objects that can be located and traced.

[0038] Step 402: Generate the document structure diagram based on the document unit; wherein, the document structure diagram is used to represent the hierarchical relationship, sequential relationship, referencing relationship, implicit association relationship, and table text association relationship of the long document to be reviewed.

[0039] After completing multi-granularity parsing, a document structure graph is constructed based on the set of document units. This document structure graph is used to express the organizational relationships, reading order, citation relationships, and table relationships within long documents. This document structure graph is not only used to retrieve adjacent text, but also to track the chapter sources, table sources, and appendix sources upon which a conclusion depends.

[0040] The document structure diagram is shown as follows: (3) in, A collection of document structure nodes. For the document structure edge set, This is a collection of attributes for nodes and edges. Document structure nodes are obtained by mapping document units: (4) in, Indicates the first i Each document unit This indicates that the document unit is mapped. The obtained document structure nodes, U This represents the set of document units obtained after multi-granularity parsing.

[0041] Document structure edges include hierarchical containment edges, sequential connection edges, explicit reference edges, implicit reference edges, and table text association edges, which can be represented as: (5) in, This indicates that the hierarchy contains a set of edges. Represents the set of sequentially connected edges. This indicates an explicit reference to the edge set. This represents the set of implicitly referenced edges. This represents the set of edges associated with the table text.

[0042] Hierarchical and sequential edges are directly established by parent-child relationships and adjacency relationships within the original text; explicit reference edges are established by anchor points such as "see Chapter X" and "as shown in Table Y"; implicit reference edges are determined by a comprehensive score based on the entity, parameter, unit, semantics, chapter distance, and document order consistency of candidate unit pairs. Established in time; the table text association edge is used in the overall score of the name, value, unit, row and column, and table title of the text and table cells. Established in a timely manner. Each side type is available for subsequent supplementary certification, selectively opened based on the gaps.

[0043] In some embodiments of this application, see Figure 4 A collaborative review method for long documents based on evidence slot gap constraints, comprising two graphs, further comprising: Step 500: Generate the domain knowledge graph; then, see... Figure 5 Step 500 includes: Step 501: Determine the entities, parameters, rules, rule versions, scope of application, exceptions, and logical constraints of the long document to be reviewed based on the rule basis, applicable conditions, and semantic relationships of the review rules; Step 502: Generate the domain knowledge graph based on the entity, the parameter, the rule, the rule version, the scope of application, the exception conditions, and the logical constraints.

[0044] Document structure diagrams answer the question, "Where is the evidence located in the document, and which chapters or tables is it related to?" Domain knowledge graphs answer the question, "What entities, parameters, rules, versions, scope of application, and exceptions does the evidence involve?" Both organize evidence from the dimensions of document structure and domain semantics, respectively.

[0045] Let the rule knowledge base be... ,in, r k Indicates the first kA rule object. Each rule object can be represented as: (6) in, rid k For rule numbering, c k For the content of the rules and regulations, scope k For the scope of application, ver k For the rules version, Effective date The expiration or obsolescence date. As an exception, These can be logical constraints or numerical constraints. This is the weight of the rule's authority.

[0046] Domain knowledge graph representation is as follows: (7) in, V K A collection of domain knowledge nodes. E K For the set of domain knowledge edges, A K It is a set of attributes for domain knowledge nodes and edges.

[0047] Entity rule mapping edges in mapping score Established in a timely manner. Explicit modeling of rule version validity and applicability: (8) (9) in, For indicator functions, t q To review the time corresponding to the task, when the rule When no expiration time is set, let , Scoring is given based on the applicability of the matching object, scene, region, and condition. When It is not used as a definitive basis, and the version gap or the scope of application gap is increased in the gap calculation.

[0048] In some embodiments of this application, a collaborative review method for long documents with two graphs based on evidence slot gap constraints further includes: Establish the relationship between the document structure diagram and the domain knowledge graph.

[0049] Specifically, after constructing the document structure graph and the domain knowledge graph, it is necessary to establish the relationship between the two graphs. If a domain entity, parameter, or rule originates from a document unit, an anchor edge is established between the corresponding document structure node and the domain knowledge node. The anchor edge is represented as follows: (10) in, Represents document structure nodes, This represents entity nodes, parameter nodes, rule nodes, version nodes, scope nodes, or exception condition nodes in the domain knowledge graph. Anchoring strength is calculated based on entity occurrence strength, location matching degree, node type consistency, and contextual consistency. (11) When the anchoring strength reaches a preset threshold, the anchoring relationship is retained: (12) in, B This represents the set of anchor edges between the document structure graph and the domain knowledge graph.

[0050] By using dual-graph anchoring, two types of evidence paths can be simultaneously accessed from the text fragment under review: one path searches for relevant chapters, tables, appendices, and citations along the document structure graph; the other path searches for relevant entities, parameters, rules, versions, and exceptions along the domain knowledge graph. These two paths complement each other, ensuring that candidate conclusions no longer rely solely on single text similarity scores.

