A knowledge-enhanced bidding regulation clause intelligent auditing method and system

Abstract

The application relates to the technical field of large language models, and particularly discloses a method and system for intelligently auditing bidding regulations and clauses based on knowledge enhancement. The application constructs a regulation knowledge base by fusing a regulation knowledge graph and a vector database, performs natural language understanding analysis on received bidding auditing items, extracts semantic branches, combines the semantic branches to construct an auditing item semantic tree, and forms a semantic condition set. Based on the semantic condition set and the regulation knowledge base, double-channel regulation retrieval is performed, candidate regulation clauses are screened, and a fusion and sorting candidate set is output after weighted fusion processing. For the fusion and sorting candidate set, conflict detection and priority determination are performed using relationship information in the regulation knowledge graph to determine final applicable regulation clauses. A compliance reasoning chain is generated according to the final applicable regulation clauses, and a structured auditing report is output based on the compliance reasoning chain. The application solves the problem of semantic omission in pure keyword matching and makes the auditing result traceable.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, and more specifically, to a method and system for intelligent review of bidding regulations and clauses based on knowledge enhancement. Background Technology

[0002] With the rapid development of electronic bidding and digital procurement management, government procurement, engineering construction bidding, and corporate procurement activities involve a large number of laws, regulations, departmental rules, and industry standards. The qualification requirements, technical parameters, commercial terms, and evaluation rules in bidding documents all need to be compared and verified against current regulations to ensure the legality and compliance of bidding activities. However, at present, the review of bidding regulations still mainly relies on professionals manually reviewing regulations and verifying them item by item. This is not only inefficient but also prone to inconsistencies in review results due to differences in the experience of reviewers, making it difficult to meet the needs of large-scale bidding operations for automated and intelligent review.

[0003] To improve review efficiency, existing technologies include regulatory review systems based on keyword rule matching. However, due to the prevalence of synonyms, diverse expressions, and complex constraints in the bidding and tendering field, relying solely on keyword matching is insufficient to accurately understand the true semantics of review items, leading to missed and false positives, and lacking effective explanation mechanisms and traceability capabilities. Furthermore, some systems employ large language models combined with Retrieval Augmentation (RAG) for regulatory question answering and compliance review. While this approach improves semantic understanding, it still suffers from problems such as the illusion of a large model, insufficient timeliness of regulations, and weak handling of regulatory applicability relationships. Particularly in bidding and tendering legal scenarios, the model may generate clauses not found in the regulations or provide incorrect interpretations, increasing review risks. Moreover, most existing RAG solutions are based on text vector similarity retrieval, making it difficult to handle citation relationships between regulations, exceptions, and priority rules between higher and lower-level laws. They also lack a complete display of reasoning basis and a review traceability mechanism, making it difficult for users to verify the accuracy and reliability of review conclusions.

[0004] Therefore, it is necessary to provide a knowledge-enhanced intelligent review method and system for bidding regulations and clauses to solve the above-mentioned technical problems. In order to solve the above problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, this invention provides a knowledge-enhanced intelligent review method and system for bidding and tendering regulations, which addresses the problems of model illusion, insufficient timeliness of regulations, weak ability to handle applicable relationships, difficulty in handling regulatory citations and priority rules in retrieval mechanisms, and lack of complete reasoning basis and traceability mechanisms in current legal scenarios such as bidding and tendering.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A knowledge-enhanced intelligent review method for bidding regulations clauses, the specific steps of which are as follows: By acquiring and structurally parsing legal texts in the bidding and tendering field, a legal knowledge graph and a vector database are obtained, and the legal knowledge graph and vector database are integrated to construct a legal knowledge base. By receiving bidding and tender review items, natural language understanding analysis is performed to extract semantic branches and construct a semantic tree for the review items, forming a set of semantic conditions; Based on semantic condition sets and legal knowledge bases, dual-channel legal retrieval is performed to filter candidate legal clauses, and after weighted fusion processing, a fused and ranked candidate set is output. For the candidate set of fusion ranking, conflict detection and priority determination are performed using relational information in the regulatory knowledge graph to determine the finally applicable regulatory clauses. Generate a compliance reasoning chain based on the applicable regulatory provisions, and output a structured audit report based on the compliance reasoning chain.

[0007] As a further aspect of the present invention, the semantic branch combination includes the audit object, audit attribute, and audit threshold.

[0008] As a further aspect of the present invention, the dual-channel regulatory retrieval includes a first retrieval channel and a second retrieval channel.

[0009] As a further aspect of the present invention, the first retrieval channel is a semantic vector retrieval channel, which encodes the entire set of semantic conditions into an audit semantic vector, performs similarity calculation with the legal clause vectors in the vector database, and recalls candidate legal clauses based on cosine similarity as a class of candidate legal clauses. The second retrieval channel is the knowledge graph precise matching channel, which extracts the legal clauses of the review item from the semantic tree of the review item, performs entity retrieval and relationship matching in the legal knowledge graph, and obtains the matched legal clauses as the second type of candidate legal clauses.

[0010] As a further aspect of the present invention, based on a semantic condition set and a regulatory knowledge base, a dual-channel regulatory retrieval is performed to filter candidate regulatory clauses. After weighted fusion processing, a fused and ranked candidate set is output, as detailed below: The fusion ranking module integrates two categories of candidate regulatory clauses. Channel weights are dynamically configured based on the category of bidding and tendering review items. A weighted voting mechanism is used to calculate the fusion score of each candidate regulatory clause, and the clauses are ranked according to their fusion scores to obtain a fusion-ranked candidate set. When the review item belongs to the qualification review category, the weight of the exact matching channel is increased; when the review item belongs to the technical parameter review category, the weight of the semantic retrieval channel is increased.

