A knowledge graph-based intelligent review method and system for tender documents

Abstract

This invention discloses a knowledge graph-based intelligent review method and system for tender documents. The method includes: determining the object to be reviewed based on the target tender document, and generating several logical association paths for the target content to be reviewed; loading structured rule sets and related knowledge points corresponding to the logical association paths of the target content to be reviewed, and performing a rationality judgment to obtain the rationality evaluation results of all logical association paths of the content to be reviewed; performing centralized comparison and consensus verification of the rationality evaluation results of all logical association paths of the content to be reviewed, and performing auxiliary review of the clauses to be reviewed in the target tender document based on the verification results to obtain the intelligent review result of the target tender document. This invention can improve the reliability of review reports and decision-making efficiency. As a knowledge graph-based intelligent review method and system for tender documents, this invention can be widely applied in the field of document content review technology.

Description

Technical Field

[0001] This invention relates to the field of document content review technology, and in particular to a knowledge graph-based intelligent review method and system for tender documents. Background Technology

[0002] In the field of public procurement, the review of tender documents is a core link in ensuring the compliance, fairness, and smooth implementation of projects. The traditional review model relies on human experts to check a massive amount of laws, regulations, industry standards, and internal rules line by line. When faced with increasingly complex cross-sectoral projects, this model has inherent defects such as low efficiency, difficulty in fully covering the correlation between clauses, and inconsistent understanding of standards by different reviewers, which can easily lead to biases in compliance judgments.

[0003] To address these issues, knowledge-based intelligent review systems have emerged, achieving automated information identification and compliance verification through preset rules, demonstrating high reliability in routine projects within a single domain. However, with the increasing number of large-scale integrated projects such as smart cities and cross-regional energy networks, existing systems have gradually revealed significant limitations: First, they struggle to effectively model implicit semantic relationships between cross-domain technical specifications, failing to capture unexpressed standard references and industry practices; second, they have low accuracy in identifying emerging non-standardized terms in cutting-edge interdisciplinary fields, making it difficult to map them to existing knowledge nodes; third, they cannot handle highly embedded non-linear logical dependencies between clauses, and simple rules cannot support multi-step context-sensitive reasoning; fourth, they lack mechanisms to adapt to fragmented regulatory updates, making it difficult to accurately integrate supplementary explanations, precedents, and other content into the reasoning chain; and fifth, complex multi-step queries suffer from severe performance bottlenecks, failing to meet the rapid decision-making needs of procurement processes.

[0004] The aforementioned issues result in insufficient accuracy and consistency of intelligent review systems in comprehensive projects, making them unable to effectively replace human review in completing high-quality reviews, thus becoming a key technological bottleneck restricting the digital transformation of public procurement. Summary of the Invention

[0005] To address the aforementioned technical problems, the purpose of this invention is to provide a knowledge graph-based intelligent review method and system for tender documents, which can improve the reliability of review reports and decision-making efficiency.

[0006] The first technical solution adopted in this invention is: a knowledge graph-based intelligent review method for tender documents, comprising the following steps: Based on the target tender documents, identify the objects to be reviewed and generate several logical connection paths for the target content to be reviewed; Load the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and make a reasonableness judgment to obtain the reasonableness assessment results of all logically related paths of the content to be reviewed. The reasonableness assessment results of all logically related paths of the content to be reviewed are centrally compared and verified for consensus. Based on the verification results, the clauses to be reviewed in the target tender documents are assisted in the review, and the intelligent review results of the target tender documents are obtained.

[0007] Furthermore, the step of identifying the objects to be reviewed based on the target tender documents and generating several logically related paths for the target content to be reviewed specifically includes: Based on the target tender documents, the objects to be reviewed were identified; If the object to be reviewed is the entire target tender document, then the project context features of the entire target tender document are extracted and a conflict solution map is constructed. Based on the nonlinear mapping relationship between the project context features and the conflict solution map, several logical association paths of the target content to be reviewed are generated. If the object to be reviewed is a clause in the target tender document, the information source of the clause to be reviewed is first determined, and then semantic analysis is performed on the clause to be reviewed to generate several logical connection paths of the target content to be reviewed.

[0008] Furthermore, if the object to be reviewed is the entire target tender document, then the step of extracting project context features from the entire target tender document and constructing a conflict solution map, and generating several logically related paths for the target content to be reviewed based on the nonlinear mapping relationship between the project context features and the conflict solution map, specifically includes: If the object to be reviewed is the entire target tender document, then the target tender document is parsed and entity recognized by natural language processing methods to obtain the contextual features of the target tender document, which include project name, technical keywords and time period; The contextual features of the target tender documents are vectorized and verified against preset geographic information system data to obtain the verified contextual feature vector. Based on the preset potential regulatory conflict types, vectorization processing is performed to generate context modifiers. The preset potential regulatory conflict types include the type of the issuing entity of the target tender document, the main core concepts of the target tender document, and the legal basis of the target tender document. Context simulation is performed on context modifiers, all context simulation conflict patterns are obtained and matched with the verified context feature vectors to construct a conflict solution map; The nonlinear mapping relationship between the conflict solution map and the preset standard legal provisions is obtained, and several logical association paths of the target content to be reviewed are generated. The nonlinear mapping relationship represents the explanatory basis for finding the corresponding conflict solution map based on the preset standard legal provisions.

[0009] Furthermore, the step of first determining the information source of the clause to be reviewed in the target tender document, and then performing semantic analysis on the clause to generate several logical connection paths of the target content to be reviewed, if the object to be reviewed is a clause to be reviewed in the target tender document, specifically includes: If the object to be reviewed is a clause to be reviewed in the target tender document, it is compared with a preset list of informal information sources, which indicates that the clause to be reviewed does not originate from preset standard legal clauses. If the clauses to be reviewed in the target tender document are not in the list of informal information sources, then semantic analysis is performed on the clauses to be reviewed in the target tender document, and domain knowledge is classified according to the results of the semantic analysis to generate several logical connection paths of the target content to be reviewed. If the clauses to be reviewed in the target tender document belong to an informal information source list, then the contextual semantics of each word in the clauses to be reviewed are obtained and contextual semantics are extracted and intent consensus is generated to obtain several logical connection paths of the target content to be reviewed.