[0051] In some embodiments of this application, step 100 includes the following: For target unit u t ,exist G D take the radius h D Internal structural evidence anchored node set Enter G K Take radius h K semantic evidence within The ranking is based on a combination of text similarity, entity mapping strength, applicability, and authority. Obtain TopK rule evidence The initial set of evidence can be represented as:

[0052] In some embodiments of this application, see Figure 6 Step 200 includes: Step 201: Generate candidate review conclusions based on the initial evidence and review tasks; Specifically, the target document unit, initial evidence set, and review task are organized into structured input: (13) in, q This indicates the description of the review task, the type of review rule, or the review question entered by the user. This represents the evidence serialization function, used to organize target text, structural evidence, semantic evidence, rule-based evidence, version information, scope of application, and exceptions into inputs that the review model can process.

[0053] The review model outputs a set of candidate conclusions: (14) in, Indicates the review model, Indicates targeting a unit of document u t The generated first j Each candidate review conclusion includes the issue location, issue type, involved entities, involved rules, preliminary judgment, explanation of reasons, and suggested modifications, which can be represented as: (15) It should be noted that the candidate conclusions are only preliminary judgments formed under the current evidentiary conditions. Since the initial evidence set may still have issues such as missing structural evidence, unconfirmed rule versions, omission of table bases, or failure to retrieve exception conditions, the candidate conclusions cannot be directly output as the final results, but must enter the evidence verification archive and gap diagnosis stage.

[0054] Step 202: Generate an evidence gap vector for the long document to be reviewed based on the candidate review conclusion; wherein the evidence gap vector is used to characterize the degree of lack of the candidate review conclusion in terms of structural evidence, substantive evidence, rule evidence, version evidence, scope of application evidence, exception condition evidence, tabular evidence, and rebuttal evidence.

[0055] The recalled evidence is mapped to the necessary evidence slots. The coverage status is determined based on the matching function of each evidence slot. The evidence gap vector is calculated according to the dimensions of structural evidence, rule evidence, version evidence, scope of application evidence, tabular evidence, exception condition evidence, and counter-evidence.

[0056] In some embodiments of this application, see Figure 7 Step 202 includes: Step 2021: Generate a set of necessary evidence slots based on the candidate review conclusions; Based on the question type, involved entities, and associated rules of the candidate review conclusion, the corresponding evidence slot template is invoked, and placeholders such as objects, parameters, rules, versions, scope of application, and exceptions in the template are instantiated to generate the necessary evidence slot set corresponding to the candidate conclusion.

[0057] Specifically, an evidence verification file is first generated: for each candidate conclusion, a corresponding evidence verification file is generated. The evidence verification file is used to record the supporting evidence on which the candidate conclusion depends, the association rules, the rule version, the scope of application, the exception conditions, the path of rebuttal, the evidence gap, and the current verification status.

[0058] The evidence verification file is represented as follows: (16) in, Number the candidate conclusions. For the target document unit, To support the evidence path set, For the set of paths to prove by contradiction, For a set of association rules, For the rule version verification results, For the scope of application verification results, For results of the exception condition search, For the evidence gap vector, Scoring for the completeness of evidence, This is the status for verifying the conclusion.

[0059] The evidence verification archive transforms the natural language candidate conclusions generated by the model into computable, traceable, and updatable evidence objects. Subsequent processes do not require the model to self-reflect, but rather involve external evidence verification based on the evidence status recorded in the evidence verification archive.

[0060] Next, maintain the evidence slot template library Lib, indexed by question type. Given the question type of candidate conclusions... First Search template Obtain the slot skeleton, then use , Instantiate the objects, parameters, and rule placeholders in the skeleton, and output the necessary evidence slot set: (17) Each evidence slot can be represented as: (18) in, Indicates the name of the evidence slot. Indicates the gap category to which the evidence slot belongs. This indicates whether the evidence slot is a critical evidence slot. This indicates the types of graph paths that the evidence slot is allowed to match. This represents the matching function for the evidence slot. This indicates the weight of the evidence slot.

[0061] Matching function by In specific terms: Numerical slots (parameter value slots) use numerical comparison and unit normalization, while version slots are called. Scope of application: slot invocation The structure / table slots use path reachability and name-row-column matching. Taking the parameter threshold class as an example, Lib provides object entity slots, parameter name slots, parameter value slots, unit slots, table source slots, rule threshold slots, version slots, scope slots, exception condition slots, and proof-of-contrast check slots.

[0062] Step 2022: Map the initial evidence to the evidence slot set, and determine the coverage status of each necessary evidence slot in the necessary evidence slot set; For the archives P j Any one of the evidence items Its counterpart to the groove The coverage is: (19) in, For the first k The coverage threshold of each evidence slot is determined by the maximum value of the evidence item. .

[0063] The evidence slot coverage status in step 2022 can be determined using binary coverage, continuous coverage score, weighted coverage score, or multi-threshold coverage. The gap vector can be calculated according to dimensions such as structure, rules, version, scope of application, table, exception conditions, and evidence against contradiction, and gap dimensions can be added or removed according to specific industry scenarios.