[0011] As a further aspect of the present invention, the first retrieval channel is a semantic vector retrieval channel, which encodes the entire set of semantic conditions into an audit semantic vector, calculates the similarity between the vector vector and the legal clause vectors in the vector database, and recalls legal clauses based on cosine similarity as a class of candidate legal clauses. The specific steps are as follows: Extracting semantic condition sets ,in, For the i-th semantic branch combination, The number of semantic branch combinations; Inputting the set of semantic conditions into a pre-trained model in the legal domain yields the corresponding review semantic vector. ,in, Let i be the semantic vector of the audit item corresponding to the i-th semantic branch combination. For legal semantic encoding model; Extract the legal clause vectors from the vector database, and use cosine similarity to calculate the semantic matching degree between the audit semantic vector and the legal clause vector; Calculate and obtain all semantic retrieval contribution values, sort the semantic retrieval contribution values ​​in descending order to obtain a sequence of candidate legal clauses, and obtain the legal clauses corresponding to a preset number as a class of candidate legal clauses.

[0012] As a further aspect of the present invention, the second retrieval channel is a knowledge graph precise matching channel. It extracts the regulatory clauses of the review item from the semantic tree of the review item, performs entity retrieval and relation matching in the regulatory knowledge graph, and obtains the matched regulatory clauses as second-class candidate regulatory clauses. The specific steps are as follows: Extract the semantic branch combination corresponding to the legal clauses of the audit item from the semantic tree of the audit item, extract the audit keyword set and the audit scenario tag set, perform retrieval and matching in the legal knowledge graph, and obtain the matched legal clauses as the second type of candidate legal clauses. The keyword matching rate is calculated based on the set of approved keywords, and the tag matching rate is calculated based on the set of approved scenario tags. The graph matching contribution value is obtained by weighting the keyword matching rate and the tag matching rate.

[0013] As a further aspect of the present invention, the candidate legal provisions of category one and category two are fused based on the fusion sorting module. Channel weights are dynamically configured according to the category of bidding and tendering review items. The fusion score of each candidate legal provision is calculated using a weighted voting mechanism, and the fusion scores are sorted to obtain a fusion sorted candidate set. The specific steps are as follows: Identify the category to which the regulatory clauses of the review items belong, and dynamically configure the channel weight based on the category; the categories include qualification review, technical parameter review, commercial terms review, and bid evaluation rule review; The candidate legal provisions of Category I and Category II are merged to obtain a merged and ranked candidate set. Extract semantic retrieval contribution value and graph matching contribution value, combine with dynamically configured channel weights to calculate the fusion score of each candidate legal clause, and sort them in descending order of fusion score to obtain a fusion ranking candidate set.

[0014] As a further aspect of the present invention, conflict detection specifically involves determining whether there is a citation relationship, conflict relationship, or conflict of applicable conditions among candidate legal provisions.

[0015] A knowledge-enhanced intelligent review system for bidding and tendering regulations includes a regulatory knowledge graph construction module, a semantic parsing module for review items, a dual-channel regulatory retrieval module, a conflict detection and judgment module, and a compliance reasoning and tracing module. The regulatory knowledge graph construction module is used to obtain the regulatory texts in the bidding and tendering field, perform structured parsing and processing, obtain the regulatory knowledge graph and vector database, and integrate the regulatory knowledge graph and vector database to construct a regulatory knowledge base; The semantic parsing module for review items is used to perform natural language understanding analysis on the received bidding review items, extract semantic branches and combine them to construct a semantic tree for the review items, and form a set of semantic conditions; The dual-channel regulatory retrieval module is used to perform dual-channel regulatory retrieval based on a semantic condition set and a regulatory knowledge base, filter candidate regulatory clauses, and output a fused and ranked candidate set after weighted fusion processing. The conflict detection and judgment module is used to perform conflict detection and priority determination on the candidate set of fusion and ranking, and to determine the finally applicable legal provisions. The compliance reasoning traceability module is used to generate a compliance reasoning chain based on the applicable regulatory provisions, and output a structured audit report based on the compliance reasoning chain.

[0016] The technical effects and advantages of this invention, a knowledge-enhanced intelligent review method and system for bidding and tendering regulations, are as follows: This invention uses a semantic tree to structurally parse the review object, review attributes, and review thresholds, improving the standardization and accuracy of review condition extraction; it utilizes a semantic vector retrieval channel to overcome the limitations of traditional keyword matching, enabling the identification of the same legal meaning under different expressions, thus improving the recall and coverage of regulation retrieval; it uses a knowledge graph precise matching channel combined with keywords, scenario tags, and regulatory relationships for precise screening, improving the accuracy and applicability of regulation matching; and it uses a fusion ranking module to dynamically adjust the weight of the retrieval channel according to the review item category, and combines a weighted voting mechanism to comprehensively rank candidate regulatory clauses, ensuring that the final recommended regulatory clauses simultaneously consider semantic relevance and legal applicability. Through the mutual supplementation and verification of dual-channel retrieval results, the risk of regulatory omissions and false positives is effectively reduced, improving the reliability and interpretability of regulatory matching results, thereby reducing the workload of manual review and improving the efficiency, quality, legality, and traceability of bidding and tendering review results. This invention achieves high recall, high accuracy, and high applicability in regulatory retrieval, which not only improves the automation and intelligence of bidding review, but also enhances the legality, standardization, and interpretability of the review results, demonstrating significant technical effects and practical application value. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a knowledge-enhanced intelligent review method for bidding regulations and clauses, provided for an embodiment of the present invention; Figure 2 This is a system block diagram of a knowledge-enhanced intelligent review system for bidding regulations and clauses, provided as an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described technical solutions are only a part of this invention, and not all of it. All other technical solutions obtained by those skilled in the art based on the technical solutions of this invention without inventive effort are within the scope of protection of this invention.

