Group regulation domain ontology gate-controlled system double-chain propagation retrieval method, system and terminal

Abstract

The application discloses a system and terminal for double-chain propagation retrieval of regulations based on group regulation domain ontology gating, and relates to the technical field of information retrieval. The technical solution comprises the following steps: processing group multi-level regulation texts to form clause-level atomic units, a unified field system and a group regulation domain ontology model; inducing longitudinal effect chain candidate relationships and horizontal correlation chain candidate relationships between clauses, and performing gating determination and controlled solidification to construct a constrained clause knowledge network; for a target matter, performing longitudinal effect chain propagation and horizontal correlation chain expansion to form a candidate clause combination, and determining an optimal applicable clause combination from the candidate clause combination through a clause combination convergence function; and self-adaptingly updating parameters. The application can be used in a group enterprise multi-level and multi-correlation regulation text environment to accurately and stably determine an applicable clause combination under the simultaneous constraint of an upper regulation and a side special regulation for the same business matter.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and more specifically, to a method, system, and terminal for institutional dual-chain propagation retrieval based on ontology gating in the domain of group regulations. Background Technology

[0002] Group enterprises typically form a multi-tiered system of regulations, from "supervisory unit—group headquarters—subsidiary—department / project." These regulations are stored and continuously revised in the form of management methods, implementation rules, business specifications, and templates. For the same business matter, it is often simultaneously constrained by higher-level regulations and by secondary, specific regulations related to funding, procurement, budgeting, auditing, and construction projects. Therefore, there are not only vertical relationships of effectiveness such as hierarchical inheritance, authorization linkage, coverage substitution, and exception application between regulations, but also horizontal relationships such as citation, supplementation, pre-conditions, parallel constraints, and association with the same topic. This complex interplay of relationships makes group regulation retrieval not a simple text search problem, but a complex technical issue involving the structured organization of regulations, relationship identification, propagation constraints, and result convergence.

[0003] Currently, there are several main approaches to the retrieval and management of institutional texts, but all of them have significant limitations: (1) Traditional techniques based on institutional document databases and keyword retrieval have a retrieval granularity that remains at the document or chapter level, making it difficult to accurately locate the combination of clauses applicable to specific matters, especially unable to handle joint constraint requirements across levels and themes. (2) Techniques based on full-text retrieval, vector retrieval, or retrieval enhancement generation can improve the recall rate of semantically relevant content, but due to the lack of binding of clause-level identifiers and their attributes such as level and time, they often recall relevant but inapplicable fragments, and cannot determine whether a legitimate propagation link can be formed between different fragments, resulting in unstable retrieval results. (3) Approaches that extract information by constructing rule bases, knowledge graphs, or using large models rely heavily on manual pre-set templates, full-scale sorting and annotation in relation construction, which not only has high implementation and maintenance costs, but is also prone to failure when revising institutional texts. More importantly, such solutions typically focus on identifying the presence or absence of relationships and static graph construction, lacking a unified domain ontology model to perform pre-gating checks on the legitimacy, consistency, and propagation boundaries of both vertical validity and horizontal association relationships before they are incorporated into the graph. This leads to erroneous or weakly related relationships easily infiltrating the knowledge network and continuing to spread during subsequent multi-hop retrieval and expansion processes, causing problems such as candidate clause inflation, path drift, and non-convergence of results. Ultimately, it is difficult to output stable, interpretable, and auditable combinations of applicable clauses.

[0004] Therefore, researching and designing a dual-chain propagation retrieval method, system, and terminal based on ontology gating of group regulations that can overcome the above-mentioned defects is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, and terminal for institutional dual-chain propagation retrieval based on ontology gating in the domain of group regulations. This method, based on clause-level atomization, forms a unified field system through a phased field learning path tailored to the group regulations scenario, and constructs an ontology model of the group regulations domain. This ontology model serves as the basis for relation graph gating and propagation constraints. Simultaneously, it transforms the original relation construction method, which relied on manual full-scale analysis and direct edge construction, into a dual-chain construction method combining automatic candidate relation summarization, gating solidification, and minimal manual confirmation, thus suppressing erroneous relations from entering the propagation network at the source. Furthermore, through vertical effect propagation, horizontal association expansion, and clause combination convergence mechanisms, it outputs applicable clause combinations, dual-chain paths, and evidence summaries corresponding to the target matter, thereby improving the stability, interpretability, and engineering feasibility of the institutional retrieval results.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: Firstly, a two-chain propagation retrieval method for regulations based on ontology gating in the group regulations domain is provided, including the following steps: The group's multi-level institutional documents are processed to form clause-level atomic units, a unified field system, and an ontology model for the group's regulations. Based on the aforementioned ontology model of the group regulations domain, the candidate relationships of vertical validity chains and horizontal association chains between clauses are automatically summarized, and the candidate relationships are gating and controlled to solidify, so as to construct a knowledge network of constrained clauses. On the knowledge network of the bound clauses, for the input target matter, vertical effect chain propagation and horizontal association chain expansion are performed. Based on the propagation and expansion results, candidate clause combinations are formed, and the optimal applicable clause combination is determined from the candidate clause combinations by calculating the clause combination convergence function. Based on text processing, network construction, and combination of feedback information during the execution process, the parameters for field learning, relation induction, gating determination, and propagation control are adaptively updated.

[0007] Furthermore, the processing of multi-level group policy texts to form clause-level atomic units, a unified field system, and a group regulation domain ontology model includes: Multi-source policy documents are collected and preprocessed for standardization to obtain standardized policy texts; The standardized policy text is parsed at the clause level to generate clause atomic units and corresponding clause anchor point information; The clause atomic units are processed through a phased field learning path, which includes a local field learning stage for a single policy document and a unified cross-document learning stage for all policies across the group. In the single-system document local field learning stage, semantic parsing and candidate field summarization are performed on the atomic units of clauses of a single system document to obtain the candidate field set of the corresponding system document. During the unified learning phase of all group policies across documents, the candidate field sets of all policy documents are merged across documents, normalized, mapped, and validated for consistency to form the unified field system. Based on the unified field system, the domain ontology model of the group regulations is constructed. The domain ontology model of the group regulations is used to describe the subject category, matter category, scope of application, hierarchical attributes, relationship type and propagation constraint rules of the system clauses.

[0008] Furthermore, based on the aforementioned ontology model of the group regulations domain, the candidate relationships of vertical validity chains and horizontal association chains between clauses are automatically summarized, and the candidate relationships are gating and controlled to solidify, in order to construct a knowledge network of bound clauses, including: Based on the aforementioned clause atomic units, unified field system, and clause anchor information, vertical effect chain candidate relationships are summarized to form a set of vertical effect chain candidate relationships, wherein the vertical effect chain candidate relationships include at least one of the following: superior constraint, subordinate refinement, coverage substitution, authorization connection, and exception application relationship. Based on the aforementioned clause atomic units, unified field system, and clause anchor information, horizontal association chain candidate relationships are summarized to form a set of horizontal association chain candidate relationships. The horizontal association chain candidate relationships include at least one of the following: reference, supplement, precondition, parallel constraint, same topic association, and suspected conflict relationship. The candidate relationships of the vertical power chain and the candidate relationships of the horizontal association chain are encapsulated in a structured manner to form candidate relationship records that include relationship type, evidence vector and relationship confidence. Based on the domain ontology model of the group regulations, the candidate relationship records are gating the judgment. The gating judgment includes verifying at least one of subject consistency, matter domain consistency, hierarchical propagation and time applicability.

[0009] Furthermore, the gating determination of the candidate relationship records is achieved by calculating the candidate relationship gating function value, specifically including: The candidate relation gating function value is the weighted sum of semantic association score, ontology consistency score, hierarchical propagation consistency score, time applicability score and evidence sufficiency score, minus a normalized result after deducting an uncertainty penalty term; Based on the comparison between the candidate relationship gating function value and the first threshold and the second threshold, the candidate relationship is determined as a fixed relationship, a relationship to be confirmed, or is eliminated, wherein the first threshold is greater than the second threshold.

[0010] Furthermore, on the bound clause knowledge network, for the input target matter, vertical validity chain propagation and horizontal association chain expansion are performed, and a candidate clause combination is formed based on the propagation and expansion results, including: On the knowledge network of the bound clauses, based on the organizational level, applicable time and authority boundaries corresponding to the target matter, vertical effect chain propagation is performed to obtain the vertical propagation result; On the knowledge network of the constrained clauses, based on the subject matter, constraints and keywords corresponding to the target matter, a horizontal association chain expansion is performed to obtain the horizontal expansion result; Based on the consistency constraints of the subject domain, the consistency constraints of the applicable objects, and the hierarchical boundary constraints defined by the domain ontology model of the group regulations, the clauses in the vertical propagation results and horizontal expansion results are jointly constrained and screened to form a set of candidate clause combinations.

[0011] Furthermore, the step of determining the optimal applicable clause combination from the candidate clause combinations by calculating the clause combination convergence function includes: For a candidate combination of applicable clauses, the convergence function of the clause combination is the weighted sum of the longitudinal effect propagation score, the horizontal association extension score, and the evidence chain integrity score, minus a conflict and diffusion penalty term. The optimal combination of applicable clauses is the combination that maximizes the convergence function value among all candidate combinations of applicable clauses.