[0064] Step 2023: Generate the evidence gap vector of the long document to be reviewed based on the coverage status.

[0065] By category d Aggregate the uncovered weights to obtain the gaps in each dimension: (20) in, This is a smoothing term. Therefore, the candidate review conclusion is obtained. c j Evidence gap vector: (twenty one) in, Indicates a structural gap. Indicates a gap in the entity. Indicates a rule gap. Indicates a version gap. Indicates a gap in the scope of application. This indicates an exception condition gap. Indicates a gap in the table of evidence. This indicates a gap for rebuttal.

[0066] The set of critical gaps can be represented as: (twenty two) when It is prohibited to directly output definitive conclusions.

[0067] In some embodiments of this application, see Figure 8 Step 300 includes: Step 301: If the initial evidence is determined to be incomplete based on the evidence gap vector, generate path type constraints based on the type of the evidence gap vector and the uncovered necessary evidence slots; Step 302: Perform a supplementary verification search on the anchored edge according to the path type constraint; Specifically, based on the evidence gap vector and the path type constraint generated by the uncovered key evidence slots, candidate supplementary evidence paths are searched on the document structure graph, the domain knowledge graph and their anchor edges, and the supplementary evidence path is selected with the gap reduction benefit, and the evidence slot coverage status and gap vector are updated.

[0068] The supplementary evidence search can employ methods such as bundle search, heuristic search, shortest path search, Top-K path search, greedy search, or reinforcement learning-based path selection. As long as the selection of the supplementary evidence path includes an evaluation of the benefit of reducing the evidence gap, the purpose of supplementary evidence in this invention can be achieved.

[0069] Step 303: Update the coverage state and the evidence gap vector based on the results of the supplementary evidence search; In steps 301 to 303, when the evidence gap vector indicates that the current evidence is incomplete, firstly, path type constraints are generated based on the gap type and the uncovered key evidence slots, and a supplementary evidence search is performed on the document structure graph, the domain knowledge graph, and the anchor edges between them. Next, with the goal of reducing the evidence gap, candidate supplementary evidence paths are evaluated, and evidence paths that can reduce the current evidence gap, fill in the key evidence slots, and have low redundancy are prioritized and added to the evidence verification archive. Finally, after the supplementary evidence is completed, the evidence slot coverage status and the evidence gap vector are updated.

[0070] During the supplementary evidence process, a counter-evidence search is performed based on the question type, involved entities, associated rules, and evidence gap status of the candidate review conclusion to find evidence that may overturn, limit, or modify the candidate review conclusion. If valid counter-evidence exists, the candidate review conclusion is revoked, marked as conflicting, subject to conditional restrictions, or downgraded according to the strength of the counter-evidence, rule version, scope of application, and source of evidence.

[0071] Step 304: Determine the completeness of evidence, strength of supporting evidence, strength of rebuttal evidence, and residual gaps based on the updated evidence gap vector; Step 305: Review the long document to be reviewed based on the initial evidence, the updated coverage status, the integrity of the evidence, the strength of the supporting evidence, the strength of the counter-evidence, and the residual gap.

[0072] Specifically, based on the coverage status of key evidence slots, the completeness of evidence, the strength of supporting evidence, the strength of rebuttal evidence, and remaining gaps, the candidate review conclusions are gating-based, and the final review result is output. The final review result includes the issue location, conclusion status, review judgment, rule basis, supporting evidence path, rebuttal evidence path, evidence gap information, and modification suggestions. The conclusion status includes definitive conclusion, insufficient evidence, conflicting evidence, and need for manual review.

[0073] In some embodiments of this application, the evidence integrity scoring step 304 includes: calculating a comprehensive evidence integrity score based on the evidence gap vector. (twenty three) in, Represents the set of evidence gap types. Candidate conclusions c j In the d Gap value under the evidence dimension This is the weight for this type of gap.

[0074] when When this occurs, it indicates that the candidate conclusion meets the basic conditions for entering the final verification; when If this occurs, it indicates that the current chain of evidence is still incomplete, and the process enters the gap-driven targeted supplementary evidence process.

[0075] In some embodiments of this application, the gap-driven supplementary proof and disproven proof verification steps in step 304 include: First, supplementary proof using a dual-graph path with the goal of reducing the gap gain: Path type mask: composed of gap vector and key gap K j Generate mask: (twenty four) According to the uncovered groove path k Take the union of sets and sort them by... Sort by size and category, and open the corresponding node and edge types: If the version gap is high, open rule version edge, alternative rule edge, and effective time node; if the table evidence gap is high, open table text association edge, table row / cell node, and appendix parameter node; if the exception gap is high, open exception condition edge, special working condition and manual review condition node; if the structure gap is high, open hierarchy, order, explicit and implicit reference edge. M j Limit the set of types that can be accessed in this round.