[0019] Example 1: The executing entity of this application may include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment may include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment may include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not impose any limitations on this. Steps S1 to S5 are detailed below: Step S1 involves obtaining and structurally parsing legal texts in the bidding and tendering field to acquire a legal knowledge graph and a vector database. The legal knowledge graph and vector database are then integrated to construct a legal knowledge base. Specifically, the structural parsing process involves parsing legal clauses into five-tuples, each including the legal name, chapter number, original text of the clause, effective status, and applicable scenario label. The "applicable scenario label" is obtained through domain expert annotation or automatic classification, such as "pre-qualification," "bid evaluation method," and "objection handling." Using legal clauses as entity nodes in the knowledge graph, semantic relationships are established between these nodes to form the legal knowledge graph. Entity nodes can be effective status or applicable scenario labels. Semantic relationships include citation relationships, conflict relationships, completion relationships, exception relationships, and hierarchical relationships (i.e., higher-level laws prevail over lower-level laws, special laws prevail over general laws, and new laws prevail over old laws). A pre-trained legal domain model is used to semantically encode the original text, applicable scenario label, and legal relationships of each legal clause, generating a legal clause vector, which is then stored in the vector database. Step S2 involves receiving bidding and tendering review items and performing natural language understanding analysis to extract semantic branches and construct a semantic tree for the review items, forming a set of semantic conditions. The semantic branches include the review object, review attributes, and review threshold. Step S3: Based on the semantic condition set and the legal knowledge base, perform dual-channel legal retrieval, filter candidate legal clauses, and output a fused and ranked candidate set after weighted fusion processing; Step S4: For the candidate set of fusion and ranking, use the relationship information in the regulatory knowledge graph to perform conflict detection and priority determination to determine the finally applicable regulatory clauses. Step S5: Generate a compliance reasoning chain based on the applicable regulatory provisions, and output a structured audit report based on the compliance reasoning chain. The compliance reasoning chain includes the original content of the audit items, extraction of key compliance elements, dual-channel matching results, regulatory application process, logical deduction process, and final audit conclusion. Simultaneously, the audit confidence level is calculated from three dimensions: regulatory timeliness, regulatory matching accuracy, and clause consistency, forming a three-level confidence evaluation result. For low-confidence audit results, a manual review prompt is automatically generated.

[0020] Preferably, the dual-channel regulatory search includes a first search channel and a second search channel; The first retrieval channel is the semantic vector retrieval channel, which encodes the entire set of semantic conditions into an audit semantic vector, calculates the similarity between the vector vector and the legal clause vector in the vector database, and recalls candidate legal clauses based on the cosine similarity as a class of candidate legal clauses. The second retrieval channel is the knowledge graph precise matching channel, which extracts the legal clauses of the review item from the semantic tree of the review item, performs entity retrieval and relation matching in the legal knowledge graph, and obtains the matched legal clauses as the second type of candidate legal clauses. The fusion ranking module integrates two categories of candidate regulatory clauses. Channel weights are dynamically configured based on the category of bidding and tendering review items. A weighted voting mechanism is used to calculate the fusion score of each candidate regulatory clause, and the clauses are ranked according to their fusion scores to obtain a fusion-ranked candidate set. When the review item belongs to the qualification review category, the weight of the exact matching channel is increased; when the review item belongs to the technical parameter review category, the weight of the semantic retrieval channel is increased.

[0021] In this embodiment of the invention, the dual-channel regulatory retrieval mechanism achieves comprehensive and accurate regulatory matching for bidding and tendering review items through the synergistic effect of the first and second retrieval channels. The first retrieval channel is a semantic vector retrieval channel, whose core function is to encode the entire set of semantic conditions for the review items into a vector representation, extract the review semantic vectors using a pre-trained model in the legal domain, and calculate the similarity between the review conditions and the regulatory clause vectors in the vector database. The cosine similarity is used to measure the semantic relevance between the review conditions and the regulatory clauses, thereby recalling candidate regulatory clauses with high semantic matching, forming a class of candidate regulatory clauses. This channel can capture the complex semantic information and implicit logical relationships expressed in natural language, and is particularly suitable for review items such as technical indicators and descriptive clauses that have lower requirements for precise text matching but require understanding of semantic meaning.

[0022] The second retrieval channel is a knowledge graph precise matching channel, which utilizes the structured regulatory information stored in the regulatory knowledge graph for rigorous matching. Keywords and scenario tags, such as the audit object, audit attributes, and thresholds, are extracted from the semantic tree of the audit item. Entity nodes are retrieved in the knowledge graph, and the relationships between nodes, such as citation, supplementary, exception, and hierarchical relationships, are combined to obtain regulatory clauses directly related to the audit conditions, forming two categories of candidate regulatory clauses. This ensures the accuracy and traceability of the regulatory clauses, and is particularly suitable for audit items involving qualification levels, eligibility requirements, and legally mandatory clauses that require strict adherence to the regulatory text.

[0023] The fusion ranking module comprehensively ranks candidate regulatory clauses based on the dual-channel search results. By identifying the review item category, such as qualification review, technical parameter review, commercial clause review, or bidding rule review, it dynamically adjusts the weights of the two channels to emphasize semantic understanding or precise matching capabilities according to review needs. For example, for qualification review, the weight of the knowledge graph channel is increased to enhance regulatory accuracy; for technical parameter review, the weight of the semantic vector channel is increased to enhance the ability to capture semantic descriptions and indicator matching. The fusion ranking uses a weighted voting mechanism to calculate the final fusion score by combining the semantic contribution value and knowledge graph matching contribution value of each candidate clause, and ranks them according to the score to obtain the fusion-ranked candidate regulatory clause set, providing a reliable basis for subsequent conflict detection, priority determination, and final regulatory application.

[0024] Overall, the dual-channel regulatory retrieval combines the advantages of semantic understanding and structured rules, which not only improves the comprehensiveness and accuracy of regulatory matching, but also achieves intelligent weight adjustment and sorting optimization for different review item categories through the integrated sorting module. This ensures the applicability, legality and traceability of regulations in the bidding and tendering review process, and significantly improves the efficiency and automation level of regulatory review.

