Bipolar argument theory-based bid invitation procurement compliance and risk decision-making method

Abstract

The invention discloses a bid invitation purchasing compliance and risk decision-making method based on a bipolar discussion theory. The method comprises the following steps: firstly, carrying out structuring and atomization processing on bid invitation documents and laws and regulations through a DeepOCR and a big language model; fusing the internal atomic arguments and external evidences, and constructing a bidirectional argument graph containing a support relationship and an attack relationship; and iteratively calculating the standing intensity score of each argument through an aggregation influence model to realize quantitative risk assessment. And finally generating a structured auditing opinion with a traceable evidence chain, and automatically distributing the structured auditing opinion to an expert review or automatic approval channel according to the risk level. Through a man-machine cooperation feedback mechanism, expert correction opinions are settled as new knowledge, and continuous self-evolution of the bidding knowledge ontology network is realized. According to the method, the accuracy, interpretability and adaptability of compliance judgment are remarkably improved, and the method is suitable for government purchase, project bid invitation, enterprise internal control and other scenes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically to a bidding and procurement compliance and risk decision-making method based on bipolar argumentation theory. Background Technology

[0002] As a crucial method for enterprise resource allocation, bidding and procurement must ensure legal procedures and fair outcomes. However, in practice, the quality of bidding documents varies greatly, often leading to bid rigging, complaints, and even legal disputes due to issues such as vague clauses, inappropriate qualification requirements, and unreasonable scoring standards. Traditional compliance reviews mainly rely on manual experts to verify regulations line by line, which suffers from bottlenecks such as low efficiency, strong subjectivity, and limited knowledge coverage.

[0003] In recent years, some institutions have attempted to introduce Natural Language Processing (NLP) and knowledge graph technologies to assist in review. For example, they use rule engines to match keywords to check legal citations, or utilize pre-trained language models to identify high-risk expressions. However, these methods generally have the following limitations: (1) Lack of deep reasoning ability: It can only perform surface semantic matching and cannot model the supporting / conflicting logical relationships between arguments; (2) Separation of internal and external knowledge: The internal bidding terms and external laws, cases and expert experience have not formed a unified reasoning framework; (3) The results are unexplainable: the large model outputs a "black box" judgment, which makes it difficult to provide a traceable chain of evidence for expert review; (4) Poor dynamic adaptability: It cannot automatically update the knowledge system based on newly released regulations or typical cases.

[0004] Furthermore, existing systems often employ a one-way "applicability assessment," neglecting the two-way argumentative structure prevalent in bidding and tendering scenarios—that is, the same clause may be supported or refuted by multiple authoritative sources simultaneously. Quantifying this complex impact and generating audit opinions with definitive assessments has become a key technical challenge in the field of intelligent compliance. Summary of the Invention

[0005] In view of the above problems, this invention is proposed to provide a bidding and procurement compliance and risk decision-making method based on the theory of bipolar argumentation to overcome or at least partially solve the above problems. To achieve the above objectives, the present invention adopts the following technical solution: A bidding and procurement compliance and risk decision-making method based on bipolar argumentation theory includes the following steps: S1: Obtain electronic documents of relevant laws, regulations, and historical bidding cases related to bidding and procurement; S2: Perform structured preprocessing on the electronic document to generate structured text with clear hierarchy; S3: Input structured text into a large language model, break it down into atomic logical units, and convert it into structured JSON data; S4: Based on the structured JSON data, construct a bidding knowledge ontology network in the graph database; S5: Receive electronic documents of bidding requirements submitted by the procurement requesting party; S6: Use a large language model to extract the set of atomic arguments from the electronic document of the tender requirements; S7: Analyze the support and attack relationships within the set of atomic arguments to generate an internal evidence set; S8: Based on each atomic argument in the atomic argument set, retrieve relevant external evidence in the bidding knowledge ontology network and generate an external evidence set; S9: Merge the internal evidence set and the external evidence set to construct a bipolar weighted argument diagram that includes internal and external connections; S10: Run the aggregation influence model on the bipolar weighted argument graph to calculate the position strength of each argument and obtain the final position strength mapping table; S11: In the bipolar weighted argument diagram, combined with the final position strength mapping table, the risk level is assessed based on the position strength of each argument, an influence tracing algorithm is constructed, and the core causal evidence chain is traced. S12: Based on the aforementioned causal evidence chain, generate a structured audit opinion text.