[0076] Constrained path search: In M j Under constraints, starting from the nodes already covered by the archives, in G D , G K With anchoring edge B Perform bounded depth priority expansion (only along) M j Allowed type expansion, sorted by edge weight and expected contribution to the target slot (bundle search), yielding a set of candidate supplementary proof paths: (25) in, Indicates the first t Evidence verification files prior to supplementary evidence submission. Indicates the first t A set of candidate supplementary evidence paths generated in rounds.

[0077] Profit estimation: for path The endpoint evidence is hypothetically added to the archive, and the coverage of each slot is recalculated according to S3.2 to obtain the predicted gap. Expected reduction in the gap: (26) The overall benefit is: (27) in, This indicates the path completeness of the candidate supplementary evidence path. This indicates the potential of candidate supplementary proof paths to discover disproven evidence or limiting conditions. This indicates the redundancy between the candidate supplementary evidence path and the current evidence verification file. This represents the context cost occupied by the candidate supplementary evidence path. to These are the weighting coefficients.

[0078] Greedy optimization and updating: In each round, select the path with the maximum marginal benefit and update the file. (27) (28) according to Add evidence refresh slot coverage, and The final stopping condition can be expressed as: (29) It is understandable that this process can transform the supplementary proof into a bi-graph path selection problem under mask constraints, with gap reduction as the objective function.

[0079] In some embodiments of this application, evidence gain and context reconstruction include the following: after each round of supplementary evidence, the newly added evidence gain is calculated: (30) in, This represents the evidence completeness score after the t-th round of supplementary evidence.

[0080] Whether to continue supplementing evidence is determined based on the gain of newly added evidence, the completeness score of evidence, and the maximum number of expansion rounds. After supplementing evidence, the evidence set also needs to be sorted, deduplicated, and compressed. Due to the limited input length of the large language model or the censorship reasoning module, it is not possible to simply input all the expanded evidence into the model. Instead, priority should be given to retaining evidence that is highly relevant to the candidate conclusion, has authoritative sources, complete paths, relatively new rules, and low redundancy.

[0081] The evidence value function can be expressed as: (31) in, Candidate evidence items can be document structure nodes, domain knowledge nodes, rule nodes, or evidence paths. This indicates the relevance of the evidence to the candidate conclusion. This indicates the authority of the evidence source. Indicates the completeness of the evidence path. Indicates the freshness of rules or data. Indicates the redundancy of evidence.

[0082] Within the context budget, select a subset of high-value evidence and reconstruct the model input. This avoids interference with review and judgment caused by introducing a large amount of weakly relevant content during the supplementary evidence process.

[0083] In some embodiments of this application, the counter-evidence retrieval and conflict verification include the following: according to , , , Generate reverse evidence slot set : (32) Constraints along its path Search for counter-evidence candidates N j .right The strength of the counter-evidence is calculated based on semantic contradictions, exception hits, inconsistent scope of application, version conflicts, unit conversion conflicts, and critical slot coverage. ,Pick Write .

[0084] If valid counter-evidence hits a key reverse slot and its version / scope / source priority is higher than the supporting path, it is revoked or marked as conflicting; if it only constitutes a conditional limitation, it is retained with applicable conditions attached; if the evidence is insufficient, it is downgraded.

[0085] In some embodiments of this application, the conclusion gating based on key evidence slots in step 305 includes: performing hard gating of key evidence slots before calculating the credibility of the conclusion. For candidate examination conclusions... c j If any uncovered evidence slots exist within the critical evidence slot set, the candidate review conclusion is not allowed to be output as a definitive conclusion. The hard gating of critical evidence slots is represented as follows: (33) in, Indicates the candidate review conclusion c j The set of uncovered key evidence slots.

[0086] After passing the hard gate of the critical evidence slot, calculate the credibility of the conclusion: (34) in, Scoring for the completeness of evidence, To support the strength of the evidence, For effective counter-evidence strength, For the strength of the remaining critical gap.

[0087] It should be noted that conclusion gating can be implemented using rule-based thresholds, scoring models, classification models, or a combination of rules and models. The output status can also be set to "definite," "insufficient evidence," "conflict," "manual review," or other equivalent statuses according to business needs.

[0088] In some embodiments of this application, the output state is determined according to the following rules: 1. When , , At that time, output a deterministic conclusion; 2. When If the key evidence slot is not covered and belongs to the rule version slot, scope of application slot, table source slot or exception condition slot, output the manual review conclusion. 3. When a valid path of rebuttal exists, and the valid path of rebuttal cannot be distinguished from the path of supporting evidence by rule version, scope of application, object type, or time condition, output a conflict conclusion; 4. When the evidence completeness score is below the threshold, but the uncovered evidence slot is not a key evidence slot, output the conclusion of insufficient evidence and list the missing evidence slots; 5. When the evidence-rebuttal path only restricts the conditions for forming the candidate review conclusion, output a definitive conclusion with applicable conditions or a conclusion of insufficient evidence.