[0025] Preferably, the first retrieval channel is a semantic vector retrieval channel, which encodes the entire set of semantic conditions into a review semantic vector, calculates the similarity between this vector and the legal clause vectors in the vector database, and recalls legal clauses based on cosine similarity as a class of candidate legal clauses. The specific steps are as follows: Extracting semantic condition sets ,in, For the i-th semantic branch combination, The number of semantic branch combinations; ,in, For the audit object corresponding to the i-th semantic branch combination, For the i-th semantic branch combination, the corresponding audit attribute is... The review threshold corresponding to the i-th semantic branch combination; Inputting the set of semantic conditions into a pre-trained model in the legal domain yields the corresponding review semantic vector. ,in, Let i be the semantic vector of the audit item corresponding to the i-th semantic branch combination. For legal semantic encoding model; Extracting regulatory clause vectors from the vector database ,in, For the j-th review semantic vector, the original text of the regulation is... For the j-th review semantic vector, the applicable scenario tag is... The legal relationship corresponding to the j-th audit semantic vector; The semantic matching degree between the audit semantic vector and the legal clause vector is calculated using cosine similarity. The calculation formula is as follows: ; In the formula: The semantic retrieval contribution value between the i-th audit semantic vector and the j-th regulatory clause vector; Calculate and obtain all semantic retrieval contribution values, sort the semantic retrieval contribution values ​​in descending order to obtain a sequence of candidate legal clauses, and obtain the legal clauses corresponding to a preset number as a class of candidate legal clauses.

[0026] In this embodiment of the invention, the first retrieval channel uses a semantic vector retrieval method to match regulations for review items. Its core function is to overcome the limitations of traditional keyword retrieval, which heavily relies on textual expression, and to achieve a deeper understanding of the true legal meaning and business intent of the review items. During the bidding and tendering review process, different tender documents may use different expressions to describe the same review requirements. For example, expressions such as "possessing a Class A general contracting qualification for building construction," "possessing a Class A general contracting qualification," and "holding a Class A general contracting qualification certificate for building construction" may differ in wording, but they essentially express the same review requirements. By constructing a semantic condition set, the review object, review attributes, and review threshold are structurally represented, and a pre-trained model in the legal domain is used to uniformly encode the combinations of semantic branches. This converts textual expressions into review semantic vectors with legal semantic features, thereby enabling the identification of the same legal meaning implied in different expressions and improving the intelligence level of regulation retrieval.

[0027] After the semantic vectors for review are generated, they are matched with the legal clause vectors in the vector database. The legal clause vectors not only contain the original text of the regulations but also integrate information such as applicable scenario tags and regulatory relationships, enabling them to more comprehensively reflect the legal meaning and scope of application of the legal clauses. By mapping the semantic vectors of the review items and the legal clause vectors to the same semantic space, even if there are no completely identical keywords between the review item text and the original regulations, the relationship between the two can still be identified based on semantic similarity. For example, if the review item states "no major safety accidents in the past three years," the regulations might state "no major or above production safety liability accidents in the past three years." Traditional keyword retrieval struggles to accurately establish the connection between the two, while semantic vector retrieval can discover a high degree of consistency in legal meaning through distance relationships in the semantic space, thus achieving accurate matching.

[0028] Using cosine similarity to calculate the semantic matching degree between the review item's semantic vector and the regulatory clause's vector can effectively measure the directional consistency of the two types of vectors in a high-dimensional semantic space. A higher cosine similarity indicates a closer legal semantic relationship between the review item and the regulatory clause, and a stronger applicability of the regulatory clause to the current review item. By calculating the semantic retrieval contribution value of all regulatory clauses and sorting them according to their contribution value, the most relevant regulatory clauses to the review conditions can be quickly selected from a massive regulatory database, forming a set of candidate regulatory clauses. This process avoids traversing the entire text of regulations clause by clause, significantly improving regulatory retrieval efficiency while reducing the time required for manual law searches.

[0029] Furthermore, the semantic vector retrieval channel possesses strong generalization and scalability capabilities. When the regulatory database is continuously updated or the expression of review items changes, there is no need to redesign complex rules; simply re-encoding using a pre-trained model is sufficient to complete the semantic association analysis between the new regulations and the new review items. Therefore, this retrieval method not only improves the recall and coverage of regulatory matching but also enhances the system's adaptability to complex legal semantics, implicit logical relationships, and diverse textual expressions. This provides a high-quality foundation of candidate regulations for subsequent precise knowledge graph matching, fusion ranking, and regulatory applicability determination, thereby improving the accuracy, intelligence level, and reliability of the entire bidding and tendering regulatory review process.

[0030] Preferably, the second retrieval channel is a knowledge graph precise matching channel. It extracts the regulatory clauses for the review items from the semantic tree of the review items, performs entity retrieval and relationship matching in the regulatory knowledge graph, and obtains the matched regulatory clauses as second-class candidate regulatory clauses. The specific steps are as follows: Extract the semantic branch combination corresponding to the legal clauses of the audit item from the semantic tree of the audit item, extract the audit keyword set and the audit scenario tag set, perform retrieval and matching in the legal knowledge graph, and obtain the matched legal clauses as the second type of candidate legal clauses. The keyword matching rate is calculated based on the set of approved keywords, and the tag matching rate is calculated based on the set of approved scenario tags. The graph matching contribution value is obtained by weighting the keyword matching rate and the tag matching rate.

[0031] In this embodiment of the invention, the second retrieval channel, namely the knowledge graph precise matching channel, is mainly used in bidding review to achieve high-precision matching of legal clauses, ensuring that candidate legal clauses are not only semantically relevant but also logically and in accordance with applicable rules. Its core function is to utilize the structured semantic branch combination in the semantic tree of the review item to extract key information such as the review object, review attributes, and thresholds, forming a set of review keywords and a set of scenario tags. This information accurately describes the legal requirements and applicable scenarios of the review conditions. For example, for the review item "The bidder must possess a Class A general contracting qualification for building construction and have no major safety accidents in the past three years," keywords such as "general contracting," "Class A qualification," "safety accidents," and "past three years" are extracted, and scenario tags such as qualification level, pre-qualification, and safety records are generated, thereby providing a precise indexing basis for subsequent knowledge graph retrieval.

[0032] During the knowledge graph retrieval process, the review keywords are compared with the legal entity nodes in the legal knowledge graph. Precise matching identifies which legal clauses contain the key elements corresponding to the review requirements. Simultaneously, scenario tags are used for secondary filtering, calculating the match degree between legal nodes and review items in applicable scenarios, thereby eliminating legal clauses that, although containing keywords, have an incompatible scope of application. This mechanism ensures the candidate legal clauses are accurate in both semantics and applicability, improving the accuracy and reliability of legal retrieval.