[0006] Preferably, based on the structured JSON data, a bidding knowledge ontology network is constructed in a graph database, including: S401: For each historical bidding case node, the key-value pairs in the structured JSON data are used as attributes of the historical bidding case node, and the original text of the disputed focus is converted into a semantic embedding vector through a large language model and saved as a node vector attribute; S402: Traverse all historical bidding case nodes, find the corresponding legal and regulatory provisions nodes based on the legal basis attributes of each historical bidding case node, and create an interpretation relationship between the two nodes; S403: Iterate through the risk label array of all historical bidding case nodes. For each label in the array, create a risk relationship between the historical bidding case node and the preset risk label node in the graph, so that all relevant cases can be quickly retrieved by risk type in the future. S404: Repeat the above processing flow until all historical bidding case nodes have completed the graph mapping; S405: For disputed clauses that have been reviewed and ultimately adopted by past experts, create an expert experience knowledge node and embed the original text of the expert opinion and the text of the core opinion into a vector, along with the expert ID, project ID, and timestamp as attributes of the expert experience knowledge node; S406: For each expert experience knowledge node, traverse every legal and regulatory provision node in the graph, calculate the cosine similarity between the text embedding vector of the expert experience knowledge node and the text vector of the legal and regulatory provision node, and when the similarity is greater than a preset threshold, create a support relationship between the expert experience knowledge node and the legal and regulatory provision node. S407: For each expert experience knowledge node, traverse every historical bidding case node in the graph, and use LLM analysis to determine whether there is an applicable relationship, a contradictory relationship, or no relation between the expert experience knowledge node and the historical bidding case node. If it is an applicable relationship, create an applicable relationship between the two nodes; if it is a contradictory relationship, create a contradictory relationship between the two nodes. S408: Traverse all legal and regulatory clause nodes in the graph database and establish conflict relationships between two legal and regulatory clause nodes.

[0007] Preferably, the bidding knowledge ontology network constructed in S4 includes the following node types and weighted relationships between nodes: Node types include: legal and regulatory provisions nodes, historical bidding case nodes, and expert experience and knowledge nodes; The weighted relationships between nodes include: (1) Interpretive Relationship: Pointing from the historical bidding case node to the legal provisions node indicates that the case constitutes a specific interpretation of the legal provisions. The weight of the interpretation relationship is set to 0.9. (2) Support relationship: The relationship between the expert experience knowledge node and the legal provisions node indicates that the expert experience supports the applicability of the legal provisions. The weight of the support relationship is set to 0.7. (3) Conflict relationship: Established between two legal provisions, indicating that there is a direct legal conflict between them, and the weight of the conflict relationship is set to 1.0; (4) Applicability Relationship: The expert experience knowledge node points to the historical bidding case node, indicating that the expert experience is applicable to the case context. The weight of the applicability relationship is set to 0.8. (5) Refutation relationship: The historical bidding case node points to the expert experience knowledge node, indicating that the case refutes the expert experience. The weight of the refutation relationship is set to 0.85.

[0008] Preferably, the specific process of establishing a conflict relationship between two legal and regulatory provisions includes the following steps: S4081: Traverse all legal and regulatory clause nodes in the graph database, generate text vectors based on the rule-based descriptions corresponding to each legal and regulatory clause node, calculate the semantic similarity between any two legal and regulatory clause nodes, and select all semantically similar or topic-related node pairs as candidate processing objects. S4082: For the currently processed node pair, determine whether their respective spatial jurisdictions overlap; if the two jurisdictions do not overlap, it is determined that there is no conflict and the node pair is skipped, and the next node pair is processed; if there is an overlap, proceed to S4083. S4083: Compare the validity levels of the node pair to determine if they are at the same validity level: If the levels of effectiveness are different, the law with the higher level of effectiveness is determined as the superior law and the law with the lower level of effectiveness is determined as the inferior law according to the preset rules of effectiveness priority. When the superior law and the inferior law are semantically similar but have potential conflicts, a conflict relationship from the inferior law to the superior law is created; otherwise, a conflict relationship from the superior law to the inferior law is created. When the validity levels are the same, compare the effective dates of the node pairs: If the effective dates are different, the legal provisions with the later effective date are regarded as new laws, and the legal provisions with the earlier effective date are regarded as old laws. A relationship edge is created in the graph database from the old law node to the new law node, indicating that the old law has been replaced by the new law. If the effective dates are the same, no relationship will be created; S4084: After processing one node pair and creating the corresponding relationship, continue processing the next node pair until all candidate node pairs have been processed.