[0089] In some embodiments of this application, the determination of the conclusion status in step 305 includes: determining the final status of the candidate conclusion based on gating conditions. When the evidence integrity score reaches a threshold, the supporting evidence strength reaches a threshold, the rebuttal strength is lower than the conflict threshold, and the conclusion credibility reaches a threshold, the final review result is output: (35) in, Number the conclusions. For the location of the problem, This is the conclusion. For the purpose of review and judgment, As the basis for the rules, To support the evidence path, For the path of proof by contradiction, For the evidence gap vector, To ensure the credibility of the conclusion, Suggestions for revision.

[0090] In some embodiments of this application, the conclusion status This includes conclusions that are definitive, lack sufficient evidence, involve conflicts, or require manual review. For definitive non-compliance conclusions, output the location of the problem, the reason for the error, the rule basis, the evidence path, and suggested modifications. For conclusions with insufficient evidence, output the type of missing evidence and suggested supplementary information. For conflicting conclusions, output supporting evidence paths, counter-evidence paths, and the reasons for the conflict. For conclusions requiring manual review, output a checklist for manual review.

[0091] As described above, this application provides a collaborative review method for long documents based on evidence slot gap constraints using a dual-graph approach. By constructing a document structure graph and a domain knowledge graph, necessary evidence slots are generated for candidate review conclusions. An evidence gap vector is calculated based on the evidence slot coverage state. A dual-graph path supplementary evidence search is performed with the goal of reducing the benefit of evidence gaps, thereby completing evidence supplementation, verifying counter-evidence, and gating the conclusions of candidate review conclusions. This application solves the problems in existing long document review methods, such as candidate conclusions relying on initial recall evidence, difficulty in locating missing evidence, unclear supplementary evidence directions, and a lack of sufficient evidence constraints in conclusion output. It achieves the effect of improving the evidence completeness, traceability, accuracy, and reliability of long document review conclusions. Specifically, compared with the prior art, this application has the following beneficial effects: Improve the integrity and traceability of evidence organization in long documents: By constructing document structure diagrams and domain knowledge graphs, chapters, paragraphs, tables, appendices, and citation relationships in long documents are associated with domain knowledge such as entities, parameters, rules, versions, scope of application, and exceptions, so that review evidence can be traced back to the original text and correspond to specific rules and bases.

[0092] Improve the accuracy of evidence gap location: Calculate the evidence gap vector based on the evidence slot coverage status, which can clearly identify the missing information of the current candidate conclusion in dimensions such as structure, entity, rule, version, scope of application, exception conditions, table and counter-evidence, so that the evidence insufficiency has a calculable and locationable expression.

[0093] Enhance the relevance and effectiveness of the supplementary evidence process: With the goal of reducing the benefits of evidence gaps, perform supplementary evidence path search on the document structure graph, domain knowledge graph and the anchor edges between the two, so as to prioritize the supplementation of evidence that plays a key role in the establishment, refuting or downgrading of the conclusion, and reduce the introduction of weakly relevant evidence.

[0094] Improve the reliability and traceability of the final review results: By using key evidence slot gating, rebuttal evidence retrieval and conflict verification, the candidate review conclusions are determined as definitive conclusions, insufficient evidence, conflicting, or requiring manual review. Supporting evidence paths, rebuttal evidence paths and gap information are output, thereby improving the accuracy, reliability and interpretability of long document review results.

[0095] For further explanation of this plan, please refer to [link / reference]. Figure 9 as well as Figure 10 This application also provides a specific implementation of a collaborative review method for long documents based on evidence slot gap constraints, which includes the following:

[0096] S1: Construct a document structure graph and a domain knowledge graph.

[0097] First, the long document to be reviewed is parsed at multiple granularities to obtain document units such as chapters, paragraphs, clauses, tables, table cells, and appendices, while retaining the hierarchy, position, title, body text, page number, table coordinates, and semantic representation of each document unit. Based on these document units, a document structure diagram is constructed to represent the hierarchical relationships, sequential relationships, referencing relationships, implicit relationships, and table text relationships within the document.

[0098] Next, a domain knowledge graph is constructed. This graph includes knowledge nodes such as entities, parameters, rules, rule versions, scope of application, exceptions, and logical constraints, used to represent the rule base, applicable conditions, and semantic relationships involved in the review process. Anchor edges are established between the document structure graph and the domain knowledge graph, creating a correspondence between the evidence locations in the document and the entities, parameters, rules, versions, scope of application, or exceptions in the domain knowledge graph, thus obtaining a dual-graph evidence base.

[0099] Document structure graphs and domain knowledge graphs can be implemented using graph databases, relational databases, adjacency lists, inverted indexes, and vector indexes, as long as they can express the structural relationships between document units and the correspondence between document evidence and entities, parameters, rules, versions, scope of application, and exceptions. This application is not limited to these methods.

[0100] S2: Retrieve initial evidence based on the document structure diagram and domain knowledge graph, and generate candidate review conclusions in conjunction with the review task.

[0101] Each candidate review conclusion should include at least the location of the issue, the type of issue, the entities involved, the relevant rules, the preliminary judgment, the explanation of the reasons, and the suggested modifications. Candidate review conclusions are used for subsequent evidence verification.