[0033] To quantify the degree of regulatory matching, keyword matching rate and scenario tag matching rate were calculated separately. Keyword matching rate reflects the proportion of regulatory clauses containing keywords relevant to the review item, demonstrating the direct correlation between the regulatory clauses and the review conditions. Scenario tag matching rate reflects the degree of fit between the legal scenario applicable to the regulatory clauses and the scenario required for the review item, demonstrating the accuracy of the regulatory applicability. By weighting these two matching rates, a final graph matching contribution value was obtained. This contribution value is used to measure the overall applicability of the regulatory clauses to the current review item. Regulatory clauses with high contribution values ​​will be prioritized for inclusion in the second category of candidate regulations, providing a reliable basis for subsequent fusion ranking and conflict detection.

[0034] This invention fully utilizes the structured legal relationships and tag information in a regulatory knowledge graph to achieve precise screening of candidate regulatory clauses. Unlike the first search channel, which focuses on semantic similarity, the knowledge graph channel ensures that candidate regulatory clauses strictly adhere to the legal text and its logical relationships, preventing semantically similar but inapplicable or repealed clauses from entering the candidate set. Simultaneously, by combining keyword and scenario tag matching methods, it can quickly and accurately identify regulatory clauses that meet the review requirements in complex legal environments, improving the legality, compliance, and intelligence of bidding and tendering reviews, and generating traceable and explainable regulatory matching results.

[0035] Preferably, the candidate legal provisions of category one and category two are fused based on the fusion sorting module. Channel weights are dynamically configured according to the category of bidding and tendering review items. The fusion score of each candidate legal provision is calculated using a weighted voting mechanism, and the fusion scores are sorted to obtain the fusion sorted candidate set. The specific steps are as follows: Identify the category to which the regulatory clauses of the review items belong, and dynamically configure the channel weight based on the category; the categories include qualification review, technical parameter review, commercial terms review, and bid evaluation rule review; The candidate legal provisions of Category I and Category II are merged to obtain a merged and ranked candidate set. Extract semantic retrieval contribution value and graph matching contribution value, combine with dynamically configured channel weights to calculate the fusion score of each candidate legal clause, and sort them in descending order of fusion score to obtain a fusion ranking candidate set.

[0036] The purpose of this invention's fusion sorting module in processing first-class and second-class candidate regulatory clauses is to intelligently integrate multi-channel regulatory retrieval results, thereby improving the accuracy and applicability of candidate regulatory clauses. In actual bidding and tendering review scenarios, different review items have different requirements for regulatory matching. For example, qualification reviews emphasize the accuracy of the regulatory text and its legal binding force, while technical parameter reviews emphasize semantic understanding and indicator correspondence capabilities. The fusion sorting module first identifies the category to which the review item belongs, such as qualification review, technical parameter review, commercial clause review, or bid evaluation rule review, and then dynamically allocates the weights of the first and second search channels for each category of review item, adjusting the contribution ratio of the search channels according to the characteristics of the review item. For example, qualification reviews tend to increase the weight of the precise matching channel to ensure the accuracy of the regulatory clauses in terms of legal effect and applicable scenarios, while technical parameter reviews increase the weight of the semantic vector channel to fully understand indicator descriptions and semantic meanings.

[0037] After the channel weights are determined, the first and second categories of candidate legal provisions are merged to form a fusion ranking candidate set. Through a weighted voting mechanism, the contribution value of each legal provision in the semantic retrieval channel and the graph precise matching channel is weighted according to the dynamically configured channel weights to calculate the fusion score. This mechanism not only considers the matching effect of legal provisions in a single channel but also comprehensively evaluates their overall performance across multiple channels, ensuring that provisions that perform well in both channels receive higher fusion scores and are thus given priority in the ranking results.

[0038] Candidate legal provisions are ranked from highest to lowest according to their fusion scores, resulting in a fused ranking candidate set. This ranking ensures semantic relevance while also considering the accuracy and logical coherence of the legal text, thus providing a reliable basis for subsequent legal conflict detection, priority determination, and the selection of the final applicable legal provisions. The fusion ranking module enables intelligent integration and dynamic weight adjustment of multi-channel legal information, improving not only the accuracy and coverage of legal matching but also enhancing the automation and interpretability of the bidding and tendering review process, providing a scientific, reasonable, and traceable basis for review decisions.

[0039] Preferably, conflict detection specifically involves determining whether there are any citation relationships, conflicting relationships, or conflicting applicable conditions among candidate legal provisions.

[0040] The priority determination process is as follows: when a conflict is detected, priority is determined according to the pre-set rules for the application of laws and regulations in the legal knowledge graph, including rules such as superior laws taking precedence over subordinate laws, special laws taking precedence over general laws, and new laws taking precedence over old laws. The final applicable law is determined in conjunction with the effective status of the law. In cases where automatic determination is not possible, the case is marked as a case requiring manual review to ensure the reliability of the review conclusion.

[0041] Preferably, a compliance reasoning chain is generated based on the applicable regulatory provisions, and a structured audit report is output based on the compliance reasoning chain. The specific steps are as follows: A compliance reasoning chain is generated based on the applicable regulatory provisions. The reasoning chain includes the original content of the audit item, the semantic parsing results, the key audit element extraction results, the dual-channel retrieval results, the regulatory application process, the logical deduction process, and the final audit conclusion. The confidence level of the final applicable legal clauses is calculated from three dimensions: the timeliness of the regulations, the accuracy of the regulations matching, and the consistency of the clauses, forming a three-level confidence evaluation result. The lowest level of confidence review result will automatically prompt for manual review. Based on the inference chain results, a structured audit report is output. This report includes audit conclusions, applicable regulatory clauses, conflict resolution results, confidence levels, and an inference chain summary. It also records the regulatory version number, matching algorithm parameters, audit timestamp, and audit file identification information, generating a traceable audit file. When regulations are updated, an impact analysis of historical audit results is performed based on the traceable file to determine the degree of impact of regulatory changes on historical audit conclusions, thereby achieving continuous compliance management under dynamically updated regulations.