[0012] Furthermore, the method also includes: Output the dual-chain relationship path, clause source description, and evidence summary corresponding to the optimal combination of applicable clauses; The dual-chain relationship path records the sequence of vertical validity relationships and horizontal association relationships from the target matter to the optimal combination of applicable terms.

[0013] Furthermore, the adaptive updating of parameters for field learning, relation induction, gating determination, and propagation control based on feedback information obtained during the text processing, network construction, and combination determination process includes: Based on feedback information from field learning and usage, update the field summarization strategy and synonym merging rules in the phased field learning path; Based on feedback information from candidate relationship induction, gating determination, and manual confirmation, the constraint parameters and candidate relationship identification parameters in the domain ontology model of the group regulations are updated. Based on the feedback information from vertical propagation, horizontal expansion, and clause combination convergence, update the propagation control parameters and the weight parameters in the clause combination convergence function; Key intermediate and final results in the text processing, network construction, and combination process are uniformly recorded to form audit logs.

[0014] Secondly, it provides a dual-chain propagation retrieval system for regulations based on ontology gating in the field of group regulations, including: The text processing module is used to process multi-level institutional documents of the group to form clause-level atomic units, a unified field system, and an ontology model of the group's regulations. The network construction module is used to automatically summarize the vertical validity chain candidate relationship and the horizontal association chain candidate relationship between clauses based on the domain ontology model of the group regulations, and to perform gating judgment and controlled solidification of the candidate relationship in order to construct a knowledge network of constrained clauses. The combination optimization module is used to perform vertical validity chain propagation and horizontal association chain expansion for the input target matter on the constrained clause knowledge network, form candidate clause combinations based on the propagation and expansion results, and determine the optimal applicable clause combination from the candidate clause combinations by calculating the clause combination convergence function. The parameter update module is used to adaptively update the parameters for field learning, relation induction, gating determination, and propagation control based on feedback information determined during the execution process through text processing, network construction, and combination.

[0015] Thirdly, a computer terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the institutional double-chain propagation retrieval method based on group regulation domain ontology gating as described in any one of the first aspects.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The institutional dual-chain propagation retrieval method based on domain ontology gating provided by this invention provides a unified, domain-oriented constraint benchmark for the entire retrieval process through the constructed Domain Ontology Model of Group Regulations (GRDO), ensuring the semantic and logical consistency of all subsequent operations. Furthermore, by utilizing this GRDO to pre-gated and controlled solidify the automatically inferred vertical validity chains and horizontal association chains, a highly reliable constrained clause knowledge network (RKN) is constructed from the source, fundamentally suppressing the introduction and spread of erroneous relational edges. In addition, targeted vertical propagation and horizontal expansion are performed on this clean, structured RKN, and a quantified convergence function is used for optimization selection. This transforms the open and drift-prone graph retrieval problem into an optimal path finding problem on a constrained network, ensuring the stability and reproducibility of the output results. The discretized, static, and highly manual-dependent post-processing workflow is reconstructed into an automated, dynamic, and model-driven feedforward-feedback coupling mechanism, significantly improving the technical feasibility of generating accurate, complete, and interpretable combinations of applicable clauses in complex institutional environments.

[0017] 2. This invention, through a phased learning process from local induction to global unification, enables the generated unified field system and GRDO to align with the actual terminology and structure of a specific group, rather than relying on pre-set fixed templates, greatly enhancing the system's generalization ability and adaptability. Based on this, institutional relationships are clearly distinguished into two categories: vertical effectiveness chains and horizontal association chains, and inductively classified separately. This aligns with the inherent logic of hierarchical effectiveness and thematic association within a group's institutional structure. Furthermore, the use of GRDO rich in domain knowledge for automatic gating of candidate relationships essentially encodes the review experience of human experts into computable rules, achieving high-throughput, standardized quality control of massive candidate relationships. This focuses human intervention on a small number of highly uncertain edges, thereby reducing the human cost of relationship construction and maintenance while ensuring the quality of the relationship network.

[0018] 3. This invention decomposes the abstract concept of relation legitimacy into a comprehensive score across multiple computable dimensions, including semantics, ontology, hierarchy, time, and evidence, through a candidate relation gating function. A threshold-based hierarchical processing mechanism provides a precise benchmark for relation edge quality, crucial for constructing a high-quality RKN. Based on RKN, vertical propagation and horizontal expansion are no longer blind graph traversals but directed searches under GRDO constraints, effectively avoiding invalid paths. The clause combination convergence function comprehensively evaluates candidate combinations selected through propagation from multiple perspectives, including vertical validity satisfaction, horizontal association coverage, and evidence chain integrity, and actively suppresses unreasonable combinations through conflict penalty terms. The clause combination convergence function and the optimization logic together constitute a clear optimization objective, transforming the fuzzy problem of finding the optimal clause combination into a clear mathematical optimization problem, thus ensuring that the output is not only relevant but also the most optimal and stable solution under the current knowledge network and constraints.

[0019] 4. This invention outputs a complete double-chain evidence path and establishes a feedback-based adaptive update closed loop. It uses real feedback from application scenarios as a supervisory signal to adjust the field learning strategy, GRDO constraint parameters, relationship identification weights, and propagation control parameters in reverse. Through this closed-loop learning, it can gradually adapt to the dynamic changes in the group's system and continuously correct its deviations in field understanding, relationship induction, and propagation judgment. This will continuously improve the accuracy and stability of retrieval in long-term operation, forming a virtuous cycle of continuous self-reinforcement. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an embodiment of the present invention; Figure 2 This is a flowchart of step S1 in Embodiment 1 of the present invention; Figure 3 This is a flowchart of step S2 in Embodiment 1 of the present invention; Figure 4 This is a flowchart of step S3 in Embodiment 1 of the present invention; Figure 5 This is a flowchart of step S4 in Embodiment 1 of the present invention; Figure 6 This is a system block diagram in Embodiment 3 of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0022] Example 1: A two-chain propagation retrieval method for regulations based on ontology gating in the group regulations domain, such as... Figure 1 As shown, Figure 1 As shown, this is specifically achieved through steps S1-S4.

[0023] Step S1: Ontology modeling of group regulations and atomic parsing of institutional clauses based on a phased field learning path, such as... Figure 2 As shown. Step S1 is used to convert the multi-level institutional text of the group into a computable representation of clause-level atomic units, a unified field system, and a group regulation domain ontology model. This provides a unified input basis for subsequent ontology-gated dual-chain candidate relationship induction, relationship solidification, vertical effect propagation, horizontal association expansion, and clause combination convergence retrieval. The output of step S1 forms the first product set P1, which includes at least clause atomic units CU, candidate field set FS, unified field system UFS, group regulation domain ontology model (GRDO), and clause anchor information Anchor.

[0024] S1-1: Collection and standardization preprocessing of multi-source institutional documents.

[0025] We collect policy documents and attachments from different levels of government, including supervisory units, group headquarters, subsidiaries, and departments / projects. We then perform standardized preprocessing on the documents, including format unification, coding standardization, title style standardization, numbering rule alignment, header and footer noise removal, watermark noise elimination, table and paragraph continuity restoration, and text structure adjustment, to obtain standardized policy documents that can be used for subsequent clause segmentation and field learning.

[0026] S1-2: Clause-level atomization parsing and anchor point generation.

[0027] The standardized institutional text is parsed at the clause level to identify structural levels such as chapters, sections, articles, clauses, and items, and the content of the institutional text is broken down into clause atomic units (CUs) with stable identifiers. At the same time, clause anchor information is generated for each clause atomic unit. The clause anchor information includes at least structural path anchors, original text position anchors, and reference positioning anchors. It is used to record the hierarchical path, original text position, and possible reference position of the clause in the original institutional document, thereby providing a positioning basis for subsequent candidate relationship induction and evidence path playback.

[0028] S1-3: Construction of a phased field learning path for the ontology of group regulations domain.

[0029] To accommodate the differences in institutional structure, terminology, hierarchical naming, and specific institutional settings among different group companies, this invention does not employ the method of manually enumerating all field templates in advance and filling them in item by item. Instead, it constructs a phased field learning path oriented towards the group's regulations ontology. The phased field learning path includes at least a local field learning stage for a single institutional document and a unified cross-document learning stage for all group-wide institutional regulations.

[0030] In some optional examples, the phased field learning path also includes the following guidance: First, clause-level learning granularity constraints are used to limit field learning to use clause atomic units (CUs) as the basic objects; Secondly, a candidate field classification prior framework is used to limit field learning to focus on the source of the system, scope of application, subject matter, constraint content and relationship clues in the context of the group regulations. Third, a unified learning objective across documents is used to merge, normalize, merge synonyms, and verify consistency of candidate fields in different institutional documents.

[0031] By adopting a phased field learning path, the field learning process can possess both AI-driven adaptive inductive capabilities and constrained direction for ontology modeling in the group's regulations domain.

[0032] S1-4: Learning about local fields in single-system documents.

[0033] For each policy document, an AI-driven semantic parsing and candidate field induction method is used to learn the semantic information in the atomic unit CU of the clause and output the candidate field set FS corresponding to the policy document.

