A rule engine-based retrieval enhanced generated archive auditing method and system
By using structured processing based on a rule engine and a traceable rule knowledge base, combined with semantic interpretation of a large language model, the problems of rule omissions and unstable results in the open archive review process have been solved, achieving efficient, reliable and interpretable review results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-30
AI Technical Summary
Existing methods for reviewing open archives rely on manual review, which is inefficient and labor-intensive. Furthermore, existing technologies suffer from rule omissions and unstable review results in rule-based review scenarios, making it difficult to meet the needs of high compliance scenarios.
By using a rule engine-based retrieval enhancement method, the review rules are structured and a traceable rule knowledge base is built. The rule engine is used for deterministic matching, and a large language model is combined for semantic interpretation to generate a traceable review report.
This improved the reliability, consistency, and traceability of document compliance review, reduced fluctuations caused by probabilistic reasoning in the model, and ensured the completeness of rule recall and the interpretability of review results.
Smart Images
Figure CN122309687A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of document review, and in particular to a document review method and system based on rule engine retrieval enhancement. Background Technology
[0002] With the continuous advancement of informatization and digitalization, the scale of archival resources continues to grow, and the demand for open access to archives is increasing. As a crucial aspect of archival management, the review of archival access requires examination of archival content based on numerous institutional documents and management rules to ensure that the opening of archives complies with relevant laws, regulations, and management requirements.
[0003] Current methods for reviewing open access to archives still primarily rely on manual review. Reviewers need to meticulously compare each rule against the text, resulting in low efficiency, high workload, and difficulty in ensuring consistency of review results. To improve review efficiency, some technical solutions in recent years have begun to incorporate natural language processing (NLP) and large language models to semantically understand archive content and generate supplementary review opinions. However, directly relying on large language models for rule judgment is limited by the model's contextual capacity and the uncertainty of the reasoning process, making it difficult to guarantee the stability and auditability of review results in high-compliance scenarios.
[0004] To alleviate the conflict between rule size and model context constraints, existing technologies have introduced a Retrieval-Augmented Generation (RAG) architecture. This architecture retrieves relevant content from external rule or policy databases and uses it as context input to expand the model's knowledge coverage. This approach has proven effective in applications like question answering, where semantic relevance is the primary goal. However, in rule review scenarios, rule applicability depends on whether conditions are met, rather than on its semantic similarity to document content. Top-k retrieval mechanisms based on vector similarity ranking suffer from approximation characteristics. When the number of rules is large or their expressions are diverse, they are prone to rule omissions, making it difficult to guarantee the completeness of rule applicability and thus affecting the compliance of the review results.
[0005] On the other hand, rule engines are characterized by strong determinism and good interpretability in compliance review and business decision-making. However, traditional rule systems lack flexibility when processing natural language text and are difficult to adapt to large-scale unstructured archive content. Existing solutions that combine rule engines with large language models mostly focus on constraining or validating the model output results, without systematically improving the recall method of rules in the review process, nor solving the applicability problem of similarity retrieval in rule review scenarios.
[0006] Therefore, existing technologies urgently need a new archival review technology solution that can achieve intelligent analysis and interpretation of archival content while ensuring the completeness of rule recall and auditability, so as to meet the actual needs of high-compliance application scenarios such as open archival review. Summary of the Invention
[0007] The purpose of this application is to provide a document review method and system based on rule engine retrieval enhancement, which can avoid the rule omission problem caused by similarity retrieval, improve the reliability, consistency and traceability of document compliance review, and reduce the fluctuations caused by probabilistic reasoning of the model.
[0008] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for document review based on rule engine retrieval enhancement generation. The method includes: receiving target review task identification information and initializing the review task runtime environment; performing structured processing and metadata binding on the document opening review rules corresponding to the target review task to generate multiple traceable rule units and construct a rule knowledge base; acquiring the target document text and updating the target document text through preprocessing; extracting factual information from the target document text and mapping the factual information to rule matching parameters; the factual information is used for rule condition judgment; constructing a rule matching object based on the target document text and performing rule review on the rule matching object by calling the rule engine retrieval enhancement generation mechanism through the rule matching parameters to obtain the review explanation result; summarizing the target document text, rule matching object, and review explanation result, and outputting a document opening review analysis report.