[0102] S3: Based on the question type, involved entities, and associated rules of the candidate review conclusion, generate a set of necessary evidence slots corresponding to the candidate review conclusion.

[0103] Necessary evidence slots describe the evidentiary elements required for a candidate review conclusion to be established, overturned, or downgraded. These include one or more of the following: object entity slots, parameter value slots, unit slots, table source slots, rule basis slots, rule version slots, scope of application slots, exception condition slots, and rebuttal check slots. Different issue types correspond to different evidence slot templates. Based on the information in the candidate review conclusion, the evidence slot template entities and rules are instantiated to obtain the set of necessary evidence slots for that candidate review conclusion.

[0104] It should be noted that evidence slots can be generated through a preset template library, or through rule configuration tables, domain ontology, prompt word templates, or model-assisted generation. Any method that can determine the necessary evidentiary elements required for the candidate review conclusion based on the question type, involved entities, and associated rules is considered an equivalent implementation.

[0105] S4: Map the recalled document evidence, rule evidence, table evidence, version evidence, scope of application evidence, exception condition evidence, and counter-evidence evidence to the necessary evidence slots, and determine the coverage status of each necessary evidence slot.

[0106] Based on the coverage status of each necessary evidence slot, calculate the evidence gap vector for the candidate review conclusion. The evidence gap vector represents the degree of absence of the candidate review conclusion in dimensions such as structural evidence, substantive evidence, rule evidence, version evidence, scope of application evidence, exception condition evidence, tabular evidence, and rebuttal evidence. If a key evidence slot is not covered, the candidate review conclusion cannot be output as a definitive conclusion.

[0107] S5: If the initial evidence is determined to be incomplete based on the evidence gap vector, a supplementary evidence search is performed between the document structure graph and the domain knowledge graph by anchoring edges.

[0108] Specifically, when the evidence gap vector indicates that the current evidence is incomplete, path type constraints are generated based on the gap type and the uncovered key evidence slots. Supplementary evidence search is then performed on the document structure graph, the domain knowledge graph, and the anchor edges between them. With the goal of reducing evidence gaps, candidate supplementary evidence paths are evaluated, and those that can reduce the current evidence gaps, fill in key evidence slots, and have low redundancy are prioritized for addition to the evidence verification archive. After supplementary evidence is completed, the evidence slot coverage status and the evidence gap vector are updated.

[0109] During the supplementary evidence process, a counter-evidence search is performed based on the question type, involved entities, associated rules, and evidence gap status of the candidate review conclusion to find evidence that may overturn, limit, or modify the candidate review conclusion. If valid counter-evidence exists, the candidate review conclusion is revoked, marked as conflicting, subject to conditional restrictions, or downgraded according to the strength of the counter-evidence, rule version, scope of application, and source of evidence.

[0110] S6: Based on the coverage status of key evidence slots, the integrity of evidence, the strength of supporting evidence, the strength of rebuttal evidence, and the remaining gaps, perform gating judgment on the candidate review conclusions and output the final review results.

[0111] The final review results include the location of the issue, the status of the conclusion, the review judgment, the basis of the rules, the path of supporting evidence, the path of rebuttal evidence, information on evidence gaps, and suggestions for modification. The status of the conclusion includes a definitive conclusion, insufficient evidence, conflicting evidence, and the need for manual review.

[0112] As described above, the specific implementation of this application provides a collaborative review method for long documents based on evidence slot gap constraints, comprising: First, constructing a dual-graph evidence base: constructing a document structure graph and a domain knowledge graph, and associating document evidence positions with domain knowledge nodes through anchor edges; Next, generating candidate review conclusions: generating candidate review conclusions containing question types, involved entities, and association rules based on target document units, dual-graph recalled evidence, and review tasks; Then, generating necessary evidence slots: instantiating and generating corresponding necessary evidence slot sets according to the question type, involved entities, and association rules of the candidate review conclusions; Calculating evidence gap vectors: mapping existing evidence to necessary evidence slots, and calculating gap vectors in dimensions such as structure, rules, version, scope of application, tables, exception conditions, and counter-evidence based on slot coverage status; Performing dual-graph path supplementary evidence search: generating path type constraints based on evidence gap vectors, and selecting supplementary evidence paths on the document structure graph and domain knowledge graph with the goal of reducing gap gains; Finally, performing counter-evidence verification and conclusion gating: retrieving counter-evidence evidence, and determining the output status of candidate conclusions based on key evidence slot coverage status, evidence integrity, supporting evidence, and counter-evidence strength. The method provided in this application improves the accuracy, interpretability, and reliability of the review conclusions for long documents.