[0034] In some optional examples, the candidate fields include at least the following five categories: (1) The source and level attribute fields of the system should include at least the system level and the issuing entity; (2) The fields for applicable objects and scope of application should include at least the applicable objects and the time range; (3) The semantic and topic expression fields for the matter should include at least the matter topic and keyword items; (4) The constraint content and execution condition fields shall include at least the constraint conditions, amount thresholds and permission elements; (5) Reference and relationship clue fields, including at least the reference information.

[0035] In this stage, candidate fields are not extracted one by one according to fixed rules. Instead, by semantically learning the institutional statements, contextual dependencies, corresponding expressions, limiting conditions and clause structures, local field representation results are formed from within a single institutional document, thereby improving the adaptability to different institutional writing styles and different special institutional expression methods.

[0036] S1-5: Unified learning of all group policies across documents and formation of a unified field system.

[0037] A unified cross-document learning process is performed on the candidate field set FS of each institutional document. Based on the domain knowledge of the group's regulations and sample feedback, the field names, field semantics and field categories are merged, normalized, synonyms are merged and consistency is verified to form a unified field system UFS.

[0038] In some optional examples, during this process, fields with similar semantics but different expressions in different institutional documents are merged into a unified field category. For example, the group headquarters, group headquarters, and headquarters functional departments are merged into a unified subject category; expenses, fund payments, and fund allocations are merged into a unified matter category; and approval authority, signing authority, and decision-making authority are merged into a unified authority element category, so as to improve the consistency of subsequent candidate relationship summarization and propagation calculation.

[0039] This step enables the transformation from summarizing local fields in a single file to representing fields in a unified manner across the entire group's policies, thus avoiding excessive reliance on a single file structure or a single existing template within the group for the field system.

[0040] S1-6: Ontology Model Construction in the Group Regulations Domain.

[0041] Based on the Unified Field System (UFS), a Group Regulations Domain Ontology Model (GRDO) is constructed to uniformly describe the subject categories, matter categories, scope of application, hierarchical attributes, relationship types, propagation boundaries, and conflict constraint rules of institutional clauses. GRDO serves as the unified constraint foundation for subsequent automatic candidate relationship summarization, ontology gating verification, and institutional double-chain propagation calculation.

[0042] In some optional examples, GRDO includes at least a set of subject categories, a set of event categories, a set of scope constraints, a set of hierarchical attributes, a set of relationship types, and a set of propagation boundary and conflict constraint rules.

[0043] The system includes the following categories: Subject Category Set describes the organizational, management, and implementing entities corresponding to the system clauses; Item Category Set describes the business themes, management matters, and binding objects corresponding to the system clauses; Scope of Application Constraint Set describes the applicable objects, applicable time, applicable organizational level, and applicable business domain; Hierarchical Attribute Set describes the system hierarchy and its hierarchical relationships; Relationship Type Set describes the vertical validity relationship types and horizontal association relationship types required for subsequent dual-chain candidate relationship summarization; and Propagation Boundary and Conflict Constraint Rules Set constrains subsequent relationship entry gate control, dual-chain propagation scope, and clause combination convergence conditions.

[0044] Through step S1, this invention achieves local field learning for a single policy document and cross-document unified learning for all policy documents of the group, and forms GRDO accordingly. This avoids the need for manual full-scale field template pre-setting and manual organization of all fields, and enhances the adaptability to different group policy systems. At the same time, it transforms the group policy text from raw natural language expression into a computable clause knowledge representation for group regulations, and provides a unified input basis for the subsequent double-chain candidate relationship induction and ontology gating in step S2, and the double-chain propagation and clause combination convergence in step S3.

[0045] Step S2: Summarize and solidify the dual-chain candidate relationship based on the ontology gating of the group regulations domain, such as... Figure 3 As shown. Step S2 is used to summarize the possible vertical validity chain relationships and horizontal association chain relationships between clauses based on the clause atomic units CU, candidate field set FS, unified field system UFS, and group rule domain ontology model formed in step S1, and to achieve controlled solidification through GRDO gating. The output of step S2 forms the second product set P2, which includes at least the vertical validity chain candidate relationship set CV, the horizontal association chain candidate relationship set CH, the solidified relationship set FR, the relationship to be confirmed set QR, the relationship evidence vector REV, and the constrained clause knowledge network RKN.

[0046] The vertical effectiveness chain is used to represent the effectiveness transfer, coverage substitution, authorization connection, and exception application relationships between different institutional levels; the horizontal association chain is used to represent the reference, supplementation, pre-constraint, parallel constraint, thematic association, and suspected conflict relationships between different institutional themes. By separating and summarizing the two chains and unifying gating, institutional relationships are no longer limited to a static "whether there is an association" mapping method, but instead form two types of computable relational objects oriented towards subsequent propagation calculations.

[0047] S2-1: Induction of candidate relationships in the vertical power chain.

[0048] Based on the clause atomic units (CU), unified field system (UFS), and clause anchor information (Anchor) formed in step S1, an AI-driven automatic candidate relationship induction method is used to identify and generate potential vertical effect relationships between different institutional levels, forming a set of vertical effect chain candidate relationships (CV). These vertical effect chain candidate relationships are primarily used to characterize the effect transmission paths and applicable boundaries between clauses at different levels within the group's institutional system.

[0049] In some optional examples, the candidate relationships of the vertical effect chain include at least the superior binding relationship, the subordinate detailing relationship, the covering and substitution relationship, the authorizing connection relationship, and the exception application relationship. Among them, the superior binding relationship is used to characterize the binding relationship between the superior system clause and the subordinate system clause; the subordinate detailing relationship is used to characterize the detailing and supplementing relationship between the subordinate system clause and the superior system clause; the covering and substitution relationship is used to characterize the covering or substitution relationship between the subsequently issued or higher-level clause and the existing clause; the authorizing connection relationship is used to characterize the relationship between the superior system authorizing the implementation or supplementary provisions of the subordinate system; and the exception application relationship is used to characterize the application boundary relationship between the general clause and the specific exception clause.

[0050] The induction of candidate relationships in the vertical effectiveness chain should take into account at least the following information: institutional level, issuing entity, applicable objects, time range, authority elements, reference information, structural path anchor points, and the mapping results between higher and lower levels of the institutional hierarchy.

[0051] In some optional examples, AI-driven automatic candidate relationship induction methods identify candidate relationships with the potential for vertical propagation by jointly analyzing the semantic representation, hierarchical attribute representation, and temporal validity representation of clauses, rather than relying on manual analysis of hierarchical relationships between regulations and manual establishment of relationship edges. This allows the vertical validity chain to have a candidate determination basis for hierarchical propagation, validity constraints, and temporal validity from the generation stage.

[0052] S2-2: Induction of candidate relationships for horizontal association chains.

[0053] Based on the clause atomic units (CU), unified field system (UFS), and clause anchor information (Anchor) formed in step S1, an AI-driven automatic candidate relationship induction method is used to identify and generate potential horizontal relationships between different institutional themes, forming a set of candidate horizontal relationship chains (CH). These candidate horizontal relationship chains are primarily used to characterize the constraint relationships formed around the same issue between different institutional documents, different institutional themes, or different specific institutional provisions.

[0054] In some optional examples, the candidate horizontal association chains include at least the following: reference relationship, supplementary relationship, precondition relationship, parallel constraint relationship, same-topic association relationship, and suspected conflict relationship. Specifically, a reference relationship indicates that one clause references or refers to other institutional clauses; a supplementary relationship indicates that other institutional clauses form a supplementary constraint on the current clause; a precondition relationship indicates that the application of the current clause is premised on the establishment of a precondition clause; a parallel constraint relationship indicates that multiple institutional clauses form a parallel application constraint on the same matter; a same-topic association relationship indicates a cross-institutional association relationship formed around the same subject matter; and a suspected conflict relationship indicates a relationship with a risk of conflict in the applicable objects, binding actions, threshold conditions, or authority boundaries.

[0055] The induction of candidate relationships for horizontal association chains should take into account at least the following information: subject matter, keywords, constraints, monetary thresholds, permission elements, reference information, structural path, relationship clue fields, and cross-document field mapping results.

[0056] In some optional examples, AI-driven automatic candidate relation induction methods identify potential constraint relationships between different institutional themes by jointly learning the semantics of matters, constraints, and cross-institutional reference expressions, thereby avoiding reliance solely on explicit references or human experience to establish lateral relation edges.

[0057] S2-3: Structured representation of candidate relations.

[0058] To facilitate subsequent GRDO gating and bi-chain propagation calculations, this invention performs structured encapsulation on each candidate relation in the vertical effectiveness chain candidate relation set CV and the horizontal association chain candidate relation set CH to form a candidate relation record CRR.

[0059] In some optional examples, each candidate relation record (CRR) includes at least the source clause identifier (CID_s), target clause identifier (CID_t), relation type identifier (RID), relation direction identifier (DIR), relation confidence level (CONF), propagation attribute (PROP), conflict risk marker (RISK), confirmation priority (PRI), and relation evidence vector (REV). The relation evidence vector (REV) records the sources and strength of evidence supporting the generation of the candidate relation. These sources include at least semantic matching evidence, field mapping evidence, referencing anchor evidence, structural path evidence, hierarchical attribute evidence, and time applicability evidence.