[0009] Secondly, this application also provides a document review system based on rule engine retrieval enhancement. The document review system based on rule engine retrieval enhancement utilizes the methods described above and includes: an initialization unit, which receives target review task identification information and initializes the review task's operating environment; a rule knowledge base construction unit, which performs structured processing and metadata binding on the document openness review rules corresponding to the target review task, generating multiple traceable rule units and constructing a rule knowledge base; and a preprocessing unit, which acquires the target document text and updates it through preprocessing. The system comprises: a target archive text; a rule matching parameter extraction unit, which extracts factual information from the target archive text and maps the factual information to rule matching parameters; the factual information is used for rule condition judgment; a rule engine review and interpretation unit, which constructs a rule matching object based on the target archive text and calls the rule engine retrieval enhancement generation mechanism through the rule matching parameters to review the rule matching object and obtain the review and interpretation results; and an archive opening review and analysis report generation unit, which summarizes the target archive text, rule matching object, and review and interpretation results and outputs an archive opening review and analysis report.
[0010] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application improves the interpretability and traceability of review results by reconstructing the technical division of labor between rule recall and semantic interpretation, ensuring the completeness of rule applicability while maintaining the accuracy and traceability of review results. First, it receives review tasks and initializes the runtime environment. It performs structured parsing and metadata binding on the open archive review rules, converting the rules into executable rule units containing conditional expressions and constraint descriptions. This constructs a rule knowledge base that supports independent maintenance at the unit level, eliminating the need for model retraining for rule updates and significantly reducing maintenance costs. Next, it acquires and preprocesses the target archive text, extracting factual parameters and mapping them to rule matching parameters. After constructing a rule matching object based on the archive text, it calls the rule engine to perform deterministic matching on all rule units in the rule knowledge base. The satisfaction of rule applicability conditions is used as the recall criterion to identify all matching rules, thus avoiding the risk of missed rule detections due to similarity retrieval. Finally, it constructs a rule matching object based on the target archive text and uses the rule matching parameters to call the rule engine's enhanced generation mechanism to review the rule matching object, obtaining the review interpretation result. This separates rule matching from semantic interpretation, reducing fluctuations caused by probabilistic model inference and ensuring consistency of review results. In this application, rule judgment is triggered by the rule engine, avoiding the omission of key information due to contextual limitations. Finally, a complete audit analysis report is generated, allowing each audit conclusion to be traced back to the specific rule clause and its applicable reasons. In summary, this application avoids the rule omission problem caused by similarity retrieval, improves the reliability, consistency, and traceability of document compliance audits, and reduces the fluctuations caused by probabilistic inference in the model. Attached Figure Description
[0011] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 This is a flowchart illustrating a document review method based on rule engine retrieval enhancement provided in an embodiment of this application.
[0013] Figure 2 This is a schematic diagram of the structure of a document review system based on rule engine retrieval enhancement provided in an embodiment of this application.
[0014] Figure label: 101. Initialization Unit; 102. Rule Knowledge Base Construction Unit; 103. Preprocessing Unit; 104. Rule Matching Parameter Extraction Unit; 105. Rule Engine Review and Interpretation Unit; 106. Archive Open Review and Analysis Report Generation Unit. Detailed Implementation
[0015] 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, and 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.
[0016] While existing technologies for reviewing open access to archives have incorporated deep learning models, intensive retrieval models, and large language models to assist in the judgment process, they still have limitations overall.
[0017] 1. In long text scenarios, existing solutions are prone to information loss due to the omission of key paragraphs or semantic compression, resulting in weak stability in the review of long text archives.
[0018] 2. Existing RAG technology uses semantic similarity as the recall criterion for review rules. However, the applicability of a rule depends on whether the conditions are met, rather than on its semantic similarity to the content of the file.
[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0020] Example 1, such as Figure 1 As shown in the figure, this embodiment provides a method for reviewing archives generated based on rule engine retrieval enhancement. The method includes: S1. Receive the target audit task identifier information and initialize the audit task runtime environment.