[0009] Preferably, S9 specifically includes: S901: Initialize an empty graph data structure to store nodes and edges; S902: Traverse the set of atomic arguments, add each atomic argument as an independent internal argument node to the node set of the graph data structure, and carry all its attribute information; S903: Traverse the external evidence set, add each external evidence as an independent external evidence node to the node set of the graph data structure, and carry all its attributes in the bidding knowledge ontology network. S904: Traverse the internal evidence set, and for each internal relation, create a directed edge between the corresponding internal argument nodes in the graph data structure. The edge type is either support or attack, and set its weight value. S905: Traverse the merged internal and external evidence sets, and for each internal-external intersection relationship, create a directed edge between the corresponding internal argument node and external evidence node in the graph data structure; S906: Traverse all pairs of external evidence nodes in the external evidence set, and determine whether there is a weighted relationship between the nodes by querying the graph database. If there is, add the relationship as an edge to the graph data structure, keeping the original relationship type and weight unchanged. S907: Once all nodes and relationships have been successfully added, return the complete bipolar weighted argument graph.

[0010] Preferably, S10 includes: S1001: Input the bipolar weighted argument graph, the basic strength mapping table, the influence factor α, the maximum number of iterations, and the convergence threshold; S1002: Initialize the position strength mapping table, and set the initial position strength of each node to its corresponding base strength; S1003: For each node in the bipolar weighted argument graph, traverse all its supporters and calculate the total support influence. Iterate through all its attackers and calculate the total opposing influence; S1004: Calculate the node's position strength in the next round based on its base strength, total support influence, and total opposition influence; S1005: The calculated field strength is normalized using the Tanh activation function, compressing it to the interval [-1, 1]. S1006: Store the normalized position strength into the position strength mapping table as the position strength of the node after this round of update; S1007: Compare the absolute difference in the strength of the position before and after the update. If the change of any node is greater than the preset convergence threshold, then set the convergence flag to false. After one round of iteration is completed, check whether the convergence flag is still true. If it is true, it is determined that the entire bipolar weighted argument graph has reached a stable state, the iteration is terminated early, and the final position strength mapping table is output. If it is false, repeat S1003 to S1007 until the convergence condition is met or the maximum number of iterations is reached, and the final position strength mapping table is output.

[0011] Preferably, the formula for calculating the position strength of a node in the next round is:

[0012] in, Let the position strength of node e be in the next round. The strength of the position in the current round, BS(e) Based on the basic strength, α is the influence factor. and These are the weights for supportive and attacking relationships, respectively. and These represent the strength of the positions of supporters and attackers in the current round, respectively.

[0013] Preferably, the tracing of the core causal evidence chain includes: S1101: Input the set of atomic arguments to be explained, the bipolar weighted argument graph, the final position strength mapping table, and the maximum backtracking depth; use the atomic arguments as the current target risk nodes, initialize the attack chain and support chain, which are used to store the key nodes in the attack path and support path, respectively, and execute S1102 and S1104 respectively. S1102: Among all the input edges of the current target risk node, filter out the relationship edges of the type of attack, calculate the actual attack influence of the source node of each edge, select the node with the largest actual attack influence as the strongest attacker in this layer, and add it to the attack chain. S1103: Take the strongest attacker as the new current target risk node, repeat S1102, and continue to trace back to its upstream attacker until there are no effective attackers or the maximum backtracking depth is reached, thus completing the construction of the attack chain. S1104: Among all the input edges of the current target risk node, filter out the relationship edges of type support, calculate the actual support influence of the source node of each edge, select the node with the largest actual support influence as the strongest supporter of this layer, and add it to the support chain. S1105: Take the strongest supporter as the new current target risk node, repeat S1104, and continue to trace back to its upstream supporters until there are no effective supporters or the maximum backtracking depth is reached, thus completing the construction of the support chain. S1106: Once the attack chain and support chain have been traced, the target risk nodes that need to be explained, the final position strength score, support chain, and attack chain are encapsulated into a structured argument summary.