[0060] By using a structured representation of candidate relationships, the candidate relationships of the vertical power chain and the candidate relationships of the horizontal association chain are no longer simple relationship labels, but computable relationship objects with propagation attributes, evidence vectors, and risk labels, thus providing a unified input for subsequent GRDO gating decisions and two-chain propagation.

[0061] S2-4: Candidate Relationship Gating Decision Based on GRDO.

[0062] Based on the Group Regulation Domain Ontology Model (GRDO), ontology gating is performed on candidate relation records (CRR) to determine whether the candidate relations meet the inclusion criteria. In some optional examples, GRDO gating verifies candidate relations from at least the following aspects: subject consistency, matter domain consistency, hierarchical propagation, time applicability, relation type legality, and propagation attribute completeness and evidentiary sufficiency.

[0063] Among them, the subject consistency check is used to determine whether the mapping relationship between the source clause and the target clause in terms of subject category and organizational level meets the GRDO constraints; the matter domain consistency check is used to determine whether the semantic consistency between the source clause and the target clause in terms of matter category, business subject and constraint object meets the GRDO constraints; the hierarchical propagation check is used to determine whether the candidate relationship conforms to the propagation boundary of the vertical effect chain or the horizontal association chain; the time applicability check is used to determine the degree of consistency of the candidate relationship in terms of effective time, expiration time and applicable period; the relationship type legality check is used to determine whether the candidate relationship type belongs to the legal relationship type defined in GRDO; and the propagation attribute integrity and evidence sufficiency check is used to determine whether the candidate relationship has the attributes and evidence support that meet the requirements of the two-chain propagation calculation.

[0064] Candidate relationships that meet the GRDO gating conditions are solidified; candidate relationships that do not meet the GRDO gating conditions are eliminated or downgraded; critical edges and suspected conflict edges that are in a high uncertainty range and may affect subsequent propagation results are output as pending confirmation prompts and included in the pending confirmation relationship set QR. Through this mechanism, the present invention limits the manual confirmation to a small number of high-impact relationship edges, rather than performing a full manual review of all candidate relationships.

[0065] S2-5: Calculation of candidate relation gating function.

[0066] In a preferred implementation, a candidate relation gating function is calculated to determine the graph entry validity of clause atomic units i and j under candidate relation type r. As shown in the formula: ; in, The semantic relevance score is used to characterize the degree of semantic matching between the source clause and the target clause in terms of subject matter, binding actions, applicable objects, and keywords. The ontology consistency score represents the degree of consistency between candidate relations and the subject categories, item categories, and scope constraints in GRDO. The hierarchical propagation consistency score is used to characterize whether candidate relationships satisfy the propagation boundary constraints of the vertical effectiveness chain or the horizontal association chain. The time applicability score is used to characterize the consistency of candidate relationships in terms of their effective, invalid, and applicable periods. The evidence sufficiency score is used to characterize the degree of support from the anchor points of reference, field mappings, structural paths, and relational clues of candidate relations. This represents an uncertainty penalty term, used to characterize the degree of uncertainty in candidate relations regarding relation type identification, field mapping, or evidence integrity. Indicates the corresponding weight parameters; This is the normalization function.

[0067] In some optional examples, a candidate relation-based gating function is used. Set the first threshold Second threshold And satisfy .when When the corresponding candidate relationship is determined to be a fixed relationship, it is included in the fixed relationship set FR; when When the corresponding candidate relationship is determined to be a relationship to be confirmed, it is included in the set of relationships to be confirmed, QR; when When necessary, the corresponding candidate relation will be removed or downgraded. Through the above-mentioned hierarchical gating mechanism, the controlled entry of candidate relations into the graph can be achieved, and the spread of erroneous relations in subsequent double-chain propagation can be suppressed.

[0068] S2-6: Construction of a knowledge network based on dual-chain solidification and constrained clauses.

[0069] After the candidate relationships are gated by GRDO, the vertical validity chain candidate relationships and the horizontal association chain candidate relationships through the gated process will be solidified to form a vertical validity sub-network and a horizontal association sub-network, and then the constrained clause knowledge network RKN will be constructed.

[0070] In some optional examples, the vertical validity subnetwork is used to support the vertical validity propagation in subsequent step S3; the horizontal association subnetwork is used to support the horizontal association expansion in subsequent step S3. Each relation edge in the RKN retains its corresponding relation type, direction attribute, gating value, propagation attribute, and relation evidence vector REV, serving as the input basis for subsequent clause combination convergence calculation and evidence path generation. By solidifying and uniformly incorporating the vertical validity chain and the horizontal association chain into the RKN, this invention achieves the transformation from automatic candidate relation summarization to dual-chain controlled network construction, so that the subsequent propagation process is no longer based on unconstrained general graph relations, but on a dedicated institutional relation network that has been gated and filtered by GRDO.

[0071] Through step S2, this invention transforms the original relationship construction method, which relies on manual full-scale sorting and direct edge construction, into a dual-chain construction method based on GRDO constraints. This method combines automatic summarization of candidate relationships for vertical validity chains, automatic summarization of candidate relationships for horizontal association chains, ontology gating and solidification, and minimal manual confirmation. This allows candidate relationships to undergo constraint verification such as subject consistency, matter domain consistency, hierarchical propagation, time applicability, and evidence sufficiency before entering the knowledge network. This prevents erroneous relationships from entering the subsequent propagation network from the source and provides a controlled and computable relationship foundation for vertical validity propagation, horizontal association expansion, and clause combination convergence retrieval in step S3.

[0072] Step S3: Propagation computation and clause convergence retrieval based on ontology-gated bichain network, such as... Figure 4 As shown. Step S3 is used to perform institutional double-chain propagation and clause convergence retrieval on the bound clause knowledge network RKN formed in step S2, so as to obtain the applicable clause combination and its evidence path corresponding to the target matter. The output of step S3 forms a third product set P3, which includes at least the vertical propagation result VP, the horizontal expansion result HP, the candidate clause combination set CA, and the optimal applicable clause combination. And the double-chain evidence path EP.

[0073] Specifically, the vertical propagation result VP characterizes the boundaries of applicable clauses and the transmission of their effectiveness at different institutional levels for the target matter; the horizontal expansion result HP characterizes the set of lateral constraint clauses formed around the target matter and their related relationships; the candidate clause combination set CA characterizes multiple candidate applicable clause combinations formed on the basis of double-chain propagation; and the double-chain evidence path EP records the formation path, relationship type sequence, clause source, and evidence summary of the optimal applicable clause combination. Through step S3, this invention achieves a stable output transformation from scattered hits of several related clauses to applicable clause combinations.

[0074] S3-1: Vertical effectiveness chain propagation.

[0075] Based on the solidified vertical validity subnetwork in step S2, vertical validity chain propagation is performed to determine the applicable clause boundaries of the target matter at different institutional levels, and to identify propagation results such as higher-level constraints, lower-level refinements, coverage substitutions, and exceptions, forming the vertical propagation result VP.

[0076] In some optional examples, the vertical power chain propagation should take into account at least the following factors: the organizational level, institutional level, issuing entity, applicable objects, applicable time, and authority boundaries of the target matter; and in the process of propagation, the hierarchical propagation boundaries and applicable scope constraints defined by GRDO should be combined to prune or stop the propagation of relationship edges that do not meet the propagation conditions.

[0077] The vertical propagation of the effectiveness chain is not simply an extension along the hierarchical lines of higher and lower-level systems. Instead, it employs differentiated propagation strategies for different types of effectiveness relationships: for higher-level binding relationships, the basic constraints of the higher-level system are retained first; for lower-level detailed relationships, detailed clauses are added without violating the higher-level constraints; for coverage and substitution relationships, existing clauses are replaced with clauses of higher effectiveness or those issued later; for authorization and connection relationships, the connection path between authorization clauses and implementation clauses is preserved; and for exception application relationships, general clauses are covered or modified when the exception triggering conditions are met. In this way, the vertical propagation result (VP) accurately reflects the effectiveness transmission structure within the group's multi-level system of systems.

[0078] S3-2: Horizontal association chain expansion.

[0079] Based on the solidified horizontal association subnetwork in step S2, horizontal association chain expansion is performed to obtain a set of side constraint clauses related to the target matter, and to identify reference clauses, supplementary clauses, pre-constraint clauses, parallel constraint clauses and suspected conflict clauses, forming the horizontal expansion result HP.

[0080] In some optional examples, the horizontal association chain expansion should take into account at least the following factors: the subject of the target matter, keywords, constraints, amount thresholds, permission elements, reference information, and cross-system field mapping results; and in the process of expansion, the GRDO-defined matter category boundaries, business domain boundaries, and relationship type legality constraints should be combined to truncate or reduce the weight of weakly related, cross-domain distorted, or insufficiently evidenced relationship edges.

[0081] Horizontal linkage chain expansion is used to discover lateral, specific institutional constraints beyond those directly related to the target matter. For example, when the target matter falls under the construction project management scenario, in addition to project management regulations, it can be further expanded to include lateral regulations related to fund disbursement, budget control, procurement implementation, and audit supervision, thus forming a more complete set of constraints. Through this method, the horizontal expansion result (HP) can reflect the parallel, preceding, and supplementary constraint relationships formed between different institutional themes around the same matter.