[0042] Example 2: Figure 1 This is a flowchart illustrating a knowledge-enhanced intelligent review method for bidding regulations and clauses provided in this embodiment of the invention. This embodiment uses a qualification requirement review item from a construction project bidding document as input. The original text of the review item is: "The bidder must possess a Class A or higher qualification for general contracting of building construction projects, and must not have had any major safety production accidents in the past three years (2023 to 2025)." This review item represents a typical qualification requirement review scenario, involving two dimensions: qualification level review and safety production record review. Upon receiving this review item, it is input into the knowledge-enhanced regulatory review platform as the object to be reviewed.

[0043] The initialization steps for the regulatory knowledge base are executed. During the deployment phase, the construction of the regulatory knowledge graph and vector database is completed. Texts of regulations such as the *Regulations on the Management of Qualifications of Construction Enterprises*, the *Implementation Regulations of the Bidding Law*, the *Production Safety Law*, and local regulations on construction market management are pre-collected, and each regulatory clause is parsed into entity nodes and stored in the knowledge graph. For example, the knowledge graph contains regulatory nodes such as N1 Article 8 of the *Regulations on the Management of Qualifications of Construction Enterprises*, N2 Article 32 of the *Implementation Regulations of the Bidding Law*, N3 Article 19 of the *Production Safety Law*, N4 Article 3 of the Ministry of Housing and Urban-Rural Development Order No. 22 on Construction Engineering Qualification Standards, and N5 Article 15 of a certain province's construction market management regulations. Simultaneously, reference relationships, refinement relationships, conflict relationships, and parallel application relationships are established between regulations. For example, a qualification level refinement relationship is established between N1 and N4, a reference relationship is established between N2 and N1, and a parallel application relationship and conflict relationship are established between N3 and N5, thus forming a complete regulatory logical network.

[0044] The process then moves to the semantic analysis stage for the audit items. A legal natural language processing model is used to segment the audit item text, identify entities, perform dependency analysis, and extract conditions. Semantic branch combinations are extracted to construct a semantic tree for the audit items. The audit items are broken down into three levels: audit object, audit attribute, and audit threshold. The audit object is identified as "bidder (general contractor)"; the audit attribute is identified as two parallel conditions: "qualification level" and "safety production record"; and the audit threshold is identified as "qualification level ≥ Level 1" and "number of safety accidents in the past three years = 0". Through this analysis, a structured set of semantic conditions is formed, realizing the transformation from unstructured bidding text to structured audit conditions.

[0045] After obtaining the semantic condition set, the dual-channel regulatory retrieval stage begins. The first retrieval channel is the semantic vector retrieval channel. The semantic condition set is encoded into a 512-dimensional semantic vector using the Lawformer pre-trained legal model, and its cosine similarity is calculated with 5238 regulatory clause vectors in the vector database. A similarity threshold of 0.75 is set, and all regulatory clauses are sorted, with the top 5 candidate regulations retrieved. The search results show that the similarity is 0.94 for Article 3 of the Ministry of Housing and Urban-Rural Development Order No. 22, the similarity is 0.91 for Article 8 of the "Regulations on the Management of Qualifications of Construction Enterprises," the similarity is 0.87 for Article 32 of the "Regulations for the Implementation of the Bidding Law," the similarity is 0.83 for Article 19 of the "Production Safety Law," and the similarity is 0.79 for Article 15 of the Local Construction Market Management Regulations. This forms the first category of candidate regulatory clauses.

[0046] Simultaneously, the second search channel initiated a precise matching process using a knowledge graph. The keyword set "general contracting, Level 1 and above, qualification, safety accident, past three years" was extracted from the semantic tree, along with the scenario tag set "qualification level, pre-qualification, safety record." Entity indexing and matching were then performed within the regulatory knowledge graph using these keywords and scenario tags. The search results showed that the tag "general contracting" in Article 8 of the "Regulations on the Management of Qualifications of Construction Enterprises" perfectly matched the audit item; the keyword "Level 1 and above qualification" in Article 3 of the Ministry of Housing and Urban-Rural Development Order No. 22 was completely consistent with the audit conditions; and the content of "major safety accidents" and "three-year safety record" in Article 19 of the "Production Safety Law" highly matched the audit item. Therefore, the above-mentioned regulatory nodes were successfully matched, forming a second set of candidate regulatory clauses.

[0047] After completing the dual-channel retrieval, a fusion and ranking process is performed. Since the current review item pertains to qualification-related review matters, the semantic vector channel weight is set to 0.4 and the knowledge graph precise matching channel weight to 0.6, according to the pre-set weight configuration scheme for qualification review scenarios. The fusion and ranking module obtains the scores from both channels and calculates the fusion score based on a weighted voting mechanism. It is found that Article 3 of Order No. 22 of the Ministry of Housing and Urban-Rural Development (N4) received the highest score in both channels, resulting in a fusion score of 0.955; Article 8 of the "Regulations on the Management of Qualifications of Construction Enterprises" (N1) received a fusion score of 0.930; Article 19 of the "Production Safety Law" (N3) received a fusion score of 0.872; and Article 32 of the "Implementation Regulations of the Bidding Law" (N2) also received a high score. Therefore, the system generates the final candidate regulatory clause ranking results based on the fusion scores.

[0048] The process then proceeds to the regulatory conflict detection and priority determination stage. Candidate regulations are mapped onto a regulatory knowledge graph, and the logical relationships between regulatory nodes are analyzed. During the detection process, it was found that Article 19 of the N3 "Production Safety Law" and Article 15 of a certain province's construction market management regulations both involve the identification of major safety accidents. However, the standards for identifying major safety accidents stipulated in the local regulation N5 are lower than those stipulated in the national law N3, thus indicating a regulatory conflict. The conflict determination is automatically executed according to the pre-set regulatory hierarchy rules in the knowledge graph. Based on the principle of "higher-level law prevails over lower-level law," the "Production Safety Law" is a higher-level law, and its legal effect is higher than that of lower-level regulations. Therefore, N3 is retained as the applicable regulation, while N5 is marked as a conflicting regulation and removed, retaining only reference information. This process ensures the accuracy and legality of the application of regulations.