[0014] Preferably, it further includes: Create two empty task lists: a high-priority task list and a low-priority task list, to store review tasks that require in-depth expert involvement and review tasks that can be approved quickly, respectively. Iterate through each audit opinion in the structured audit opinion text, and based on the position strength score corresponding to the audit opinion, determine that the audit opinion is highly controversial for medium-risk audit opinions and add it to the high-priority task list; otherwise, determine that the audit opinion is highly credible and add it to the low-priority task list. The audit opinions in the high-priority task list are sent to the expert in-depth review workbench for manual review and intervention; Audit comments from low-priority task lists can be directly pushed to the automated approval channel for rapid processing.

[0015] Preferably, it further includes: The adopted expert opinions are stored in the bidding knowledge ontology network in the form of expert experience knowledge nodes, so as to realize the self-evolution and updating of the knowledge graph.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a bidding and procurement compliance and risk decision-making method based on the bipolar argument theory, which has the following effects: (1) Break through surface semantic matching and realize deep compliance reasoning based on logical argumentation. To address the shortcomings of existing methods that rely solely on keyword matching or shallow semantic recognition and lack logical reasoning capabilities, this invention introduces bipolar argumentation theory for the first time. It unifies bidding terms, laws and regulations, historical cases, and expert experience into computable "atomic arguments," and dynamically constructs a bidirectional argumentation network containing supporting and refuting relationships through a large language model. The system is no longer limited to "matching" but simulates the legal argumentation process, making quantitative decisions in the clash of opposing evidence, significantly improving the logical rigor and scenario adaptability of compliance judgments.

[0017] (2) Break down internal and external knowledge barriers and build a unified multi-source collaborative reasoning framework. This invention constructs a Knowledge Ontology Network (TKON) for Bidding and Tendering, organically integrating external authoritative knowledge (legal and regulatory provisions, historical bidding and tendering cases, and expert experience) with internal bidding clauses under review into a single graph structure, and achieving cross-domain association in the subsequent bipolar weighted argumentation graph. This mechanism completely solves the problem of the separation between "internal rules" and "external evidence" in traditional systems, ensuring that every review conclusion is based on a complete logical foundation supported or challenged by both internal and external evidence.

[0018] (3) Output interpretable and traceable audit conclusions to enhance trust in human-machine collaboration. Unlike large-scale models that output information in a "black box" manner, this invention uses the Aggregate Influence Model (AIM) to calculate the position strength score (from -1 to 1) for each argument and generates a core causal evidence chain (including the strongest support path and the strongest attack path). Experts can clearly trace the reasoning process, key evidence sources, and quantitative basis for any risk assessment, achieving "evidence-based conclusions, verifiable doubts, and precise intervention," greatly improving the transparency and credibility of AI-assisted decision-making.

[0019] (4) Support dynamic knowledge evolution and realize continuous self-optimization of system capabilities. This invention designs a closed-loop knowledge update mechanism: review opinions adopted by experts can be automatically converted into new expert experience knowledge nodes and injected into TKON; simultaneously, newly released regulations or typical cases can be incrementally updated through an automated process. This mechanism enables the system to adapt to changes in the legal environment and the evolution of business practices, effectively overcoming the fundamental defect of traditional static knowledge bases that are "outdated as soon as they are deployed".