[0113] Based on the same inventive concept, this application also provides a collaborative review device for long documents with two images based on evidence slot gap constraints, which can be used to implement the method described in the above embodiments, as shown in the following embodiments. Since the principle of the collaborative review device for long documents with two images based on evidence slot gap constraints is similar to that of the collaborative review method for long documents with two images based on evidence slot gap constraints, the implementation of the collaborative review device for long documents with two images based on evidence slot gap constraints can refer to the implementation of the collaborative review method for long documents with two images based on evidence slot gap constraints, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0114] This application provides a specific implementation of a long document dual-image collaborative review device based on evidence slot gap constraints, capable of implementing a long document dual-image collaborative review method based on evidence slot gap constraints. See [link to implementation details]. Figure 11 A collaborative review device for long documents based on evidence slot gap constraints includes the following components: The initial evidence recall module 10 is used to recall the initial evidence of the long document to be reviewed based on the pre-generated document structure diagram and domain knowledge graph of the long document to be reviewed; wherein, an anchor edge is set between the document structure diagram and the domain knowledge graph; Evidence gap vector generation module 20 is used to generate an evidence gap vector for the long document to be reviewed based on the initial evidence and the review task. The long document review module 30 is used to review the long document to be reviewed based on the initial evidence and the evidence gap vector.

[0115] In some embodiments of this application, see Figure 12 The evidence gap vector generation module 20 includes: The candidate review conclusion generation unit 20a is used to generate candidate review conclusions based on the initial evidence and the review task. The evidence gap vector generation unit 20b is used to generate an evidence gap vector for the long document to be reviewed based on the candidate review conclusion; wherein, the evidence gap vector is used to characterize the degree of lack of the candidate review conclusion in terms of structural evidence, substantive evidence, rule evidence, version evidence, scope of application evidence, exception condition evidence, tabular evidence, and rebuttal evidence.

[0116] In some embodiments of this application, see Figure 13 The evidence gap vector generation unit 20b includes: Necessary evidence slot set generation unit 20b1 is used to generate a necessary evidence slot set based on the candidate review conclusion; The coverage status determination unit 20b2 is used to map the initial evidence to the evidence slot set and determine the coverage status of each necessary evidence slot in the necessary evidence slot set; Evidence gap vector generation subunit 20b3 is used to generate the evidence gap vector of the long document to be reviewed based on the coverage status.

[0117] In some embodiments of this application, see Figure 14 The long document review module 30 includes: The path type constraint generation unit 30a is used to generate path type constraints based on the type of the evidence gap vector and the uncovered necessary evidence slots when the initial evidence is determined to be incomplete according to the evidence gap vector. The supplementary evidence search unit 30b is used to perform supplementary evidence search on the anchoring edge according to the path type constraint. Coverage status update unit 30c is used to update the coverage status and the evidence gap vector according to the result of the supplementary evidence search. The evidence integrity determination unit 30d is used to determine the evidence integrity, supporting evidence strength, rebuttal evidence strength, and residual gaps based on the updated evidence gap vector. The long document review unit 30e is used to review the long document to be reviewed based on the initial evidence, the updated coverage status, the integrity of the evidence, the strength of the supporting evidence, the strength of the counter-evidence, and the residual gap.

[0118] In some embodiments of this application, see Figure 15 A collaborative review device for long documents with two diagrams based on evidence slot gap constraints, further comprising: Document structure diagram generation module 40 is used to generate the document structure diagram; see also Figure 16 The document structure diagram generation module 40 includes: The document granularity parsing unit 40a is used to perform granular parsing on the long document to be reviewed and determine the document units of the long document to be reviewed; wherein, the document units include: chapters, paragraphs, clauses, tables, table units and appendices; The document structure diagram generation unit 40b is used to generate the document structure diagram based on the document unit; wherein, the document structure diagram is used to represent the hierarchical relationship, sequential relationship, referencing relationship, implicit association relationship, and table text association relationship of the long document to be reviewed.

[0119] In some embodiments of this application, see Figure 17 A collaborative review device for long documents with two diagrams based on evidence slot gap constraints, further comprising: Domain knowledge graph generation module 50 is used to generate the domain knowledge graph; see also Figure 18 The domain knowledge graph generation module 50 includes: The entity determination unit 50a is used to determine the entities, parameters, rules, rule versions, scope of application, exceptions and logical constraints of the long document to be reviewed based on the rule basis, applicable conditions and semantic relationships of the review rules. Domain knowledge graph generation unit 50b is used to generate the domain knowledge graph based on the entity, the parameter, the rule, the rule version, the scope of application, the exception conditions, and the logical constraints.

[0120] This application also provides a specific implementation of an electronic device capable of implementing all steps in the long document dual-graph collaborative review method based on evidence slot gap constraints described in the above embodiments. See [link to implementation details]. Figure 19 The electronic devices specifically include the following: Processor 1201, memory 1202, communications interface 1203, and bus 1204; The processor 1201, memory 1202, and communication interface 1203 communicate with each other via bus 1204; the communication interface 1203 is used to realize information transmission between server-side devices, power measurement devices, and user-side devices and other related devices.