[0021] Furthermore, step S1 specifically includes: S11. Receive the target review task identifier information and determine the set of file objects associated with the target review task and the identifier of the applicable rule set.
[0022] S12. Load the process environment for the target review task; the process environment is used to display the rule matching objects and the review interpretation results.
[0023] S13. Configure permissions for viewing, referencing, and recording audit interpretation results.
[0024] S14. Bind the target file text to the target review task.
[0025] S2. The file opening review rules corresponding to the target review task are structured and metadata bound to generate multiple traceable rule units, and a rule knowledge base is constructed.
[0026] Furthermore, step S2 specifically includes: S21. Receive the file opening review rules corresponding to the target review task.
[0027] Optionally, the rules for reviewing open access to archives may include at least one of the following: an open access management system, control clauses, and review rules.
[0028] S22. Perform structured analysis on the rules for reviewing the opening of archives, identify the boundaries of clauses, applicable conditions, constraints, and hierarchical relationships in the rules for reviewing the opening of archives, and generate different rule structure units.
[0029] S23. Perform rule expression transformation on each rule structure unit to obtain the transformed rule structure unit.
[0030] Optionally, the rule expression includes at least: a condition expression and a constraint description.
[0031] S24. Bind metadata to the transformed rule structure unit to generate a traceable rule unit.
[0032] Optionally, the metadata includes: rule source, clause number, rule type, and applicable scenario.
[0033] S25. Store traceable rule units in the rule knowledge base.
[0034] Optionally, the rule knowledge base supports traceable rule unit-level control for adding, modifying, and invalidating rules.
[0035] S3. Obtain the target file text and update the target file text through preprocessing.
[0036] Furthermore, step S3 specifically includes: S31. Extract the target file text from the file.
[0037] S32. Perform paragraph division and format normalization on the target file text, and update the target file text.
[0038] S4. Extract factual information from the target file text and map the factual information to rule matching parameters; the factual information is used for rule condition judgment.
[0039] Furthermore, step S4 specifically includes: S41. Load the rule condition parameter configuration corresponding to the audit scenario in the target audit task.
[0040] S42. Extract factual information from the target archive text; the factual information is used to characterize the objective attributes of the archive.
[0041] Optionally, the factual information includes: file type, creation time, type of object involved, whether it contains personal privacy information, and whether it involves state secrets.
[0042] S43. Map factual information into rule matching parameters that the rule engine can directly call.
[0043] Furthermore, the representation of the factual information includes at least one of the following: an enumeration value, a Boolean value, or a numerical form.
[0044] Furthermore, the rule matching parameters are only used for rule condition judgment and do not directly participate in rule review.
[0045] S5. Construct a rule matching object based on the target file text, and use the rule matching parameters to call the rule engine retrieval enhancement generation mechanism to perform rule review on the rule matching object and obtain the review explanation result.
[0046] Furthermore, step S5 specifically includes: S51. Construct a rule matching object based on the target file text.
[0047] S52. Call the rule engine to perform deterministic matching on traceable rule units in the rule knowledge base, and identify all traceable rule units that meet the conditions based on the rule matching parameters to form a rule hit set.
[0048] S53. Perform deduplication and structuring on the rule hit set, and update the rule hit set.
[0049] S54. Construct a rule interpretation context based on archive summary, rule hit set and metadata integration.
[0050] S55. Input the rule interpretation context into the language analysis model, and the language analysis model will review the rule matching objects and generate a review interpretation result based on rule references.
[0051] Furthermore, the audit interpretation results include: rule hit description, rule application analysis, and evidence location information.
[0052] S6. Summarize the target file text, rule matching objects, and review interpretation results, and output a file open review analysis report.
[0053] Example 2, as Figure 2 As shown, this embodiment provides a document review system based on rule engine retrieval enhancement. The document review system based on rule engine retrieval enhancement utilizes the methods described above and includes: Initialization unit 101 is used to receive target audit task identification information and initialize the audit task running environment.
[0054] The rule knowledge base construction unit 102 is used to perform structured processing and metadata binding on the file opening review rules corresponding to the target review task, generate multiple traceable rule units, and construct a rule knowledge base.