[0020] In summary, this invention systematically addresses the key shortcomings of existing intelligent compliance technologies in terms of reasoning paradigm, knowledge architecture, interpretability, and self-evolution capability. It not only achieves an order-of-magnitude improvement in the accuracy and depth of audits, but also constructs a new generation of intelligent risk control system for bidding and procurement that is trustworthy, traceable, learnable, and usable, demonstrating outstanding technological advancement and broad industrial application value. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a flowchart of a bidding and procurement compliance and risk decision-making method based on bipolar argumentation theory provided in an embodiment of the present invention; Figure 2 This is a flowchart for establishing conflict relationships provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the construction of a bipolar weighted argument diagram containing internal and external relationships, provided in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the calculation of the position strength using the running aggregation influence model provided in this embodiment of the invention; Figure 5 This is a flowchart for constructing a core causal evidence chain provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This invention discloses a bidding and procurement compliance and risk decision-making method based on the theory of bipolar argumentation, such as... Figure 1 As shown, it includes: S1: Obtain electronic documents of relevant laws, regulations, and historical bidding cases related to bidding and procurement; S2: Perform structured preprocessing on electronic documents to generate structured text with clear hierarchy; S3: Input structured text into a large language model, break it down into atomic logical units, and convert it into structured JSON data; S4: Construct a knowledge ontology network for bidding and tendering based on structured JSON data in a graph database; S5: Receive electronic documents of bidding requirements submitted by the procurement requesting party; S6: Extract the set of atomic arguments from the electronic document of bidding requirements using a large language model; S7: Analyze the supporting and attacking relationships within the set of atomic arguments to generate an internal evidence set; S8: Based on each atomic argument in the atomic argument set, retrieve relevant external evidence in the bidding knowledge ontology network and generate an external evidence set; S9: Merge the internal evidence set and the external evidence set to construct a bipolar weighted argument diagram that includes internal and external connections; S10: Run the aggregation influence model on the bipolar weighted argument graph to calculate the position strength of each argument and obtain the final position strength mapping table; S11: In the bipolar weighted argument diagram, combined with the final position strength mapping table, the risk level is assessed based on the position strength of each argument, an influence tracing algorithm is constructed, and the core causal evidence chain is traced. S12: Generate structured audit opinion text based on the causal evidence chain.

[0025] In this embodiment, in S2, based on DeepSeek-OCR, the electronic document is preprocessed in a structured manner to generate a structured text with rich content, a clear hierarchical structure, and deep learning information. For example, for laws and regulations, hierarchical parsing is performed according to "legal level - issuing authority - effective date - clause number - clause text"; for historical bidding cases, fields such as "case number, procurement project name, disputed issues, cited legal provisions, risk labels, and judgment results" are extracted to generate a clearly structured text.

[0026] In this embodiment, in S3, an AtomicClause Splitting Agent based on a Large Language Model (LLM) is introduced to decompose complete legal provisions into the smallest semantic units—atomic clauses, which are legal provisions with a single, independent, and indivisible logical judgment meaning.

[0027] To ensure the consistency, standardization, and computability of the decomposition results, the system uses standardized prompts (PromptEngineering) to guide large language models such as Deepseek to perform atomic tasks. For legal and regulatory provisions, they are decomposed into the smallest, independent atomic clauses with a single logical judgment meaning; for historical bidding cases, the core elements, time, and scope of the cases are extracted. Example of an atomic decomposition agent: # ROLE: You are a meticulous legal text analyst.

[0028] # TASK: Please break down the following complete legal text into the smallest, independent logical units. Each unit should express a complete and indivisible legal provision.

[0029] # OUTPUT_FORMAT: Outputs as a JSON array, with each object containing 'id' (an auto-incrementing number, such as '22.1.3' representing Article 22, Section 1, Item 3) and 'text' (the original text of the atomic clause).