[0082] S3-3: Joint constraints of clause combinations and formation of candidate combinations.

[0083] Based on the vertical propagation result VP and the horizontal expansion result HP, and in accordance with the propagation boundary, matter constraints and conflict constraints defined by the Group Rule Domain Ontology Model GRDO, the candidate clauses are subject to joint constraints and result screening to form a candidate clause combination set CA.

[0084] In some optional examples, joint constraints include at least the following: domain consistency constraints, applicable object consistency constraints, applicable time consistency constraints, hierarchical boundary constraints, authority boundary constraints, and conflict handling rules. Through joint constraints, the propagation results at the single-clause level are transformed into candidate combinations of clauses oriented towards the target matter.

[0085] Each candidate clause combination in the CA set includes at least one set of core clauses retained after vertical propagation, one set of supplementary side clauses after horizontal expansion, and a double-chain relationship path corresponding to the clause combination. The formation process of candidate clause combinations not only focuses on whether the clauses are related, but also on whether they can jointly constitute a set of effective constraints on the target matter, thereby avoiding the mistaken inclusion of clauses that are only similar at the textual level but incompatible at the application level into the same combination.

[0086] S3-4: Calculation of convergence function for clause combination.

[0087] In a preferred implementation, the clause combination convergence function is calculated for candidate applicable clause combination A. As shown in the formula: ; in, The score represents the vertical propagation of effectiveness, which is used to characterize the degree to which the candidate clause combination meets the institutional level constraints, the constraints of the superior clauses, the lower-level detailed conditions, and the exception application paths. The score represents the horizontal association expansion score, which characterizes the extent to which candidate clause combinations cover references, supplements, preconditions, and parallel constraints. The evidence chain completeness score is used to characterize whether the two-chain relationship path from the target matter to the candidate clause combination is complete, closed, and replayable. The conflict and diffusion penalty term is used to characterize the degree of conflict, path drift, invalid expansion, or candidate inflation within the clause combination; , , , This represents the corresponding weight parameter.

[0088] Based on the clause combination convergence function, the candidate clause combination with the best convergence score is selected as the optimal applicable clause combination for the target matter, which can be expressed as: ; in, This represents the optimal combination of applicable clauses in the final output. Through a convergence function, the results of vertical validity chain propagation, horizontal association chain expansion, and evidence chain integrity are unified into the same evaluation framework. Furthermore, a penalty term suppresses conflict diffusion and candidate expansion, thereby improving the stability and reproducibility of the final output.

[0089] S3-5: Output of search results.

[0090] Based on the optimal combination of applicable terms Output the search results corresponding to the target matter. In some optional examples, the search results will include at least the best applicable clause combination, the two-chain relationship path, the clause source description, and the evidence summary.

[0091] Among them, the dual-chain relationship path is used to record the sequence of vertical validity relationships and horizontal association relationships from the target matter to the final combination of clauses; the clause source description is used to explain the system to which each clause belongs, the system level and version source; the evidence summary is used to explain the main sources of evidence supporting the output of the combination of clauses, including semantic matching evidence, field mapping evidence, reference anchor evidence and propagation path evidence.

[0092] Through the above output method, the present invention enables the search results to no longer be limited to providing a few relevant clauses or returning a few similar fragments, but to form stable output results that are oriented towards the target matter, have a combined structure, path basis and evidence explanation, thereby facilitating subsequent manual review, compliance review and audit traceability.

[0093] Through step S3, this invention achieves integrated computation of vertical validity chain propagation, horizontal association chain expansion, and clause combination convergence retrieval on the constrained clause knowledge network (RKN), elevating the institutional retrieval process from "clause fragment-level recall" to "applicable clause combination-level output." Compared with existing schemes based on full-text retrieval, vector retrieval, general knowledge graphs, or large-model question answering, this invention can perform controlled propagation and joint convergence of complex relationship networks formed by multi-level and multi-related institutional intertwining under GRDO constraints. This reduces candidate clause expansion and result drift problems, improves the stability, interpretability, and traceability of institutional retrieval results, and provides a clear output basis for feedback writing and adaptive updating in the subsequent step S4.

[0094] Step S4: Adaptive update of bi-linked retrieval based on ontology-gated feedback, such as... Figure 5 As shown. Step S4 records and writes back the field learning results, candidate relationship summarization results, gating judgment results, clause combination output results, and manual confirmation results from steps S1 to S3. Based on the feedback information, it adaptively updates the phased field learning path, the Group Regulation Domain Ontology Model GRDO, the candidate relationship automatic summarization strategy, and the double-chain propagation control parameters to continuously improve the accuracy, stability, and adaptability of the institutional double-chain propagation retrieval under different group institutional systems. The output of step S4 forms the fourth product set P4, which includes at least the updated field learning strategy UL, the updated ontology constraint parameters UG, the updated candidate relationship identification parameters UR, the updated propagation control parameters UP, and the updated audit trace record Log.

[0095] Specifically, the updated field learning strategy UL is used to optimize the local field learning and cross-document unified learning process in step S1; the updated ontology constraint parameter UG is used to optimize the subject category, item category, scope of application, hierarchical attribute, relationship type, propagation boundary and conflict constraint rules in GRDO; the updated candidate relationship identification parameter UR is used to optimize the vertical effectiveness chain candidate relationship induction and the horizontal association chain candidate relationship induction in step S2; the updated propagation control parameter UP is used to optimize the vertical effectiveness chain propagation, horizontal association chain expansion and clause combination convergence calculation in step S3; and the audit trace record Log is used to save the key intermediate results and final results in the current retrieval execution and feedback correction process.

[0096] S4-1: Phased field learning path update based on field learning feedback.

[0097] The learning results of local fields in a single policy document in step S1, the unified learning results of cross-document systems across the entire group, and the feedback information in the process of field mapping, field classification, and consistent use of fields in steps S2 to S3 are recorded and written back to the phased field learning path for updating field summarization strategies, field category mapping relationships, synonym merging rules, and abnormal field handling rules.

[0098] In some optional examples, the feedback information includes at least the following: first, field categories and field values ​​that are frequently used and consistently effective in the subsequent candidate relation induction process; second, high-frequency synonyms, ambiguous expressions, and anomalous expressions that appear in the cross-document unified learning process; third, records in the gating process where candidate relations are eliminated or downgraded due to inconsistent field classification; and fourth, field combinations that have proven to have high or low explanatory power in the clause combination output process.

[0099] When systemic differences are detected between different group institutional systems in terms of institutional hierarchy naming, issuing entity description, subject matter division, permission expression, or threshold expression, this invention updates the prior classification framework and cross-document unified learning objectives in the field learning path. This enables the subsequent field learning process to adapt to the new group institutional system more quickly, without requiring manual full definition of field templates. In this way, the phased field learning path in step S1 can continuously maintain its adaptive capability to group regulations scenarios.

[0100] S4-2: GRDO based on gated feedback and candidate relation induction parameter update.

[0101] The automatic candidate relation summarization results, relation evidence vector REV, gating judgment results, and manual confirmation results in step S2 are recorded and written back to the GRDO constraint model and the automatic candidate relation summarization module to update the ontology constraint parameter UG and the candidate relation identification parameter UR.

[0102] In some optional examples, the updates include at least the following: First, for vertical validity chain relationship patterns that have been repeatedly confirmed as valid by humans, enhance the identification weight of the corresponding relationship type in the automatic candidate relationship summarization; Second, for horizontal association chain relationship patterns that have been repeatedly confirmed as valid by humans, enhance the weight of related matter topic mapping, constraint mapping, and reference clue identification; Third, for candidate relationship patterns that have been repeatedly removed by GRDO gating, adjust the relationship type legality constraints, hierarchical propagation constraints, or matter domain consistency constraints; Fourth, for frequently occurring suspected conflict relationship patterns, adjust the conflict risk marking strategy and confirmation priority strategy.

[0103] When there is a discrepancy between the manually confirmed result and the automatically summarized initial candidate relation result, this invention does not simply overwrite the original result. Instead, it records the deviation pattern as a ternary feedback sample of relation summarization, gating, and confirmation, which is used to subsequently optimize relation recognition weights, gating thresholds, or evidence vector composition, thereby gradually reducing repetitive manual confirmation work. In this way, the bi-chain candidate relation summarization and GRDO gating process in step S2 can be continuously optimized, and erroneous relation edges can be prevented from repeatedly entering the knowledge network.

[0104] S4-3: Binocular propagation control and convergence parameter update based on retrieval output feedback.

[0105] The vertical propagation result VP, the horizontal expansion result HP, the candidate clause combination set CA, and the optimal applicable clause combination from step S3 are used. The dual-chain evidence path (EP) is recorded and, in conjunction with the search results, manual review results, and audit traceability results, is written back to the dual-chain propagation control module and the clause combination convergence module to update the propagation control parameter (UP).