[0055] The preprocessing unit 103 is used to acquire the target file text and update the target file text through preprocessing.
[0056] The rule matching parameter extraction unit 104 is used to extract factual information from the target file text and map the factual information to rule matching parameters; the factual information is used for rule condition judgment.
[0057] The rule engine review and interpretation unit 105 is used to construct a rule matching object based on the target file text, and to call the rule engine retrieval enhancement generation mechanism through rule matching parameters to review the rule matching object and obtain the review and interpretation result.
[0058] The document opening review and analysis report generation unit 106 is used to summarize the target document text, rule matching objects and review interpretation results, and output the document opening review and analysis report.
[0059] In practical application, this application does not involve parameter optimization or performance improvement based on existing audit models or retrieval mechanisms. Instead, it starts from the essential needs of rule auditing tasks and reconstructs the technical division of labor between rule retrieval and audit interpretation in the audit system. Specifically, this application does not adopt a rule retrieval method based on vector similarity ranking, nor does it use a small number of "most relevant rules" as the basis for audit judgment. Instead, it uses rule structuring processing and a rule engine mechanism to perform deterministic matching of audit rules, thereby identifying all rule clauses that meet the applicable conditions during the audit process and ensuring the completeness of rule retrieval. In this way, the audit results do not depend on similarity approximation judgment, avoiding the problem of rule omissions when there are many rules or diverse rule expressions. On this basis, this application introduces a large language model (language analysis model) as a rule interpretation and evidence generation component. Based on the rule hitting results output by the rule engine, the large language model performs semantic analysis, clause interpretation, and evidence location on the rule applicability relationship, thereby generating traceable audit interpretation results. By clearly distinguishing the responsibilities of rule matching and semantic interpretation, the audit process simultaneously possesses determinism and interpretability, meeting the requirements of consistency and auditability in high compliance scenarios.
[0060] The technical effects of this application are as follows: This application improves the interpretability and traceability of review results by reconstructing the technical division of labor between rule recall and semantic interpretation, ensuring the completeness of rule applicability while maintaining the accuracy and traceability of review results. First, it receives review tasks and initializes the runtime environment. It performs structured parsing and metadata binding on the open archive review rules, converting the rules into executable rule units containing conditional expressions and constraint descriptions. This constructs a rule knowledge base that supports independent maintenance at the unit level, eliminating the need for model retraining for rule updates and significantly reducing maintenance costs. Next, it acquires and preprocesses the target archive text, extracting factual parameters and mapping them to rule matching parameters. After constructing a rule matching object based on the archive text, it calls the rule engine to perform deterministic matching on all rule units in the rule knowledge base. The satisfaction of rule applicability conditions is used as the recall criterion to identify all matching rules, thus avoiding the risk of missed rule detections due to similarity retrieval. Finally, it constructs a rule matching object based on the target archive text and uses the rule matching parameters to call the rule engine's enhanced generation mechanism to review the rule matching object, obtaining the review interpretation result. This separates rule matching from semantic interpretation, reducing fluctuations caused by probabilistic model inference and ensuring consistency of review results. In this application, rule judgment is triggered by the rule engine, avoiding the omission of key information due to contextual limitations. Finally, a complete audit analysis report is generated, allowing each audit conclusion to be traced back to the specific rule clause and its applicable reasons. In summary, this application avoids the rule omission problem caused by similarity retrieval, improves the reliability, consistency, and traceability of document compliance audits, and reduces the fluctuations caused by probabilistic inference in the model.
[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0062] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for reviewing archives generated based on rule engine retrieval enhancement, characterized in that, The method includes: Receive the target audit task identifier information and initialize the audit task runtime environment; The file opening review rules corresponding to the target review task are structured and bound to metadata to generate multiple traceable rule units, thus constructing a rule knowledge base; Obtain the target file text and update the target file text through preprocessing; Factual information is extracted from the target file text and mapped to rule matching parameters; the factual information is used for rule condition judgment. A rule matching object is constructed based on the target archive text, and the rule engine retrieval enhancement generation mechanism is called through the rule matching parameters to perform rule review on the rule matching object and obtain the review and interpretation results. Summarize the target file text, rule matching objects, and review interpretation results, and output a file open review analysis report.