[0010] Furthermore, the step of performing semantic analysis on the clauses to be reviewed in the target tender document if they do not belong to the informal information source list, and then classifying them according to domain knowledge based on the semantic analysis results to generate several logical connection paths for the target content to be reviewed, specifically includes: If the clauses to be reviewed in the target tender document are not in the list of informal information sources, perform semantic analysis on the clauses to be reviewed in the target tender document, obtain the relationship between the words in the clauses to be reviewed, and perform transformation processing to generate context vectors of several words. The similarity calculation is performed between the context vectors of several words and the known concept knowledge vectors in the preset concept knowledge base to obtain the similarity calculation results. The preset concept knowledge base has several professional term knowledge vectors. Based on the similarity calculation results, domain knowledge classification is performed on each word in the clause to be reviewed to determine the associated domain of each word in the clause to be reviewed; Obtain the chapter titles in the target tender document where each term in the clause to be reviewed is located, and generate several logical connection paths for the target content to be reviewed by combining the related fields corresponding to each term in the clause to be reviewed.

[0011] Furthermore, the step of obtaining the contextual semantic extraction and intent consensus generation of the various words in the target tender document to obtain several logical connection paths of the target content to be reviewed, if the clauses to be reviewed in the target tender document belong to an informal information source list, specifically includes: If the clauses to be reviewed in the target tender document belong to an informal information source list, the keywords in the clauses to be reviewed are extracted by combining the context of the words in the clauses to be reviewed, and the key text fragments of the clauses to be reviewed are obtained. Semantic segmentation is performed on the key text fragments of the clauses to be reviewed to obtain the segmented key text fragments of the clauses to be reviewed. Semantic extraction and intent consensus generation are performed on the key text fragments of the segmented clauses to be reviewed. The semantic intent vectors of the key texts are obtained and similarity is calculated to obtain the semantic similarity score of the key texts. The semantic similarity scores of key texts are converted into the intent confidence scores of the clauses to be reviewed. Combined with the related domains corresponding to each word in the clauses to be reviewed, several logical connection paths of the target content to be reviewed are generated.

[0012] Furthermore, the step of semantic extraction and intent consensus generation of the segmented key text fragments of the clauses to be reviewed, obtaining the semantic intent vector of the key text and calculating its similarity to obtain the semantic similarity score of the key text, specifically includes: Deep semantic analysis is performed on the segmented key text fragments of the clauses to be reviewed using a pre-trained context-sensitive language model to obtain the semantic representation of the key text fragments; By using a pre-trained intent classifier and a pre-trained named entity recognition (NER) model, the semantic representation of key text fragments is subjected to intent recognition and concept extraction, respectively, to obtain the semantic intent vector and semantic concept vector of the key text. The semantic similarity score of the key text is obtained by calculating the similarity between the semantic intent vector and the semantic concept vector of the key text.

[0013] Furthermore, the step of loading the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and performing a reasonableness judgment to obtain the reasonableness assessment results of all logically related paths of the content to be reviewed, specifically includes: Load a set of structured rules and related knowledge points corresponding to several logically related paths of the target content to be reviewed. The set of structured rules consists of logical statements of condition-association-conclusion, and the related knowledge points are derived from a preset concept knowledge base. By using recursive queries and graph traversal, we obtain all logical association paths of the target words and their non-linear dependencies on the structured rule set and associated knowledge points. Based on non-linear dependencies, association weight values ​​are set, and the reasonableness of all logical association paths of the target words is judged based on the association weight values, so as to obtain the reasonableness assessment results of all logical association paths of the content to be reviewed.

[0014] Furthermore, the step of centrally comparing and verifying the reasonableness assessment results of all logically related paths of the content to be reviewed, and then conducting auxiliary review of the clauses to be reviewed in the target tender document based on the verification results to obtain the intelligent review result of the target tender document, specifically includes: The reasonableness assessment results of all logically related paths of the content to be reviewed are centrally compared and consensus verified. The reasonableness assessment results include reasonable conclusions and confidence scores. If the reasonableness assessment results of all logical connection paths of the content to be reviewed have a consistent reasonableness conclusion and the confidence score is higher than the preset score threshold, then all logical connection paths of the target word will be merged and output to obtain a consistency conclusion report. If the reasonableness assessment results of all logical connection paths of the content to be reviewed are inconsistent or the confidence score is lower than the preset score threshold, then all logical connection paths of the target word will be output independently to obtain a multi-path difference report. Based on the consistency conclusion report or the multi-path difference report, the clauses to be reviewed in the target tender document are reviewed to obtain the intelligent review result of the target tender document.

[0015] The second technical solution adopted in this invention is: a knowledge graph-based intelligent review system for tender documents, comprising: The first module is used to determine the objects to be reviewed based on the target tender documents and generate several logical association paths for the target content to be reviewed. The second module is used to load the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and to make a reasonableness judgment to obtain the reasonableness evaluation results of all logically related paths of the content to be reviewed. The third module is used to centrally compare and verify the rationality assessment results of all logically related paths of the content to be reviewed, and to conduct auxiliary review of the clauses to be reviewed in the target tender document based on the verification results, so as to obtain the intelligent review result of the target tender document.

[0016] The beneficial effects of the method and system of this invention are as follows: Based on the target tender document, this invention identifies the object to be reviewed and generates several logically related paths for the target content to be reviewed. It further loads the structured rule sets and related knowledge points corresponding to these logically related paths to perform a rationality judgment, obtaining the rationality evaluation results of all logically related paths for the content to be reviewed. By exploring multiple possible logical reasoning paths in parallel, the system can evaluate the same clause from different dimensions and based on different rule sets. When these paths lead to inconsistent conclusions, the system does not simply select one, but transparently reveals these differences and elaborates on the logical basis and knowledge points leading to the disagreement. Finally, the rationality evaluation results of all logically related paths for the content to be reviewed are centrally compared and consensus verified. Based on the verification results, the system performs auxiliary review of the clauses to be reviewed in the target tender document, obtaining the intelligent review result of the target tender document. By actively exploring multiple logical reasoning paths and performing consensus verification, the accuracy and consistency of intelligent review of tender documents can be significantly improved. It can effectively identify and make transparent the potential logical conflicts and ambiguities in complex clauses, thereby significantly reducing the review burden on human reviewers when dealing with large, comprehensive, and cross-domain tender documents due to information dispersion, implicit correlations, and nonlinear dependencies, and ultimately improving the reliability of review reports and decision-making efficiency. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of an intelligent review method for tender documents based on knowledge graphs, as described in this invention. Figure 2 This is a structural block diagram of a knowledge graph-based intelligent review system for tender documents according to the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0019] Reference Figure 1 This invention provides a knowledge graph-based intelligent review method for tender documents, which includes the following steps: S100. Based on the target tender documents, determine the objects to be reviewed and generate several logical connection paths for the target content to be reviewed; S110. Based on the target tender documents, determine the objects to be reviewed; S120. If the object to be reviewed is the entire target tender document, then the project context features of the entire target tender document are extracted and a conflict solution map is constructed. Based on the nonlinear mapping relationship between the project context features and the conflict solution map, several logical association paths of the target content to be reviewed are generated. First, it is important to clarify that in the review of large-scale, comprehensive, and cross-domain tender documents, when specific clauses (such as clauses on data interoperability) need to refer to emerging and non-standardized principles (such as cross-regional data sovereignty in smart cities), and these principles mainly exist in the form of fragmented, informal guiding documents, industry white papers, or administrative interpretations, how can we ensure that the semantic parsing unit can accurately capture the subtle semantic differences in different contexts, and support the path generator to construct highly confident and distinguishable logical association paths for these principles, so as to effectively drive the parallel inference engine to perform accurate multi-step logical inference?