[0049] After conflict resolution, a compliance inference chain is generated. The inference chain first records the original content of the audit items, then displays the semantic parsing results, including the audit object being the bidder, the audit attributes being qualification level and safety accident records, and the audit threshold being Level 1 qualification and no major safety accidents in the past three years. Subsequently, the dual-channel matching results and fusion score are displayed, with the matching regulations mainly including Article 8 of the "Regulations on the Management of Qualifications of Construction Enterprises," Article 32 of the "Implementation Regulations of the Bidding Law," and Article 19 of the "Production Safety Law." Next, the system executes logical deduction, determining that Level 1 qualification meets the requirements based on the qualification management regulations, and that the absence of major safety accidents in the past three years meets the integrity requirements based on the Production Safety Law. Simultaneously, it confirms that there are no unresolved conflicts between the various regulations, thus ensuring a complete and self-consistent logical chain.

[0050] After the inference chain is generated, the confidence level of the audit results is further calculated. A score is awarded based on factors such as the validity of the regulations, the strength of the dual-channel matching, the results of regulatory conflict resolution, and the timeliness of the regulations. Since all applicable regulations in this embodiment are currently valid, both channels successfully matched the core regulatory clauses, and regulatory conflicts have been automatically resolved using knowledge graph rules, the system assigns a "high confidence" rating.

[0051] Finally, a structured audit report is output. The report clearly states the audit conclusion as "compliant," primarily applicable regulations including Article 3 of the Ministry of Housing and Urban-Rural Development Order No. 22, Article 8 of the "Regulations on the Management of Qualifications of Construction Enterprises," and Article 19 of the "Production Safety Law." The audit confidence level is high. The reasoning chain summary indicates that both the qualification level requirements and safety accident requirements have clear legal basis and there are no applicable conflicts. It also records conflicts between N3 and N5 and completes the automatic processing according to the principle of prioritizing higher-level laws. An audit traceability file is automatically generated, saving the regulatory version number, search parameters, timestamp, and audit file ID for subsequent auditing, supervision, and historical review.

[0052] As can be seen from the above implementation process, the embodiments of the present invention fully verify the core technical aspects such as the knowledge graph and vector database fusion architecture, semantic tree parsing of review items, dual-channel retrieval, automatic determination of regulatory conflicts, and generation of interpretable reasoning chains, thereby realizing the automation, accuracy, and traceability of the bidding and tendering regulatory review process.

[0053] Example 3: A knowledge-enhanced intelligent review system for bidding regulations and clauses, comprising a regulatory knowledge graph construction module, a review item semantic parsing module, a dual-channel regulatory retrieval module, a conflict detection and judgment module, and a compliance reasoning and tracing module; the regulatory knowledge graph construction module is connected to the review item semantic parsing module, the review item semantic parsing module is connected to the dual-channel regulatory retrieval module, the dual-channel regulatory retrieval module is connected to the conflict detection and judgment module, and the conflict detection and judgment module is connected to the compliance reasoning and tracing module; The regulatory knowledge graph construction module is used to obtain the regulatory texts in the bidding and tendering field, perform structured parsing and processing, obtain the regulatory knowledge graph and vector database, and integrate the regulatory knowledge graph and vector database to construct a regulatory knowledge base; The semantic parsing module for review items is used to perform natural language understanding analysis on the received bidding review items, extract semantic branches and combine them to construct a semantic tree for the review items, and form a set of semantic conditions; The dual-channel regulatory retrieval module is used to perform dual-channel regulatory retrieval based on a semantic condition set and a regulatory knowledge base, filter candidate regulatory clauses, and output a fused and ranked candidate set after weighted fusion processing. The conflict detection and judgment module is used to perform conflict detection and priority determination on the candidate set of fusion and ranking, and to determine the finally applicable legal provisions. The compliance reasoning traceability module is used to generate a compliance reasoning chain based on the applicable regulatory provisions, and output a structured audit report based on the compliance reasoning chain.

[0054] like Figure 2 The diagram shown is a system block diagram of an intelligent review system for bidding regulations and clauses based on knowledge enhancement, according to an embodiment of the present invention. This system can be used to execute... Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0055] Through the above embodiments, this invention uses semantic trees to structurally parse the review objects, review attributes, and review thresholds, improving the standardization and accuracy of review condition extraction. It utilizes a semantic vector retrieval channel to overcome the limitations of traditional keyword matching, enabling the identification of the same legal meaning under different expressions, thus improving the recall and coverage of regulatory retrieval. It employs a knowledge graph precise matching channel, combining keywords, scenario tags, and regulatory relationships for accurate filtering, improving the accuracy and applicability of regulatory matching. A fusion ranking module dynamically adjusts the weight of the retrieval channels based on the review item category, and combines a weighted voting mechanism to comprehensively rank candidate regulatory clauses, ensuring that the final recommended regulatory clauses simultaneously consider semantic relevance and legal applicability. Through the mutual supplementation and verification of dual-channel retrieval results, the risk of missed and false detections of regulations is effectively reduced, improving the reliability and interpretability of regulatory matching results. This reduces the workload of manual review, improves the efficiency and quality of bidding and tendering reviews, and enhances the legality and traceability of review results. This invention achieves high recall, high accuracy, and high applicability in regulatory retrieval, which not only improves the automation and intelligence of bidding review, but also enhances the legality, standardization, and interpretability of the review results, demonstrating significant technical effects and practical application value.

[0056] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

[0057] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge enhancement-based intelligent auditing method for bidding regulations clauses, characterized in that, The specific steps are as follows: By acquiring and structurally parsing legal texts in the bidding and tendering field, a legal knowledge graph and a vector database are obtained, and the legal knowledge graph and vector database are integrated to construct a legal knowledge base. By receiving bidding and tender review items, natural language understanding analysis is performed to extract semantic branches and construct a semantic tree for the review items, forming a set of semantic conditions; Based on semantic condition sets and legal knowledge bases, dual-channel legal retrieval is performed to filter candidate legal clauses, and after weighted fusion processing, a fused and ranked candidate set is output. For the candidate set of fusion ranking, conflict detection and priority determination are performed using relational information in the regulatory knowledge graph to determine the finally applicable regulatory clauses. Generate a compliance reasoning chain based on the applicable regulatory provisions, and output a structured audit report based on the compliance reasoning chain.