[0121] The processor 1201 is used to call the computer program in the memory 1202. When the processor executes the computer program, it implements all the steps in the long document dual-graph collaborative review method based on evidence slot gap constraint in the above embodiment. For example, when the processor executes the computer program, it implements the following steps: Initial evidence for the long document to be reviewed is retrieved based on the pre-generated document structure graph and domain knowledge graph of the long document to be reviewed; wherein, anchor edges are set between the document structure graph and the domain knowledge graph; Based on the initial evidence and the review task, generate the evidence gap vector of the long document to be reviewed; The long document to be examined is examined based on the initial evidence and the evidence gap vector.

[0122] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the long document dual-image collaborative review method based on evidence slot gap constraints in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the long document dual-image collaborative review method based on evidence slot gap constraints in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Initial evidence for the long document to be reviewed is retrieved based on the pre-generated document structure graph and domain knowledge graph of the long document to be reviewed; wherein, anchor edges are set between the document structure graph and the domain knowledge graph; Based on the initial evidence and the review task, generate the evidence gap vector of the long document to be reviewed; The long document to be examined is examined based on the initial evidence and the evidence gap vector.

[0123] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0124] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0125] While this application provides method operation steps as shown in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the method can be executed in the order shown in the embodiments or drawings or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0126] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] This application uses specific embodiments to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A collaborative review method for long documents using two diagrams based on evidence slot gap constraints, characterized in that, include: Initial evidence for the long document to be reviewed is retrieved based on the pre-generated document structure graph and domain knowledge graph of the long document to be reviewed; wherein, anchor edges are set between the document structure graph and the domain knowledge graph; Generate an evidence gap vector for the long document to be reviewed based on the initial evidence and the review task. The long document to be examined is examined based on the initial evidence and the evidence gap vector.

2. The long document dual-image collaborative review method according to claim 1, characterized in that, Based on the initial evidence and the review task, an evidence gap vector for the long document to be reviewed is generated, including: Based on the initial evidence and the review task, candidate review conclusions are generated; An evidence gap vector is generated for the long document to be reviewed based on the candidate review conclusion; wherein the evidence gap vector is used to characterize the degree of lack of the candidate review conclusion in terms of structural evidence, substantive evidence, rule evidence, version evidence, scope of application evidence, exception condition evidence, tabular evidence, and rebuttal evidence.

3. The long document dual-image collaborative review method according to claim 2, characterized in that, Based on the candidate review conclusions, an evidence gap vector for the long document to be reviewed is generated, including: A necessary set of evidence slots is generated based on the candidate review conclusions; Map the initial evidence to the set of evidence slots, and determine the coverage status of each necessary evidence slot in the set of necessary evidence slots; The evidence gap vector of the long document to be reviewed is generated based on the coverage status.

4. The long document dual-image collaborative review method according to claim 3, characterized in that, Examining the long document to be examined based on the initial evidence and the evidence gap vector includes: If the initial evidence is determined to be incomplete based on the evidence gap vector, a path type constraint is generated based on the type of the evidence gap vector and the uncovered necessary evidence slots. A supplementary verification search is performed on the anchored edge according to the path type constraint; Update the coverage status and the evidence gap vector based on the results of the supplementary evidence search; Determine the completeness of evidence, strength of supporting evidence, strength of rebuttal evidence, and residual gaps based on the updated evidence gap vector; The long document to be reviewed is examined based on the initial evidence, the updated coverage status, the integrity of the evidence, the strength of the supporting evidence, the strength of the counter-evidence, and the residual gaps.

5. The long document dual-image collaborative review method according to claim 1, characterized in that, The steps for generating the document structure diagram include: The long document to be reviewed is analyzed at a granular level to determine its document units; wherein, the document units include: chapters, paragraphs, clauses, tables, table units, and appendices; The document structure diagram is generated based on the document units; wherein, the document structure diagram is used to represent the hierarchical relationship, sequential relationship, referencing relationship, implicit association relationship, and table text association relationship of the long document to be reviewed.

6. The long document dual-image collaborative review method according to any one of claims 1 to 5, characterized in that, The steps for generating the domain knowledge graph include: The entities, parameters, rules, rule versions, scope of application, exceptions, and logical constraints of the long document to be reviewed are determined based on the rule basis, applicable conditions, and semantic relationships of the review rules. The domain knowledge graph is generated based on the entity, the parameter, the rule, the rule version, the scope of application, the exception conditions, and the logical constraints.

7. A collaborative review device for long documents with two diagrams based on evidence slot gap constraints, characterized in that, include: The initial evidence retrieval module is used to retrieve the initial evidence of the long document to be reviewed based on the pre-generated document structure diagram and domain knowledge graph of the long document to be reviewed; wherein, an anchor edge is set between the document structure diagram and the domain knowledge graph; The evidence gap vector generation module is used to generate an evidence gap vector for the long document to be reviewed based on the initial evidence and the review task. The long document review module is used to review the long document to be reviewed based on the initial evidence and the evidence gap vector.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the long document dual-graph collaborative review method based on evidence slot gap constraints as described in any one of claims 1 to 6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the long document dual-graph collaborative review method based on evidence slot gap constraints as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the long document dual-graph collaborative review method based on evidence slot gap constraints as described in any one of claims 1 to 6.