[0020] Based on this, this embodiment addresses the problem in the review of large, comprehensive, and cross-domain tender documents where specific clauses require reference to emerging and non-standardized principles. These principles exist in fragmented and informal forms, leading to difficulties in semantic parsing and logical path construction. A context-driven temporary semantic aggregation and intent consensus extraction method is proposed. The core of this method is that when the intelligent review system identifies that a clause to be reviewed requires reference to an emerging and non-standardized principle, it does not directly rely on a pre-defined, fully structured knowledge representation. Instead, the system activates a multi-source information aggregator, selectively extracting the most relevant text fragments from multiple scattered and informal sources in real time, based on the context and semantic focus of the current clause. Subsequently, a contextual semantic extraction unit utilizes deep semantic analysis technology to perform detailed intent identification and key concept extraction on these extracted text fragments. By comparing descriptions of the same concept or behavior from different sources, it identifies and quantifies the core intent and potential compliance considerations in real time, forming a temporary intent consensus statement valid only for the current review task. This statement exists in the form of a set of logical predicates and intent descriptions, representing the core understanding of the non-standardized principle in the current specific context. The path generator will construct subsequent logically related paths based directly on this instant-generated intent consensus statement, rather than pre-defined static rules.

[0021] S121. If the object to be reviewed is the entire target tender document, then the target tender document is parsed and entity recognized by natural language processing methods to obtain the contextual features of the target tender document. The contextual features include project name, technical keywords and time period. S122. Vectorize the contextual features of the target tender document and verify them with the preset geographic information system data to obtain the verified contextual feature vector. S123. Based on the preset potential regulatory conflict types, perform vectorization processing to generate context modifiers. The preset potential regulatory conflict types include the type of the issuing entity of the target tender document, the main core concepts of the target tender document, and the legal basis of the target tender document. In this embodiment, the system maintains a "regulatory conflict pattern library," which stores known or potential types of regulatory conflicts, as well as possible solutions or priority rules for these conflicts under different context modifiers. This library not only includes conflicts between formal regulations, but also focuses on collecting and representing potential semantic discrepancies and applicability differences between informal interpretive texts (such as industry announcements and administrative memoranda).

[0022] The first is the definition of conflict patterns. Each conflict pattern definition includes: the core concepts involved in the conflict (e.g., "carbon emission cap", "data sovereignty"), the typical manifestations of the conflict (e.g., "conflict between absolute cap and offsetting mechanism", "conflict between local storage and cross-regional transmission"), and the types of issuing entities that may be involved (e.g., "government agencies", "industry alliances").

[0023] Then, a mapping between context modifiers and solutions is performed. For each conflict mode, a series of "context modifiers" are defined (e.g., "the project is located in an ecologically sensitive area," "the data involves personal privacy information"), and solutions are preset or derived for each combination of modifiers. For example, for "carbon emission cap conflict," a solution can be defined as follows: 1) IF (the scenario is "ecologically sensitive area") THEN (the solution is "strictly adhere to the absolute upper limit, and the offsetting mechanism is not applicable"); 2) IF (the scenario is “non-ecologically sensitive area”) THEN (the solution is “allowing a dynamic carbon emission offset credit mechanism”).

[0024] The final solution is stored in the form of structured logical rules or priority instructions, for example, using first-order logical predicates or decision tree nodes. By pre-building and maintaining a comprehensive conflict pattern library, the system can identify various conflicts that may arise in complex regulatory environments and provide contextualized, pre-defined resolution strategies for these conflicts, laying the foundation for subsequent predictive resolution.

[0025] S124. Perform scenario simulation on scenario modifiers, obtain all scenario simulation conflict patterns and match them with the verified scenario feature vectors to construct a conflict solution map. In this embodiment, after the project context features are extracted, the system launches a "predictive conflict solution generator." This generator iterates through all conflict patterns in the "regulatory conflict pattern library" and simulates the application of context modifiers under each conflict pattern for the current project's "project context description vector." Through this simulation, the system can pre-calculate explicit solutions for all known or potential regulatory conflicts within the overall context of the current project.

[0026] First, context matching and rule activation are performed. The generator matches the current project's "project context description vector" with the "context modifiers" of each conflict pattern in the conflict pattern library. For example, if the project context includes "ecologically sensitive area", then all conflict resolution rules related to "ecologically sensitive area" are activated.

[0027] Next comes solution derivation. For each activated conflict mode, the generator retrieves the corresponding preset solution from the library based on the matched context modifier. If multiple solutions or priority rules exist, the generator will adjudicate according to preset meta-rules (e.g., "interpretations from government agencies take precedence over guidance from industry alliances") to ensure that a unique and explicit solution is generated for each conflict.

[0028] Finally, a conflict resolution graph is constructed, integrating all pre-computed conflict solutions into a single "conflict resolution graph." This graph is stored in key-value pairs, where the key is a unique identifier of the conflict (e.g., the type of clause involved, the conflict concept), and the value is the explicit solution to the conflict in the current project context (e.g., a specific logical rule or a priority instruction).

[0029] In summary, this embodiment can resolve all potential conflicts before the chain reasoning begins, providing the chain reasoning module with a "conflict-free map" that avoids the complexity and uncertainty of real-time conflict handling during reasoning, thereby ensuring the accurate one-time construction of the logic chain.

[0030] S125. Obtain the non-linear mapping relationship between the conflict solution map and the preset standard legal clauses, and generate several logical association paths of the target content to be reviewed. The non-linear mapping relationship represents the explanatory basis for finding the corresponding conflict solution map based on the preset standard legal clauses.

[0031] In this embodiment, the "conflict solution map" generated by the "predictive conflict solution generator" is passed to the "chain reasoning module." When the chain reasoning module performs multi-step logical inference and traces the non-linear dependencies between clauses and external standards or clauses, it no longer needs to evaluate and resolve regulatory conflicts in real time. Instead, it consults this pre-generated map to directly obtain explanations of conflicts that have been resolved in the current project context.

[0032] First, conflict point identification and graph query are performed. At each step of the chain reasoning, when the module identifies a key concept or rule that may involve regulatory conflict (e.g., when assessing “carbon emission caps”), it generates a query key (e.g., “carbon emission cap conflict”) and queries the “conflict solution graph”.