2. The intelligent review method for bidding regulations and clauses based on knowledge enhancement according to claim 1, characterized in that, Semantic branch combinations include the audit object, audit attribute, and audit threshold. 3.The knowledge-enhanced bidding regulation clause intelligent auditing method according to claim 1, characterized in that, The dual-channel regulatory search includes a first search channel and a second search channel.

4. The knowledge-enhanced bidding regulation clause intelligent auditing method according to claim 3, characterized in that, The first retrieval channel is the semantic vector retrieval channel, which encodes the entire set of semantic conditions into an audit semantic vector, calculates the similarity between the vector vector and the legal clause vector in the vector database, and recalls candidate legal clauses based on the cosine similarity as a class of candidate legal clauses. The second retrieval channel is the knowledge graph precise matching channel, which extracts the legal clauses of the review item from the semantic tree of the review item, performs entity retrieval and relationship matching in the legal knowledge graph, and obtains the matched legal clauses as the second type of candidate legal clauses.

5. The knowledge-enhanced bidding regulation clause intelligent auditing method according to claim 4, characterized in that, Based on semantic condition sets and a legal knowledge base, a dual-channel legal retrieval is performed to filter candidate legal clauses. After weighted fusion processing, a fused and ranked candidate set is output, as follows: The fusion sorting module performs fusion processing on candidate legal provisions of category 1 and category 2. Channel weights are dynamically configured according to the category of bidding and tendering review items. The fusion score of each candidate legal provision is calculated using a weighted voting mechanism, and the fusion scores are sorted to obtain a fusion sorted candidate set.

6. The knowledge-enhanced bidding regulation clause intelligent auditing method according to claim 5, characterized in that, The first retrieval channel is the semantic vector retrieval channel. The entire set of semantic conditions is encoded into a review semantic vector, and its similarity is calculated with the legal clause vectors in the vector database. Legal clauses are then recalled based on cosine similarity as a class of candidate legal clauses. The specific steps are as follows: Extracting based on a set of semantic conditions wherein, is the i-th semantic branch combination, is the number of semantic branch combinations; Inputting the set of semantic conditions into a pre-trained model in the legal domain yields the corresponding review semantic vector. ,in, Let i be the semantic vector of the audit item corresponding to the i-th semantic branch combination. For legal semantic encoding model; Extract the legal clause vectors from the vector database, and use cosine similarity to calculate the semantic matching degree between the audit semantic vector and the legal clause vector; Calculate and obtain all semantic retrieval contribution values, sort the semantic retrieval contribution values ​​in descending order to obtain a sequence of candidate legal clauses, and obtain the legal clauses corresponding to a preset number as a class of candidate legal clauses.

7. The knowledge-enhanced bidding regulation clause intelligent auditing method according to claim 6, characterized in that, The second retrieval channel is the knowledge graph precise matching channel. It extracts the regulatory clauses of the review item from the semantic tree of the review item, performs entity retrieval and relation matching in the regulatory knowledge graph, and obtains the matched regulatory clauses as the second category of candidate regulatory clauses. The specific steps are as follows: Extract the semantic branch combination corresponding to the legal clauses of the audit item from the semantic tree of the audit item, extract the audit keyword set and the audit scenario tag set, and perform retrieval and matching in the legal knowledge graph to obtain the matched legal clauses as the second type of candidate legal clauses. The keyword matching rate is calculated based on the set of approved keywords, and the tag matching rate is calculated based on the set of approved scenario tags. The graph matching contribution value is obtained by weighting the keyword matching rate and the tag matching rate.

8. The knowledge-enhanced bidding regulation clause intelligent auditing method according to claim 7, characterized in that, The fusion and ranking module is used to fuse candidate legal provisions of category 1 and category 2. Channel weights are dynamically configured according to the category of bidding and tendering review items. A weighted voting mechanism is used to calculate the fusion score of each candidate legal provision, and the provisions are ranked according to the fusion score to obtain the fusion-ranked candidate set. The specific steps are as follows: Identify the category to which the regulatory clauses of the review items belong, and dynamically configure the channel weight based on the category; the categories include qualification review, technical parameter review, commercial terms review, and bid evaluation rule review; The candidate legal provisions of Category I and Category II are merged to obtain a merged and ranked candidate set. Extract semantic retrieval contribution value and graph matching contribution value, combine with dynamically configured channel weights to calculate the fusion score of each candidate legal clause, and sort them in descending order of fusion score to obtain a fusion ranking candidate set. 9.The knowledge-enhanced bidding regulation clause intelligent auditing method according to claim 1, wherein, Conflict detection specifically involves determining whether there are any citation relationships, conflicting relationships, or conflicting applicable conditions among candidate legal provisions.

10. A knowledge-enhanced intelligent auditing system for bidding regulations clauses, applied to the knowledge-enhanced intelligent auditing method for bidding regulations clauses according to any one of claims 1-9, characterized in that, It includes a regulatory knowledge graph construction module, an audit item semantic parsing module, a dual-channel regulatory retrieval module, a conflict detection and judgment module, and a compliance reasoning and tracing module; The regulatory knowledge graph construction module is used to obtain the regulatory texts in the bidding and tendering field, perform structured parsing and processing, obtain the regulatory knowledge graph and vector database, and integrate the regulatory knowledge graph and vector database to construct a regulatory knowledge base; The semantic parsing module for review items is used to perform natural language understanding analysis on the received bidding review items, extract semantic branches and combine them to construct a semantic tree for the review items, and form a set of semantic conditions; The dual-channel regulatory retrieval module is used to perform dual-channel regulatory retrieval based on a semantic condition set and a regulatory knowledge base, filter candidate regulatory clauses, and output a fused and ranked candidate set after weighted fusion processing. The conflict detection and judgment module is used to perform conflict detection and priority determination on the candidate set of fusion and ranking, and to determine the finally applicable legal provisions. The compliance reasoning traceability module is used to generate a compliance reasoning chain based on the applicable regulatory provisions, and output a structured audit report based on the compliance reasoning chain.

Images