[0106] In some optional examples, the updates include at least the following: First, adjusting the vertical validity chain propagation boundary parameters to address invalid extension paths, hierarchical out-of-bounds paths, or false triggering of exceptions in vertical propagation; Second, adjusting the horizontal association chain extension depth, extension weight, and extension truncation conditions to address issues such as excessive lateral constraints, the inclusion of weakly related clauses, or extension path drift in horizontal propagation; Third, adjusting the joint constraint conditions of the clause combination to address issues such as candidate expansion, conflict aggregation, or broken evidence chains in the candidate clause combination set CA; Fourth, adjusting the corresponding weight parameters to address the performance differences in the vertical propagation score, horizontal expansion score, evidence chain integrity score, and conflict penalty term in the clause combination convergence function, in order to improve the final optimal applicable clause combination. Stability and interpretability.

[0107] When similar optimal combinations of applicable clauses and evidence paths repeatedly appear in multiple searches for a certain type of target matter, this invention can record such combination patterns as high-confidence propagation templates for priority propagation and convergence control of subsequent similar matters. When clause conflicts or path drifts repeatedly occur for a certain type of target matter, it is recorded as a high-risk propagation pattern, and the penalty weight is increased or stricter gating conditions are triggered in subsequent propagation processes. In this way, the double-chain propagation and clause combination convergence retrieval in step S3 can be continuously optimized.

[0108] S4-4: Retrieve trace records.

[0109] In step S4, key intermediate and final results in the field learning, candidate relationship induction, GRDO gating, double-chain propagation, clause combination convergence and result output processes are uniformly recorded to form an audit record Log.

[0110] In some optional examples, the Log includes at least the original input items, the invocation status of atomic units of clauses, the candidate field summary results, the unified field system mapping results, the candidate relation set, the gating value, the set of fixed relations, the set of relations to be confirmed, the vertical propagation result VP, the horizontal expansion result HP, the candidate clause combination set CA, and the optimal applicable clause combination. And the double-chain evidence path EP.

[0111] Through step S4, this invention establishes a closed-loop optimization system with ontology gating of group regulations as its core, a phased field learning path as a pre-adaptation mechanism, and double-chain propagation convergence as its target output mechanism. This ensures that field learning, relationship summarization, ontology gating, double-chain propagation, and clause combination convergence are no longer one-time static processes, but can be continuously and adaptively updated based on different group system regulations, different types of matters, and different search result feedback. Therefore, on the one hand, the workload of manually compiling all fields and relationships can be gradually reduced; on the other hand, the accuracy, stability, interpretability, and reproducibility of the double-chain propagation retrieval of regulations under different group system scenarios can be improved.

[0112] Example 2: The system of a group-type enterprise is used as a typical application scenario for illustration. This group's system hierarchy includes systems for the supervising unit, the group headquarters, subsidiaries, and departments / projects; the system topics cover procurement and bidding, construction management, contract management, expense reimbursement, fund payment, budget control, and authorization approval. For the same business matter, it is often simultaneously constrained by both higher-level systems and lateral specific systems. Therefore, there exists a vertical chain of effectiveness between system clauses characterized by hierarchical transmission, coverage substitution, authorization connection, and exception application, as well as a horizontal chain of association characterized by citation, supplementation, pre-conditional constraints, parallel constraints, related topics, and suspected conflicts. The implementation process of this invention will be described in detail below with reference to steps S1 to S4.

[0113] In this embodiment, the system implementing the technology described in this invention is deployed in a group-level policy retrieval and compliance assistance platform. The platform collects policy documents from the group's policy database, subsidiary policy databases, and project policy databases, including documents such as procurement management regulations, construction management regulations, contract management regulations, expense reimbursement regulations, fund payment authorization forms, and budget control details. The system receives a description of the target matter input by the user, such as an expense reimbursement matter, a project investment change matter, or an urgent procurement payment matter, and outputs, based on the method of this invention, the applicable clause combination corresponding to the target matter, a dual-chain relationship path, and an evidence summary.

[0114] Step S1: Ontology modeling of group regulations and atomic parsing of institutional clauses based on phased field learning paths.

[0115] First, policy documents are collected from different levels, including supervisory units, group headquarters, subsidiaries, and departments / projects. These documents undergo standardized preprocessing, including format standardization, title style standardization, numbering rule alignment, header / footer and watermark noise removal, and table and paragraph continuity restoration, to create standardized policy documents suitable for subsequent parsing. Then, clause-level atomic parsing is performed on the standardized policy documents, identifying structural levels such as chapters, sections, articles, clauses, and items. The policy content is broken down into clause atomic units (CUs), and structural path anchors, original text location anchors, and citation location anchors are generated for each clause atomic unit.

[0116] In this embodiment, a phased field learning path is constructed for the domain ontology of group regulations. First, local field learning is performed on a single regulation document, and then cross-document unified learning is performed on all regulation documents of the group. For local field learning of a single regulation document, the system adopts an artificial intelligence-driven clause semantic parsing and candidate field induction method to extract the candidate field set FS from the clause atomic unit CU.

[0117] Candidate fields include five categories: First, the source and hierarchical attribute fields of the system, such as the system level and issuing entity; second, the applicable objects and scope fields, such as the applicable objects, applicable time, and applicable organizational level; third, the semantics and theme expression fields of the matter, such as the matter theme and keywords; fourth, the constraint content and execution conditions fields, such as constraint conditions, monetary thresholds, and permission elements; and fifth, the reference and relationship clue fields, such as reference information, precondition expression, and exception trigger clues.

[0118] For example, for the clause "a single travel expense exceeding 10,000 yuan requires approval from the department head, and an expense exceeding 20,000 yuan requires approval from the company head", the system can summarize the subject as "travel expense reimbursement", the constraint as "single expense trigger threshold", the amount threshold as "10,000 yuan, 20,000 yuan", and the permission elements as "department head, company head".

[0119] After completing the local field learning of a single policy document, a unified cross-document learning process is performed on the candidate field set (FS) of all policy documents across the group. This involves merging, standardizing mapping, synonym merging, and consistency verification of field names, semantics, and categories to form a unified field system (UFS). For example, "Group Headquarters," "Group Headquarters," and "Headquarters Functional Departments" are merged into a unified subject category; "Expenses," "Reimbursement Payments," and "Fund Allocations" are merged into a unified item category; and "Approval Authority," "Signature Authority," and "Decision-Making Authority" are merged into a unified authority element category. Based on the unified field system (UFS), the system constructs a Group Regulations Domain Ontology Model (GRDO) to uniformly describe the subject category, item category, scope of application, hierarchical attributes, relationship type, propagation boundary, and conflict constraint rules of policy clauses. This forms the first product set (P1) output from step S1, which is used for subsequent dual-chain candidate relationship summarization and ontology gating.

[0120] Step S2: Summarize and solidify the dual-chain candidate relationship based on the ontology gating of the group regulations domain.

[0121] Subsequently, based on the clause atomic unit CU, unified field system UFS, clause anchor information Anchor and GRDO obtained in step S1, candidate summaries are made for the possible vertical effect chain relationship and horizontal association chain relationship between institutional clauses.

[0122] For candidate relationships in the vertical effect chain, the system focuses on identifying the superior constraint relationship, subordinate refinement relationship, coverage and substitution relationship, authorization connection relationship, and exception application relationship between different institutional levels. For example, between the "Group Expense Reimbursement Method" and the "Subsidiary Expense Reimbursement Implementation Rules", a vertical effect chain relationship of "superior constraint - subordinate refinement" can be identified; between the "Original Payment Authorization Table" and the "Newly Revised Authorization Table", a "coverage and substitution" relationship can be identified.

[0123] For candidate horizontal association chains, the system focuses on identifying reference relationships, supplementary relationships, precondition relationships, parallel constraint relationships, relationships with the same theme, and suspected conflict relationships. For example, between the "Construction Project Management Measures" and the "Investment Management Measures," parallel constraint relationships related to budget overruns can be identified; between the "Emergency Procurement Management Measures" and the "Contract Management Measures," horizontal association relationships related to references and exception triggers can be identified.

[0124] In this embodiment, the system structurally encapsulates each candidate relationship to form a Candidate Relationship Record (CRR). The Candidate Relationship Record includes at least: source clause identifier, target clause identifier, relationship type, relationship direction, relationship confidence level, propagation attribute, conflict risk marker, confirmation priority, and Relationship Evidence Vector (REV). The Relationship Evidence Vector (REV) records at least the following evidence sources: semantic matching evidence, field mapping evidence, referencing anchor evidence, structural path evidence, hierarchical attribute evidence, and time applicability evidence.

[0125] For example, when the system identifies a clause in the "Implementation Rules for Subsidiary Expense Reimbursement" and a clause in the "Group Expense Reimbursement Method" as having a subordinate detailed relationship, the system will simultaneously save: the hierarchical differences between the two clauses, the consistency of the subject matter, the matching of the amount threshold, and the citation expressions that appear in the original text, in order to form the corresponding relational evidence vector REV.

[0126] Subsequently, ontology gating is performed on candidate relationships based on GRDO. Specifically, the system performs subject consistency checks, event domain consistency checks, hierarchical propagation checks, time applicability checks, relationship type legality checks, and propagation attribute integrity and evidence sufficiency checks on candidate relationships. Based on the results of the candidate relationship gating function calculation, the relationships are either solidified, pending confirmation, or eliminated.