[0033] If a matching solution exists in the graph, the chained reasoning module immediately applies that solution as an explicit rule for the current logical step. For example, if the graph indicates that "carbon emission caps must be strictly adhered to in ecologically sensitive areas, and offsetting is not allowed," the module will directly use this rule for judgment, without considering offsetting mechanisms from industry alliances.

[0034] Finally, based on the explicit solutions provided by the graph, the chained reasoning module can smoothly construct a highly reliable and unified logical chain, ensuring that each inference step is based on resolved, contextualized rules. This embodiment, guided by the graph, enables the chained reasoning module to avoid ambiguity and repeated attempts in complex situations, significantly reducing computational time caused by real-time conflict resolution and ensuring the accuracy and consistency of the review results.

[0035] S130. If the object to be reviewed is a clause to be reviewed in the target tender document, the information source of the clause to be reviewed shall be determined first, and then semantic analysis shall be performed on the clause to be reviewed to generate several logical connection paths of the target content to be reviewed.

[0036] First, it's important to clarify that in the review of large-scale, comprehensive, and cross-sectoral tender documents, when the "chain-reasoning module" performs multi-step logical inference, the external regulatory requirements it traces are no longer single, static official documents, but rather manifested as explanatory texts issued by different entities in fragmented and informal forms. These texts may exhibit semantic discrepancies or applicability differences in specific contexts (e.g., conflicts between industry-wide guidelines and regional administrative interpretations). The challenge lies in designing a processing method that enables the "chain-reasoning module" to effectively assess the differences in the validity of these multi-source informal information and, based on the specific context of the current review clause, accurately identify and determine the order of application of these conflicting statements. This would allow it to construct a highly reliable and unified logical chain even in a complex and evolving regulatory environment, ensuring the accuracy and consistency of the review results.

[0037] Based on this, the focus is on proactively identifying and resolving potential regulatory conflicts based on the overall project context before detailed chain reasoning begins. This ensures that the chain reasoning module can operate according to a clear and unambiguous set of rules when dealing with fragmented and context-dependent regulatory interpretations.

[0038] S131. If the object to be reviewed is a clause to be reviewed in the target tender document, then it is judged against a preset list of informal information sources, wherein the preset list of informal information sources indicates that the clause to be reviewed does not originate from preset standard legal clauses. In this embodiment, when the semantic parsing unit in the intelligent review system initially identifies that the clause to be reviewed (e.g., "the platform should have the ability to seamlessly exchange data with existing municipal systems and future third-party applications") may involve emerging and non-standardized principles (e.g., the "data sovereignty principle"), the system will activate a "multi-source information aggregator". This aggregator does not blindly collect all information, but rather strategically performs real-time data capture and text extraction from a preset list of informal information sources based on the semantic focus of the clause and the initially inferred relevant areas.

[0039] First, information sources are configured. The system maintains a configurable list of informal information sources, including but not limited to official websites of specific international organizations, white paper repositories published by industry alliances, policy explanation pages of government agencies, and selected professional forums or blogs. These sources are assigned different weights based on their authority and relevance.

[0040] When a clause triggers a query for non-standardized principles, the aggregator uses web crawling modules (e.g., the Python-based Scrapy framework or the Go-based Colly library) to access configured information sources in real time. It performs initial filtering of the crawled text content based on keywords in the clause (such as "data sovereignty," "cross-regional data," and "localization") and contextual information provided by the semantic parsing unit, retaining only text fragments highly relevant to the current review task (e.g., paragraphs or sentences containing keywords).

[0041] Finally, the aggregator further uses text segmentation techniques to break down long documents into smaller, semantically complete text fragments (e.g., sentences or phrases). These fragments will serve as input to the contextual semantic extraction unit.

[0042] S132. If the clauses to be reviewed in the target tender document do not belong to the list of informal information sources, then perform semantic analysis on the clauses to be reviewed in the target tender document, and classify them according to the domain knowledge based on the results of the semantic analysis to generate several logical connection paths of the target content to be reviewed. Specifically, if the clauses to be reviewed in the target tender document do not belong to the informal information source list, semantic analysis is performed on the clauses to be reviewed in the target tender document to obtain the contextual relationships of each word in the clauses to be reviewed, and the transformation process is performed to generate context vectors of several words; the context vectors of several words are compared with the known concept knowledge vectors in the preset concept knowledge base to obtain the similarity calculation results, wherein the preset concept knowledge base has several professional term knowledge vectors; based on the similarity calculation results, the domain knowledge is classified for each word in the clauses to be reviewed to determine the associated domains corresponding to each word in the clauses to be reviewed; the chapter titles in the target tender document where each word in the clauses to be reviewed is located are obtained, and combined with the associated domains corresponding to each word in the clauses to be reviewed, several logical association paths of the target content to be reviewed are generated.

[0043] In this embodiment, when the intelligent review system receives a clause to be reviewed from a large, comprehensive, cross-domain tender document, the system activates a "semantic parsing unit" to perform in-depth semantic analysis. This unit first uses a "context vector generator" to process the clause text. This generator analyzes the contextual relationships between words in sentences, converting each word and phrase in the clause into a numerical representation (i.e., a context vector), thereby capturing its true meaning in a specific context. This is particularly important for understanding implicit semantic connections and emerging non-standardized terms in tender documents. For example, for a term like "data interoperability," it not only recognizes its literal meaning but also, based on its appearance in the "Information Technology Requirements" section, understands its deeper meaning, such as data formats and interface standards.

[0044] Next, a "concept similarity calculation module" compares these context vectors with existing concept vectors in the system's internal "concept knowledge base." If it encounters emerging technical terms that the system's vocabulary or entity recognition program cannot accurately identify, this module will attempt to calculate their similarity to known concepts and initially categorize them into the closest technical category.

[0045] Subsequently, a "path generator" proactively explores and generates multiple possible logically related paths based on the core intent of the clause, its chapter information, and the initially identified related fields. These paths are not limited to the rules explicitly cited in the tender document, but rather infer the relevant laws, regulations, industry standards, or internal rules based on the semantic implications of the clause. For example, for a clause about "data interoperability," the path generator will query the "concept knowledge base" for pre-defined relationships related to concepts such as "data," "interoperability," "exchange," "standards," and "security," thereby generating at least two reasoning paths: one may focus on national information security standards (e.g., pointing to the "National Information Security Level Protection Regulations"), and the other may focus on international data exchange protocols (e.g., pointing to the "Open Application Programming Interface Specification"), even if the tender document does not explicitly mention these specific standards. This multi-path exploration mechanism can effectively address implicit semantic relationships and non-linear logical dependencies in large, comprehensive, and cross-domain tender documents.