2. The document review method based on rule engine retrieval enhancement according to claim 1, characterized in that, Receive the target audit task identifier information and initialize the audit task runtime environment, specifically including: Receive the target review task identifier information and determine the set of file objects associated with the target review task and the identifier of the applicable rule set; Load the process environment for the target review task; the process environment is used to display the rule matching objects and the review interpretation results; Configure permissions for viewing, referencing, and recording audit interpretation results; Bind the target file text to the target review task.
3. The document review method based on rule engine retrieval enhancement according to claim 1, characterized in that, The document access review rules corresponding to the target review task are structured and bound to metadata to generate multiple traceable rule units, thus constructing a rule knowledge base, which specifically includes: Receive the file opening review rules corresponding to the target review task; The rules for reviewing the opening of archives are analyzed in a structured manner to identify the boundaries of clauses, applicable conditions, constraints, and hierarchical relationships in the rules, and to generate different rule structure units. Each rule structure unit is transformed into a rule expression to obtain the transformed rule structure unit; Bind metadata to the transformed rule structure unit to generate a traceable rule unit; Traceable rule units are stored in the rule knowledge base.
4. The document review method based on rule engine retrieval enhancement according to claim 1, characterized in that, Obtain the target file text and update the target file text through preprocessing, specifically including: Extract the target archive text from the archive file; The target file text is divided into paragraphs and its format is normalized, and then updated.
5. The document review method based on rule engine retrieval enhancement according to claim 1, characterized in that, Extract factual information from the target archive text and map this factual information to rule matching parameters, specifically including: Load the rule condition parameter configuration corresponding to the audit scenario in the target audit task; Factual information is extracted from the target archival text; this factual information is used to characterize the objective attributes of the archival text. Factual information is mapped to rule matching parameters that the rule engine can directly call.
6. The document review method based on rule engine retrieval enhancement according to claim 5, characterized in that, The representation of the factual information includes at least one of the following: enumeration value, Boolean value, or numerical form.
7. The document review method based on rule engine retrieval enhancement according to claim 5, characterized in that, The rule matching parameters are only used for rule condition judgment and do not directly participate in rule review.
8. The document review method based on rule engine retrieval enhancement according to claim 1, characterized in that, A rule matching object is constructed based on the target archive text. The rule matching object is then reviewed by the rule engine's enhanced generation mechanism, which uses the rule matching parameters to obtain the review and interpretation results, including: Construct rule-matching objects based on the target archive text; The rule engine is invoked to perform deterministic matching on traceable rule units in the rule knowledge base. Based on the rule matching parameters, all traceable rule units that meet the conditions are identified, forming a rule hit set. The rule hit set is deduplicated and structured, and then updated. A rule interpretation context is constructed by integrating archive summaries, rule hit sets, and metadata; The rule interpretation context is input into the language analysis model, which then reviews the rule matching objects and generates a review interpretation result based on rule references.
9. The document review method based on rule engine retrieval enhancement according to claim 8, characterized in that, The audit interpretation results include: rule hit explanation, rule application analysis, and evidence location information.
10. A document review system based on rule engine retrieval enhancement, characterized in that, The document review system based on rule engine retrieval enhancement utilizes the method described in any one of claims 1-9, wherein the document review system based on rule engine retrieval enhancement includes: An initialization unit is used to receive target audit task identification information and initialize the audit task running environment. The rule knowledge base construction unit is used to perform structured processing and metadata binding on the file opening review rules corresponding to the target review task, generate multiple traceable rule units, and construct a rule knowledge base. A preprocessing unit is used to acquire the target file text and update the target file text through preprocessing. A rule matching parameter extraction unit is used to extract factual information from the target archive text and map the factual information to rule matching parameters; the factual information is used for rule condition judgment. The rule engine review and interpretation unit is used to construct a rule matching object based on the target file text, and call the rule engine retrieval enhancement generation mechanism through the rule matching parameters to review the rule matching object and obtain the review and interpretation result. The document opening review and analysis report generation unit is used to summarize the target document text, rule matching objects, and review interpretation results, and output the document opening review and analysis report.