[0127] In a preferred implementation, regarding the validity of clause atomic units i and j in the graph under candidate relation type r, the system calculates the candidate relation gating function value according to the formula in the aforementioned technical solution. If the gating value is higher than a preset high threshold, the candidate relation is included in the fixed relation set FR; if the gating value is in the pending confirmation range, it is included in the pending confirmation relation set QR for manual confirmation; if the gating value is lower than the low threshold, it is eliminated or downgraded. In this way, the system transforms the original relation construction method that relies on manual full sorting and direct edge construction into a bi-chain construction method of "automatic candidate relation summarization + GRDO gating + minimal manual confirmation".

[0128] After completing the candidate relation gating, the system solidifies the candidate relations of the vertical validity chain and the candidate relations of the horizontal association chain, forming the vertical validity subnetwork and the horizontal association subnetwork, and further constructs the constrained clause knowledge network (RKN). Each relation edge in the RKN retains its relation type, direction attribute, gating value, propagation attribute, and relation evidence vector (REV) for use in the subsequent step S3 to perform bi-chain propagation and clause combination convergence retrieval.

[0129] Step S3: Propagation computation and clause convergence retrieval based on ontology-gated bichain network.

[0130] In this embodiment, three typical issues are selected to illustrate step S3.

[0131] Example Item 1: A subsidiary's marketing department incurred a travel expense of RMB 18,000 in March 2025. It is necessary to determine the applicable combination of expense reimbursement terms.

[0132] Example Item 2: If the total investment of a project is expected to exceed the original approved budget during construction, it is necessary to determine the combination of applicable clauses related to the budget overrun.

[0133] Example Item 3: A project requires partial payment in advance due to urgent procurement. It is necessary to determine the applicable combination of terms under the constraints of contract signing and procurement process.

[0134] For Example Item 1, the system performs vertical propagation of the effect chain based on RKN, identifying the superior constraints and subordinate refinement relationships between the group's expense reimbursement procedures and the subsidiary's expense reimbursement rules, obtaining the vertical propagation result VP. Simultaneously, the system expands the authorization and approval clauses and fund payment clauses related to the reimbursement item based on the horizontal association chain, obtaining the horizontal expansion result HP. Subsequently, based on the consistency of the item domain, applicable objects, applicable time, hierarchical boundaries, and authority boundaries defined by GRDO, the system jointly constrains the candidate clauses, forming several candidate clause combinations CA.

[0135] Based on this, the system calculates a convergence score for each candidate clause combination according to the clause combination convergence function in the technical solution, and selects the clause combination with the best convergence score as the optimal applicable clause combination. The final search results include: the optimal combination of applicable clauses for "reimbursement of travel expenses of RMB 18,000", the relationship path formed by the vertical effect chain and the horizontal association chain, the system and version source of each clause, and a summary of evidence.

[0136] For Example Item 2, the system first performs vertical effect chain propagation based on the knowledge network of constrained clauses (RKN) to identify the superior constraints, detailed supplements, and coverage substitution relationships among the "Project Management Measures," "Investment Management Measures," and their superior systems regarding budget overruns, forming the vertical propagation result VP. Subsequently, the system performs horizontal association chain expansion to retrieve the fund management, approval authorization, and audit supervision clauses related to budget overruns, forming the horizontal expansion result HP.

[0137] The system then applies joint constraints to the candidate clauses based on the propagation boundaries, event constraints, and conflict constraints defined by the GRDO, forming a candidate clause combination (CA) corresponding to the "project investment over budget" event. Through convergence calculation of the candidate clause combinations, the system selects the optimal applicable clause combination. It outputs a two-chain relationship path and a summary of evidence. This output reflects the combined constraint structure formed by budget overruns across different institutional levels and specific institutional frameworks, rather than simply returning a few scattered clauses.

[0138] For Example Item 3, the system performs vertical effect chain propagation to determine the hierarchical boundaries of emergency procurement items in the group procurement system, project procurement implementation rules, and contract management system; at the same time, it performs horizontal association chain expansion to identify procurement exception clauses, contract signing clauses, fund payment clauses, and authorization approval clauses related to the "payment first, contract later" item.

[0139] During the joint constraint phase, the system filters candidate clauses based on the subject matter, applicable objects, monetary thresholds, authority boundaries, and exception triggering conditions defined by the GRDO, forming candidate clause combinations (CAs). Through the calculation of the clause combination convergence function, the system outputs the optimal applicable clause combination that best matches the subject matter. The system simultaneously outputs the bi-chain relationship path that forms the combination and a summary of evidence. This result can support subsequent manual review and compliance audit.

[0140] In this embodiment, the key to step S3 is not returning a single institutional fragment, but rather forming a combination of candidate clauses oriented towards the target matter through joint operations of vertical validity chain propagation and horizontal association chain expansion. Furthermore, a clause combination convergence function is used to suppress candidate expansion, path drift, and conflict diffusion. In this way, the system can transform scattered clause hits into stable outputs of clause combinations, significantly improving the stability, interpretability, and traceability of the institutional retrieval results.

[0141] Step S4: Adaptive update of bi-linked retrieval based on ontology-gated feedback.

[0142] Finally, step S4 is executed to uniformly record and write back the field learning results, candidate relationship summarization results, gating determination results, optimal applicable clause combination output results, and manual confirmation results from steps S1 to S3. Based on the feedback information, the phased field learning path, GRDO, candidate relationship automatic summarization strategy, and double-chain propagation control parameters are adaptively updated.

[0143] For example, for field classification patterns that have been repeatedly verified as valid in multiple matters, the system can write them back into the phased field learning path to improve the efficiency of subsequent field learning; for vertical power chain relationship patterns or horizontal association chain relationship patterns that are frequently confirmed as valid by humans, the system can write them back and enhance the corresponding relationship summarization weight; for relationship types that often lead to invalid expansion or path drift during propagation, the system can suppress their subsequent diffusion by adjusting the gating parameters and propagation control parameters.

[0144] In this embodiment, the system uniformly tracks the entire execution process, forming an audit log. The log includes at least: original event input, clause atomic unit invocation status, candidate field summary results, unified field system mapping results, candidate relationship set, gate value, fixed relationship set, unconfirmed relationship set, vertical propagation result VP, horizontal expansion result HP, candidate clause combination set CA, and optimal applicable clause combination. And the double-chain evidence path EP.

[0145] By recording data, the system can not only support manual review and audit traceability of current matters, but also serve as feedback for subsequent updates of field learning paths, GRDO constraint parameters, candidate relationship identification parameters, and propagation control parameters, thus forming a closed-loop optimization mechanism.

[0146] As can be seen from this embodiment, the present invention, through a phased field learning path, ontology modeling of group regulations, summarization and gating of dual-chain candidate relationships, dual-chain propagation and clause combination convergence, feedback writing and adaptive updating, can stably output applicable clause combinations and dual-chain evidence paths corresponding to target matters in a multi-level, multi-related institutional text environment of the group. This reduces the manual workload in the process of sorting institutional fields and building relationship edges, and improves the stability, interpretability, traceability and engineering feasibility of institutional retrieval results.

[0147] When implementing this invention, it is essential to ensure the completeness of the group's multi-level policy documents, covering at least the levels of the supervising unit, group headquarters, subsidiaries, and departments / projects. Efforts should be made to collect the policy text, attachments, authorization forms, approval forms, and revision information. Missing policy documents, incomplete versions, or unclear effective / expiration dates will affect the accuracy of clause atomic parsing, field learning, and the subsequent summarization of vertical and horizontal effect chains.

[0148] When implementing this invention, the phased field learning path and the Group's Regulations Domain Ontology Model (GRDO) should be continuously corrected and maintained in conjunction with the Group's actual institutional system. Particular attention should be paid to core fields and constraints such as subject categories, matter categories, scope of application, authority elements, monetary thresholds, and relationship types. If significant changes occur in the Group's institutional structure, terminology, or management approach, and the unified field system and GRDO are not updated in a timely manner, it may lead to biases in candidate relationship summarization, distorted gating results, or inaccurate propagation boundaries.

[0149] When implementing this invention, for critical edges, suspected conflict edges, and high-risk clause combinations whose gating values ​​are in the pending confirmation range, a manual confirmation mechanism should be retained, and the field learning results, candidate relation sets, gating judgment results, clause combination output results, and double-chain evidence paths should be uniformly logged. This is beneficial for subsequent audit review and also for continuously optimizing the field learning, relation summarization, and propagation convergence effects through feedback write-back, avoiding the repeated spread of erroneous relations or abnormal paths in the system.

[0150] Example 3: A system for retrieving institutional double-chain propagation based on domain ontology gating of group regulations. This system is used to implement the institutional double-chain propagation retrieval method based on domain ontology gating of group regulations described in Example 1, such as... Figure 6 As shown, it includes a text processing module, a network construction module, a combination optimization module, and a parameter update module.