[0046] S133. If the clauses to be reviewed in the target tender document belong to an informal information source list, then obtain the contextual semantic extraction and intent consensus of each word in the clauses to be reviewed, and obtain several logical connection paths of the target content to be reviewed.

[0047] Specifically, if the clauses to be reviewed in the target tender document belong to an informal information source list, the keywords in the clauses to be reviewed are extracted by combining the context of the words in the clauses to be reviewed, resulting in key text fragments of the clauses to be reviewed; semantic segmentation is performed on the key text fragments of the clauses to be reviewed, resulting in segmented key text fragments of the clauses to be reviewed; semantic refinement and intent consensus generation are performed on the segmented key text fragments of the clauses to be reviewed, obtaining the semantic intent vector of the key text and calculating its similarity, resulting in the semantic similarity score of the key text; the semantic similarity score of the key text is converted into the intent confidence of the clauses to be reviewed, and combined with the related domains corresponding to the words in the clauses to be reviewed, several logical association paths of the target content to be reviewed are generated.

[0048] Specifically, a pre-trained context-sensitive language model is used to perform deep semantic analysis on the segmented key text fragments of the clauses to be reviewed, resulting in semantic representations of the key text fragments. A pre-trained intent classifier and a pre-trained named entity recognition (NER) model are used to perform intent recognition and concept extraction on the semantic representations of the key text fragments, respectively, resulting in semantic intent vectors and semantic concept vectors of the key texts. The similarity between the semantic intent vectors and semantic concept vectors of the key texts is calculated to obtain the semantic similarity score of the key texts.

[0049] In this embodiment, when the intelligent review system receives a large, comprehensive, and cross-domain tender document, it activates a "project context feature extractor." This extractor deeply analyzes various parts of the tender document, as well as any accompanying project description documents, to identify and extract key context features related to the project. These features include, but are not limited to, the project's geographical location (e.g., whether it is located in an "ecologically sensitive area" or "specific economic development zone"), project type (e.g., "infrastructure construction," "smart city transformation," "energy network upgrade"), project scale, core technology areas involved, and any explicitly mentioned special requirements or restrictions.

[0050] First, text parsing and entity recognition are performed using Natural Language Processing (NLP) techniques, such as Named Entity Recognition (NER) models based on the Transformer architecture. This scans the text in the tender documents to identify entities such as geographical location, project name, technical keywords, and timeframes. For example, it identifies "ecologically sensitive area" in "Project Location: A certain ecologically sensitive area".

[0051] Then, contextual features are vectorized, transforming the identified contextual features into standardized numerical representations or labels. For example, "ecologically sensitive area" can be mapped to a Boolean value (1 for yes, 0 for no), or a more fine-grained classification label. These standardized features will constitute a "project contextual description vector".

[0052] Further contextual feature verification is performed by cross-validating with pre-defined Geographic Information System (GIS) data or industry classification standards to ensure the accuracy and completeness of the extracted contextual features. For example, verifying whether "a certain ecologically sensitive area" is indeed marked as a sensitive area in the GIS database.

[0053] Therefore, this embodiment can provide accurate and consistent input for subsequent conflict identification and solution pre-calculation through comprehensive and standardized context feature extraction, avoiding judgment bias caused by missing or ambiguous context information.

[0054] When the semantic parsing unit in the intelligent review system initially identifies that the clause to be reviewed (e.g., "the platform should have the ability to seamlessly exchange data with existing municipal systems and future third-party applications") may involve emerging and non-standardized principles (e.g., the "data sovereignty principle"), the system will activate a "multi-source information aggregator." This aggregator does not blindly collect all information, but rather strategically performs real-time data capture and text extraction from a pre-set list of informal information sources based on the semantic focus of the clause and the initially inferred relevant areas.

[0055] First, information sources are configured. The system maintains a configurable list of informal information sources, including but not limited to official websites of specific international organizations, white paper repositories published by industry alliances, policy explanation pages of government agencies, and selected professional forums or blogs. These sources are assigned different weights based on their authority and relevance.

[0056] Then, real-time crawling and filtering are performed. When a clause triggers a query for non-standardized principles, the aggregator uses web crawling modules (e.g., the Python-based Scrapy framework or the Go-based Colly library) to access the configured information sources in real time. It performs preliminary filtering on the crawled text content based on keywords in the clauses (such as "data sovereignty," "cross-regional data," and "localization") and contextual information provided by the semantic parsing unit, retaining only text fragments highly relevant to the current review task (e.g., paragraphs or sentences containing keywords).

[0057] Furthermore, for the filtered text, the aggregator uses text segmentation techniques to break down long documents into smaller, semantically complete text fragments (e.g., sentences or phrases). These fragments will serve as input to the contextual semantic extraction unit.

[0058] The text fragments extracted by the multi-source information aggregator are sent to the "contextual semantic extraction unit". The core task of this unit is to deeply analyze the semantics of these fragmented texts, identify the core intentions and key concepts contained therein, and extract a temporary, contextualized intention consensus statement for the current censorship task.

[0059] For each text segment, the unit generates its semantic representation (i.e., vector embedding) using a pre-trained context-sensitive language model (e.g., a BERT model or a variant thereof fine-tuned for the legal or technical field). These vectors are able to capture the meaning of words in a specific context.

[0060] Based on these semantic representations, the unit applies an intent classifier and a named entity recognition (NER) model. The intent classifier (e.g., a multi-label classifier whose output labels include "data localization intent", "data transfer authorization intent", "data traceability intent", etc.) identifies the core intent expressed by the text fragment. The NER model then extracts key concepts (e.g., "personally identifiable information", "anonymous data", "cloud service provider").

[0061] The unit compares descriptions of the same intent or concept from text fragments from different information sources. For example, if multiple white papers mention "data localization" but with different emphases (one emphasizing storage location, the other emphasizing processing jurisdiction), the unit calculates the semantic similarity of these descriptions (e.g., using cosine similarity).

[0062] If the semantic similarity of descriptions of a certain intent from multiple sources exceeds a preset threshold (e.g., 0.7), and these descriptions are highly relevant in the context of the current clause, the unit generates a "temporary intent consensus statement." This statement exists in the form of a set of structured logical predicates, such as: Regarding the application of the "data sovereignty principle" under the "seamless data exchange" clause: 1) IF (data type IS sensitive data) THEN (data storage location IS local); 2) IF (data transfer IS crosses regions) THEN (explicit authorization required by IS); 3) IF (Data processing IS external service) THEN (Requires IS clear responsibility).

[0063] These statements represent the core understanding of this non-standardized principle in the current censorship context, accompanied by a confidence score based on source weight and semantic similarity.

[0064] Furthermore, in this embodiment, the path generator no longer relies solely on preset static rules, but directly utilizes the "temporary intent consensus statement" generated by the context semantic extraction unit to construct specific logically related paths.

[0065] The path generator transforms each temporary intent consensus statement (logical predicate) into one or more executable censorship rules. For example, for statement 1 "IF (data type IS sensitive data) THEN (data storage location IS local)", the path generator constructs a censorship path that includes the following steps: 1) Identify the data types involved in the clauses to be reviewed.

[0066] 2) Determine whether the data type belongs to the category of "sensitive data" (by querying the concept knowledge base or further semantic analysis).

[0067] 3) If it is sensitive data, check the terms and conditions regarding the location of data storage to see if it explicitly requires "local storage".

[0068] When constructing paths, the path generator considers the confidence score of intent consensus statements. Statements with higher confidence scores generate more distinctive and explicit logical paths. If multiple complementary or slightly conflicting intent consensus statements exist, the path generator constructs an independent logical path for each statement to ensure multi-dimensional evaluation. For example, if there are two intent consensuses regarding "data sovereignty"—one focusing on "local storage" and the other on "authorized transfer"—two different censorship paths will be generated.

[0069] The constructed logical connection path (including specific review steps and judgment conditions) will be passed to the parallel inference engine, which will initiate an independent logical inference process to assess the compliance of the terms.

[0070] S200: Load the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and make a reasonableness judgment to obtain the reasonableness assessment results of all logically related paths of the content to be reviewed. S210. Load a set of structured rules and related knowledge points corresponding to several logically related paths of the target content to be reviewed. The set of structured rules consists of logical statements of condition-association-conclusion, and the related knowledge points are derived from a preset concept knowledge base. S220. By recursively querying and graph traversing, obtain all logical association paths of the target words and their non-linear dependencies on the structured rule set and associated knowledge points. S230. Based on the non-linear dependency relationship, set the association weight value, and judge the reasonableness of all logical association paths of the target word according to the association weight value, so as to obtain the reasonableness assessment result of all logical association paths of the content to be reviewed.

[0071] In this embodiment, for each logically related path identified above, the system initiates an independent logical inference process through a "parallel inference engine." This engine can utilize a multi-threaded or distributed computing architecture to ensure that multiple paths can be processed simultaneously, thereby improving review efficiency. Each inference process evaluates the compliance of the clause based on its specific set of rules and knowledge points, and generates a preliminary judgment result. These rule sets are stored in a structured form, for example, they can be represented as logical statements such as "If clause A satisfies condition X, and condition X is related to regulation Y, then clause A is compliant under regulation Y."

[0072] In each reasoning process, a "chain reasoning module" delves into the multi-step logical deduction level, tracing the non-linear dependencies between clauses and external standards. For example, when assessing the compliance of an "environmental impact mitigation plan," if a rule along a path requires checking the environmental requirements of "material procurement," this module will further query the "concept knowledge base" to find the connection between "material procurement" and the "waste management agreement," and then trace back from the "waste management agreement" to the "regional carbon emission cap" regulations. This recursive query and graph traversal approach effectively handles highly embedded and non-linear logical dependencies between clauses, ensuring a comprehensive evaluation even in large, integrated, and cross-domain projects involving multi-step logical deduction.

[0073] S300: The reasonableness assessment results of all logically related paths of the content to be reviewed are centrally compared and consensus verified. Based on the verification results, the clauses to be reviewed in the target tender documents are assisted in the review, and the intelligent review results of the target tender documents are obtained.

[0074] S310. The reasonableness assessment results of all logically related paths of the content to be reviewed are centrally compared and consensus verified. The reasonableness assessment results include reasonableness conclusions and confidence scores. S320. If the reasonableness assessment results of all logical connection paths of the content to be reviewed have a consistent reasonableness conclusion and the confidence score is higher than the preset score threshold, then all logical connection paths of the target word are merged and output to obtain a consistency conclusion report. In this embodiment, the system aggregates the preliminary judgment results generated by all parallel inference paths into a "result comparison unit" for centralized comparison and consensus verification. Each preliminary judgment result not only includes a compliance conclusion (e.g., "compliant" or "non-compliant"), but also a confidence score (e.g., between 0 and 1).

[0075] If all parallel paths arrive at a consistent compliance conclusion and their confidence scores are all above a preset threshold (e.g., 0.8), the system considers the judgment result of the clause to be clear and reliable.

[0076] S330. If the reasonableness assessment results of all logical connection paths of the content to be reviewed are inconsistent or the confidence score is lower than the preset score threshold, then all logical connection paths of the target word will be output independently to obtain a multi-path difference report. In this embodiment, if the judgment results obtained from different paths are inconsistent (for example, one path judges "compliant" and another path judges "non-compliant"), or if the confidence score of the compliance judgment of a certain key clause is lower than the threshold, the system will immediately mark it as a "potential conflict".

[0077] At this point, a "difference report generator" will activate, producing a detailed "multi-path difference report." This report not only clearly identifies the inconsistent conclusions but also details the various reasoning paths that led to these differences, the different knowledge points cited (e.g., which specific law, industry standard, or internal regulation), and the logical points of divergence between them (e.g., path A considers "data encryption strength" to be compliant, while path B considers "data transmission protocol" to be non-compliant). This report significantly reduces the workload of source tracing analysis for human review teams when dealing with complexity, which can lead to decreased accuracy and consistency in review results, thereby improving the overall reliability of the review process.

[0078] S340. Based on the consistency conclusion report or multi-path difference report, conduct auxiliary review of the clauses to be reviewed in the target tender document to obtain the intelligent review result of the target tender document.

[0079] In summary, when reviewing each clause in a tender document, this invention no longer relies on the opinion of just one "expert," but instead consults multiple "experts" from different professional backgrounds simultaneously. When a clause in a large, comprehensive, cross-disciplinary project tender document needs review, the system first carefully "reads" the clause and attempts to understand its deeper meaning, even if it contains some uncommon words or unexplained connections. Based on the context and initial understanding of the clause, it proactively envisions multiple possible review angles or bases, just as different experts interpret clauses from different perspectives such as law, technology, and management.

[0080] Next, the system will simultaneously initiate multiple independent review processes, each conducting in-depth analysis along a pre-defined review perspective (i.e., a logical path). These processes will trace the complex, non-linear relationships between the clauses and various external standards, and may even require multi-step logical deductions. For example, a clause concerning "data security" may be reviewed simultaneously from three perspectives: national law, industry standards, and international best practices. Each review process will provide a preliminary judgment based on the rules and knowledge it is based on.

[0081] Finally, the system aggregates and compares all these independent judgments. If all the "experts" agree, the compliance of the clause is clear. However, if different "experts" give inconsistent opinions, or if an "expert" lacks confidence in their judgment, the system will immediately issue a warning and generate a detailed report. This report not only points out where the disagreements exist but also clearly outlines the basis and reasoning process for each "expert's" conclusion, helping human reviewers quickly pinpoint the problem and understand the root cause of the disagreement.