[0151] The system comprises the following modules: a text processing module for processing multi-level institutional documents within the group to form clause-level atomic units, a unified field system, and a domain ontology model for the group's regulations; a network construction module for automatically summarizing candidate vertical validity chains and horizontal association chains based on the domain ontology model for the group's regulations, and for gating and solidifying these candidate relationships to construct a knowledge network of constrained clauses; a combination optimization module for performing vertical validity chain propagation and horizontal association chain expansion on the constrained clause knowledge network for the input target matter, forming candidate clause combinations based on the propagation and expansion results, and determining the optimal applicable clause combination from the candidate clause combinations through a clause combination convergence function; and a parameter update module for adaptively updating the parameters for field learning, relationship summarization, gating judgment, and propagation control based on feedback information from the text processing, network construction, and combination determination processes.

[0152] Example 4: A computer terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the institutional double-chain propagation retrieval method based on the domain ontology gating of group regulations as described in Example 1.

[0153] Working Principle: This invention, based on clause-level atomization, forms a unified field system through a phased field learning path tailored to group regulations scenarios, and constructs an ontology model for the group regulations domain. This ontology model serves as the basis for relation entry gating and propagation constraints. Simultaneously, it transforms the original relation construction method, which relied on manual full-scale analysis and direct edge building, into a dual-chain construction method combining automatic candidate relation summarization, gating solidification, and minimal manual confirmation, thus suppressing erroneous relations from entering the propagation network at the source. Furthermore, through vertical effect propagation, horizontal association expansion, and clause combination convergence mechanisms, it outputs applicable clause combinations, dual-chain paths, and evidence summaries corresponding to the target matter, thereby improving the stability, interpretability, and engineering feasibility of the regulatory retrieval results.

[0154] 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.

[0155] 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.

[0156] 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.

[0157] 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.

[0158] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dual-chain propagation retrieval method for regulations based on ontology gating in the domain of group regulations, characterized in that, Includes the following steps: The group's multi-level institutional documents are processed to form clause-level atomic units, a unified field system, and an ontology model for the group's regulations. Based on the aforementioned ontology model of the group regulations domain, the candidate relationships of vertical validity chains and horizontal association chains between clauses are automatically summarized, and the candidate relationships are gating and controlled to solidify, so as to construct a knowledge network of constrained clauses. On the knowledge network of the bound clauses, for the input target matter, vertical effect chain propagation and horizontal association chain expansion are performed. Based on the propagation and expansion results, candidate clause combinations are formed, and the optimal applicable clause combination is determined from the candidate clause combinations by calculating the clause combination convergence function. Based on text processing, network construction, and combination of feedback information during the execution process, the parameters for field learning, relation induction, gating determination, and propagation control are adaptively updated.

2. The institutional dual-chain propagation retrieval method based on ontology gating in the group regulations domain as described in claim 1, characterized in that, The processing of multi-level institutional texts within the group to form clause-level atomic units, a unified field system, and a group regulations domain ontology model includes: Multi-source policy documents are collected and preprocessed for standardization to obtain standardized policy texts; The standardized policy text is parsed at the clause level to generate clause atomic units and corresponding clause anchor point information; The clause atomic units are processed through a phased field learning path, which includes a local field learning stage for a single policy document and a unified cross-document learning stage for all policies across the group. In the single-system document local field learning stage, semantic parsing and candidate field summarization are performed on the atomic units of clauses of a single system document to obtain the candidate field set of the corresponding system document. During the unified learning phase of all group policies across documents, the candidate field sets of all policy documents are merged across documents, normalized, mapped, and validated for consistency to form the unified field system. Based on the unified field system, the domain ontology model of the group regulations is constructed. The domain ontology model of the group regulations is used to describe the subject category, matter category, scope of application, hierarchical attributes, relationship type and propagation constraint rules of the system clauses.

3. The institutional dual-chain propagation retrieval method based on ontology gating in the group regulations domain as described in claim 2, characterized in that, Based on the aforementioned ontology model of the group regulations domain, the candidate relationships of vertical validity chains and horizontal association chains between clauses are automatically summarized, and the candidate relationships are gating and controlled to solidify, in order to construct a knowledge network of bound clauses, including: Based on the aforementioned clause atomic units, unified field system, and clause anchor information, vertical effect chain candidate relationships are summarized to form a set of vertical effect chain candidate relationships, wherein the vertical effect chain candidate relationships include at least one of the following: superior constraint, subordinate refinement, coverage substitution, authorization connection, and exception application relationship. Based on the aforementioned clause atomic units, unified field system, and clause anchor information, horizontal association chain candidate relationships are summarized to form a set of horizontal association chain candidate relationships. The horizontal association chain candidate relationships include at least one of the following: reference, supplement, precondition, parallel constraint, same topic association, and suspected conflict relationship. The candidate relationships of the vertical power chain and the candidate relationships of the horizontal association chain are encapsulated in a structured manner to form candidate relationship records that include relationship type, evidence vector and relationship confidence. Based on the domain ontology model of the group regulations, the candidate relationship records are gating the judgment. The gating judgment includes verifying at least one of subject consistency, matter domain consistency, hierarchical propagation and time applicability.

4. The institutional dual-chain propagation retrieval method based on ontology gating in the group regulations domain according to claim 3, characterized in that, The gating determination of the candidate relationship records is achieved by calculating the candidate relationship gating function value, specifically including: The candidate relation gating function value is the weighted sum of semantic association score, ontology consistency score, hierarchical propagation consistency score, time applicability score and evidence sufficiency score, minus a normalized result after deducting an uncertainty penalty term; Based on the comparison between the candidate relationship gating function value and the first threshold and the second threshold, the candidate relationship is determined as a fixed relationship, a relationship to be confirmed, or is eliminated, wherein the first threshold is greater than the second threshold.

5. The institutional dual-chain propagation retrieval method based on group regulations domain ontology gating as described in claim 3, characterized in that, On the knowledge network of the bound clauses, for the input target matter, vertical validity chain propagation and horizontal association chain expansion are performed, and a combination of candidate clauses is formed based on the propagation and expansion results, including: On the knowledge network of the bound clauses, based on the organizational level, applicable time and authority boundaries corresponding to the target matter, vertical effect chain propagation is performed to obtain the vertical propagation result; On the knowledge network of the constrained clauses, based on the subject matter, constraints and keywords corresponding to the target matter, a horizontal association chain expansion is performed to obtain the horizontal expansion result; Based on the consistency constraints of the subject domain, the consistency constraints of the applicable objects, and the hierarchical boundary constraints defined by the domain ontology model of the group regulations, the clauses in the vertical propagation results and horizontal expansion results are jointly constrained and screened to form a set of candidate clause combinations.

6. The institutional double-chain propagation retrieval method based on group regulations domain ontology gating as described in claim 5, characterized in that, The step of determining the optimal applicable clause combination from the candidate clause combinations by calculating the clause combination convergence function includes: For a candidate combination of applicable clauses, the convergence function of the clause combination is the weighted sum of the longitudinal effect propagation score, the horizontal association extension score, and the evidence chain integrity score, minus a conflict and diffusion penalty term. The optimal combination of applicable clauses is the combination that maximizes the convergence function value among all candidate combinations of applicable clauses.

7. The institutional dual-chain propagation retrieval method based on ontology gating in the group regulations domain according to claim 1, characterized in that, The method also includes: Output the dual-chain relationship path, clause source description, and evidence summary corresponding to the optimal combination of applicable clauses; The dual-chain relationship path records the sequence of vertical validity relationships and horizontal association relationships from the target matter to the optimal combination of applicable terms.

8. The institutional dual-chain propagation retrieval method based on ontology gating in the group regulations domain according to claim 1, characterized in that, The method of adaptively updating parameters for field learning, relation induction, gating determination, and propagation control based on feedback information obtained during the text processing, network construction, and combination process includes: Based on feedback information from field learning and usage, update the field summarization strategy and synonym merging rules in the phased field learning path; Based on feedback information from candidate relationship induction, gating determination, and manual confirmation, the constraint parameters and candidate relationship identification parameters in the domain ontology model of the group regulations are updated. Based on the feedback information from vertical propagation, horizontal expansion, and clause combination convergence, update the propagation control parameters and the weight parameters in the clause combination convergence function; Key intermediate and final results in the text processing, network construction, and combination process are uniformly recorded to form audit logs.

9. A dual-chain propagation retrieval system for regulations based on ontology gating in the domain of group regulations, characterized in that: include: The text processing module is used to process multi-level institutional documents of the group to form clause-level atomic units, a unified field system, and an ontology model of the group's regulations. The network construction module is used to automatically summarize the vertical validity chain candidate relationship and the horizontal association chain candidate relationship between clauses based on the domain ontology model of the group regulations, and to perform gating judgment and controlled solidification of the candidate relationship in order to construct a knowledge network of constrained clauses. The combination optimization module is used to perform vertical validity chain propagation and horizontal association chain expansion for the input target matter on the constrained clause knowledge network, form candidate clause combinations based on the propagation and expansion results, and determine the optimal applicable clause combination from the candidate clause combinations by calculating the clause combination convergence function. The parameter update module is used to adaptively update the parameters for field learning, relation induction, gating determination, and propagation control based on feedback information determined during the execution process through text processing, network construction, and combination.

10. A computer terminal, 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 computer program, it implements the institutional double-chain propagation retrieval method based on the domain ontology gating of group regulations as described in any one of claims 1-8.

Images