[0082] This approach differs from conventional intelligent review schemes. Conventional schemes typically follow a pre-defined, single logical path to determine the compliance of clauses. This single-path approach is highly efficient when handling tender documents with clear structures, well-defined rules, and a single domain. For example, reviewing a standardized software procurement contract, where the clauses and corresponding laws and regulations are usually direct and explicit, a single path is sufficient to quickly provide accurate results. However, for the large, comprehensive, and cross-domain project tender documents targeted by this scheme, they often integrate multiple professional technical specifications, contain numerous implicit semantic relationships, frequently introduce emerging and non-standardized professional terms that have not yet reached a unified consensus, and have highly embedded and non-linear logical dependencies between clauses. In this complex environment, a single-path review scheme is prone to overlooking key information, misunderstanding the intent of clauses, or being unable to handle multi-step logical inferences due to its limited perspective, leading to a decrease in the accuracy and consistency of the review results. This scheme, by proactively exploring multiple paths and verifying consensus, can better cope with this complexity and uncertainty, ensuring the comprehensiveness and reliability of the review results.

[0083] In the review of tender documents for large, comprehensive, and cross-disciplinary projects, traditional methods face problems such as fragmented information, complex interrelationships between clauses, inconsistent review standards, and an excessive burden on human reviewers. Especially when dealing with tender documents that integrate multiple professional technical specifications, contain numerous implicit semantic relationships, frequently introduce emerging non-standardized terms, and have highly embedded non-linear logical dependencies between clauses, conventional intelligent review systems often struggle to provide comprehensive and consistent judgments. These systems typically rely on a single, pre-defined reasoning path, aiming to provide a clear "yes" or "no" answer. When faced with complexity, ambiguity, or potentially conflicting interpretations, they either fail to provide effective judgments or may offer a one-sided or misleading single conclusion, requiring human reviewers to still invest significant effort in verification and correction.

[0084] In the complex scenarios described above, a single, deterministic judgment is often insufficient to reflect the full picture of the problem. This embodiment no longer pursues a single, "perfect" answer, but actively embraces complexity, transforming the system's role from a "single answer provider" to a "logical conflict discoverer" and a "consensus-forming facilitator." By exploring multiple possible logical reasoning paths in parallel, the system can evaluate the same clause from different dimensions and according to different rule sets. When these paths lead to inconsistent conclusions, the system does not simply choose one, but transparently reveals these differences and elaborates on the logical basis and knowledge points leading to the disagreement.

[0085] This implementation moves and partially automates the "dispute resolution" and "standardization" tasks traditionally performed by human experts in the later stages of review. Instead of waiting for the system to completely fail or for problems to be discovered manually, it proactively identifies and exposes potential logical conflicts during the reasoning process. This is a non-linear strategy that integrates "uncertainty management" into the core reasoning process. By making these uncertainties transparent, it actually enhances the reliability and credibility of the final review results. It enables human experts to more efficiently locate problems, understand the root causes of disagreements, and ultimately make more unified and accurate decisions, effectively solving the problem of decreased accuracy and consistency in review results when dealing with complex, multi-step logical inferences.

[0086] By proactively exploring multiple logical reasoning paths and performing consensus verification, the accuracy and consistency of intelligent review of tender documents can be significantly improved. It can effectively identify and make transparent potential logical conflicts and ambiguities in complex clauses, thereby greatly reducing the review burden on human reviewers when handling large, comprehensive, and cross-domain tender documents due to information dispersion, implicit connections, and non-linear dependencies, ultimately improving the reliability of review reports and decision-making efficiency.

[0087] Reference Figure 2 A knowledge graph-based intelligent review system for tender documents includes: The first module 201 is used to determine the objects to be reviewed based on the target tender documents and generate several logical association paths for the target content to be reviewed. The second module 202 is used to load the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and to make a reasonableness judgment to obtain the reasonableness evaluation results of all logically related paths of the content to be reviewed. The third module 203 is used to centrally compare and verify the rationality assessment results of all logically related paths of the content to be reviewed, and to conduct auxiliary review of the clauses to be reviewed in the target tender document based on the verification results, so as to obtain the intelligent review result of the target tender document.

[0088] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0089] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this is not intended to limit the scope of the embodiments of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A knowledge graph-based intelligent review method for tender documents, characterized in that, Includes the following steps: Based on the target tender documents, identify the objects to be reviewed and generate several logical connection paths for the target content to be reviewed; Load the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and make a reasonableness judgment to obtain the reasonableness assessment results of all logically related paths of the content to be reviewed. The reasonableness assessment results of all logically related paths of the content to be reviewed are centrally compared and verified for consensus. Based on the verification results, the clauses to be reviewed in the target tender documents are assisted in the review, and the intelligent review results of the target tender documents are obtained.

2. The intelligent review method for tender documents based on knowledge graphs according to claim 1, characterized in that, The step of identifying the objects to be reviewed based on the target tender documents and generating several logically related paths for the target content to be reviewed specifically includes: Based on the target tender documents, the objects to be reviewed were identified; If the object to be reviewed is the entire target tender document, then the project context features of the entire target tender document are extracted and a conflict solution map is constructed. Based on the nonlinear mapping relationship between the project context features and the conflict solution map, several logical association paths of the target content to be reviewed are generated. If the object to be reviewed is a clause in the target tender document, the information source of the clause to be reviewed is first determined, and then semantic analysis is performed on the clause to be reviewed to generate several logical connection paths of the target content to be reviewed.

3. The intelligent review method for tender documents based on knowledge graphs according to claim 2, characterized in that, If the object to be reviewed is the entire target tender document, then the step of extracting project context features from the entire target tender document and constructing a conflict solution map, and generating several logical association paths for the target content to be reviewed based on the nonlinear mapping relationship between the project context features and the conflict solution map, specifically includes: If the object to be reviewed is the entire target tender document, then the target tender document is parsed and entity recognized by natural language processing methods to obtain the contextual features of the target tender document, which include project name, technical keywords and time period; The contextual features of the target tender documents are vectorized and verified against preset geographic information system data to obtain the verified contextual feature vector. Based on the preset potential regulatory conflict types, vectorization processing is performed to generate context modifiers. The preset potential regulatory conflict types include the type of the issuing entity of the target tender document, the main core concepts of the target tender document, and the legal basis of the target tender document. Context simulation is performed on context modifiers, all context simulation conflict patterns are obtained and matched with the verified context feature vectors to construct a conflict solution map; The nonlinear mapping relationship between the conflict solution map and the preset standard legal provisions is obtained, and several logical association paths of the target content to be reviewed are generated. The nonlinear mapping relationship represents the explanatory basis for finding the corresponding conflict solution map based on the preset standard legal provisions.

4. The intelligent review method for tender documents based on knowledge graphs according to claim 3, characterized in that, If the object to be reviewed is a clause in the target tender document, the step of first determining the information source of the clause to be reviewed, and then performing semantic analysis on the clause to be reviewed to generate several logical connection paths of the target content to be reviewed specifically includes: If the object to be reviewed is a clause to be reviewed in the target tender document, it is compared with a preset list of informal information sources, which indicates that the clause to be reviewed does not originate from preset standard legal clauses. If the clauses to be reviewed in the target tender document are not in the list of informal information sources, then semantic analysis is performed on the clauses to be reviewed in the target tender document, and domain knowledge is classified according to the results of the semantic analysis to generate several logical connection paths of the target content to be reviewed. If the clauses to be reviewed in the target tender document belong to an informal information source list, then the contextual semantics of each word in the clauses to be reviewed are obtained and contextual semantics are extracted and intent consensus is generated to obtain several logical connection paths of the target content to be reviewed.

5. The intelligent review method for tender documents based on knowledge graphs according to claim 4, characterized in that, The step of performing semantic analysis on the clauses to be reviewed in the target tender document if they do not belong to the informal information source list, and then classifying them according to domain knowledge based on the semantic analysis results to generate several logical connection paths for the target content to be reviewed, specifically includes: If the clauses to be reviewed in the target tender document are not in the list of informal information sources, perform semantic analysis on the clauses to be reviewed in the target tender document, obtain the relationship between the words in the clauses to be reviewed, and perform transformation processing to generate context vectors of several words. The similarity calculation is performed between the context vectors of several words and the known concept knowledge vectors in the preset concept knowledge base to obtain the similarity calculation results. The preset concept knowledge base has several professional term knowledge vectors. Based on the similarity calculation results, domain knowledge classification is performed on each word in the clause to be reviewed to determine the associated domain of each word in the clause to be reviewed; Obtain the chapter titles in the target tender document where each term in the clause to be reviewed is located, and generate several logical connection paths for the target content to be reviewed by combining the related fields corresponding to each term in the clause to be reviewed.

6. The intelligent review method for tender documents based on knowledge graphs according to claim 4, characterized in that, If the clauses to be reviewed in the target tender document belong to an informal information source list, then the step of obtaining the contextual semantics of each word in the clauses to be reviewed and performing contextual semantic extraction and intent consensus generation to obtain several logical connection paths of the target content to be reviewed specifically includes: If the clauses to be reviewed in the target tender document belong to an informal information source list, the keywords in the clauses to be reviewed are extracted by combining the context of the words in the clauses to be reviewed, and the key text fragments of the clauses to be reviewed are obtained. Semantic segmentation is performed on the key text fragments of the clauses to be reviewed to obtain the segmented key text fragments of the clauses to be reviewed. Semantic extraction and intent consensus generation are performed on the key text fragments of the segmented clauses to be reviewed. The semantic intent vectors of the key texts are obtained and similarity is calculated to obtain the semantic similarity score of the key texts. The semantic similarity scores of key texts are converted into the intent confidence scores of the clauses to be reviewed. Combined with the related domains corresponding to each word in the clauses to be reviewed, several logical connection paths of the target content to be reviewed are generated.

7. The intelligent review method for tender documents based on knowledge graphs according to claim 6, characterized in that, The step of semantic extraction and intent consensus generation of the segmented key text fragments of the clauses to be reviewed, obtaining the semantic intent vector of the key text and calculating its similarity to obtain the semantic similarity score of the key text, specifically includes: Deep semantic analysis is performed on the segmented key text fragments of the clauses to be reviewed using a pre-trained context-sensitive language model to obtain the semantic representation of the key text fragments; The semantic representation of key text segments is subjected to intent recognition and concept extraction by a pre-trained intent classifier and a pre-trained named entity recognition model, respectively, to obtain the semantic intent vector and semantic concept vector of the key text. The pre-trained named entity recognition model represents a pre-trained natural language processing model. The semantic similarity score of the key text is obtained by calculating the similarity between the semantic intent vector and the semantic concept vector of the key text.

8. The intelligent review method for tender documents based on knowledge graphs according to claim 7, characterized in that, The step of loading the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and performing a reasonableness judgment to obtain the reasonableness assessment results of all logically related paths of the content to be reviewed, specifically includes: Load a set of structured rules and related knowledge points corresponding to several logically related paths of the target content to be reviewed. The set of structured rules consists of logical statements of condition-association-conclusion, and the related knowledge points are derived from a preset concept knowledge base. By using recursive queries and graph traversal, we obtain all logical association paths of the target words and their non-linear dependencies on the structured rule set and associated knowledge points. Based on non-linear dependencies, association weight values ​​are set, and the reasonableness of all logical association paths of the target words is judged based on the association weight values, so as to obtain the reasonableness assessment results of all logical association paths of the content to be reviewed.

9. The intelligent review method for tender documents based on knowledge graphs according to claim 8, characterized in that, The step of centrally comparing and verifying the reasonableness assessment results of all logically related paths of the content to be reviewed, and then conducting auxiliary review of the clauses to be reviewed in the target tender document based on the verification results to obtain the intelligent review result of the target tender document, specifically includes: The reasonableness assessment results of all logically related paths of the content to be reviewed are centrally compared and consensus verified. The reasonableness assessment results include reasonable conclusions and confidence scores. If the reasonableness assessment results of all logical connection paths of the content to be reviewed have a consistent reasonableness conclusion and the confidence score is higher than the preset score threshold, then all logical connection paths of the target word will be merged and output to obtain a consistency conclusion report. If the reasonableness assessment results of all logical connection paths of the content to be reviewed are inconsistent or the confidence score is lower than the preset score threshold, then all logical connection paths of the target word will be output independently to obtain a multi-path difference report. Based on the consistency conclusion report or the multi-path difference report, the clauses to be reviewed in the target tender document are reviewed to obtain the intelligent review result of the target tender document.

10. A knowledge graph-based intelligent review system for tender documents, characterized in that, Includes the following modules: The first module is used to determine the objects to be reviewed based on the target tender documents and generate several logical association paths for the target content to be reviewed. The second module is used to load the structured rule set and related knowledge points corresponding to several logically related paths of the target content to be reviewed, and to make a reasonableness judgment to obtain the reasonableness evaluation results of all logically related paths of the content to be reviewed. The third module is used to centrally compare and verify the rationality assessment results of all logically related paths of the content to be reviewed, and to conduct auxiliary review of the clauses to be reviewed in the target tender document based on the verification results, so as to obtain the intelligent review result of the target tender document.

Images