[0030] # CONTEXT: Original text of the law: Article 22 Suppliers participating in government procurement activities shall meet the following conditions: (i) Possess the capacity to independently assume civil liability; (ii) Possessing a good business reputation and a sound financial accounting system; (iii) Possessing the necessary equipment and professional technical capabilities to perform the contract; ..." # OUTPUT: In this embodiment, a bidding knowledge ontology network is constructed in a graph database based on structured JSON data, including: S401: Read the structured JSON data of historical cases generated in the previous step. For each JSON file, create a historical bidding case node in the Neo4j graph database. One historical bidding case node corresponds to one historical bidding case in the structured JSON data. All key-value pairs in the structured JSON data are stored as attributes of the node for the historical bidding case. The focus text of the dispute is converted into a semantic embedding vector through a large language model and stored as the node attribute caseVector for subsequent semantic similarity retrieval. S402: Traverse all historical bidding case nodes, find the corresponding legal and regulatory provisions nodes based on the legal association attributes of each historical bidding case node, create an interpretation relationship between the two nodes, and assign a weight of 0.9; S403: Iterate through the risk label array of all historical bidding case nodes. For each label in the array, create a risk relationship between the historical bidding case node and the preset risk label node in the graph, so that all relevant cases can be quickly retrieved by risk type in the future. S404: Repeat the above processing flow until all historical bidding case nodes have completed the graph mapping; S405: For disputed clauses that have been reviewed and ultimately adopted by past experts, create an expert experience knowledge node and embed the original text of the expert opinion and the text of the core opinion into a vector, along with the expert ID, project ID, and timestamp as attributes of the expert experience knowledge node; S406: For each expert experience knowledge node, traverse every legal and regulatory provision node in the graph, calculate the cosine similarity between the text embedding vector of the expert experience knowledge node and the text vector of the legal and regulatory provision node. When the similarity is greater than the preset threshold, create a support relationship between the expert experience knowledge node and the legal and regulatory provision node, and assign a weight of 0.7. S407: For each expert experience knowledge node, traverse every historical bidding case node in the graph, and use LLM analysis to determine whether there is an applicable relationship, a contradictory relationship, or no correlation between the expert experience knowledge node and the historical bidding case node. If it is an applicable relationship, create an applicable relationship between the two nodes and set the weight to 0.8. If it is a contradictory relationship, create a contradictory relationship between the two nodes and set the weight to 0.85.

[0031] S408: Traverse all legal and regulatory clause nodes in the graph database and establish conflict relationships between two legal and regulatory clause nodes.

[0032] like Figure 2 As shown, the specific process of establishing a conflict relationship between two legal and regulatory provisions includes the following steps: S4081: Traverse all legal and regulatory clause nodes in the graph database, generate text vectors based on the rule-based descriptions corresponding to each legal and regulatory clause node, calculate the semantic similarity between any two legal and regulatory clause nodes, and select all semantically similar or topic-related node pairs as candidate processing objects. S4082: For the currently processed node pair, determine whether their respective spatial jurisdictions overlap; if the two jurisdictions do not intersect, it is determined that there is no conflict and the node pair is skipped, and the next node pair is processed; if there is an overlap, proceed to S4083; for example, one is "country" and "province A", or "province A" and "city B". If the jurisdictions of the two regulations are completely unrelated, there is no conflict or substitution relationship between them, the algorithm skips the node pair and continues to process the next one.

[0033] S4083: Compare the effectiveness level of the node pair (e.g., law > administrative regulation > departmental rules > local regulations > internal system) to determine whether they are at the same effectiveness level: If the levels of effectiveness are different, the law with the higher level of effectiveness is determined as the superior law and the law with the lower level of effectiveness is determined as the inferior law according to the preset rules of effectiveness priority. When the superior law and the inferior law are semantically similar but have potential conflicts, a conflict relationship from the inferior law to the superior law is created; otherwise, a conflict relationship from the superior law to the inferior law is created. When the validity levels are the same, compare the effective dates of the node pairs: If the effective dates are different, the legal provisions with the later effective date are regarded as new laws, and the legal provisions with the earlier effective date are regarded as old laws. A relationship edge is created in the graph database from the old law node to the new law node, indicating that the old law has been replaced by the new law. If the effective dates are the same, no relationship will be created; S4084: After processing one node pair and creating the corresponding relationship, continue processing the next node pair until all candidate node pairs have been processed.

[0034] Therefore, the knowledge ontology network (TKON) for bidding and tendering constructed in S4 includes the following node types and weighted relationships between nodes. TKON includes three types of core nodes: The LawClause node has the following attributes: unique identifier id, original text of the clause text, source document sourceDoc, hierarchyLevel, effective date effectiveDate, repeal date repeatDate, clause type clauseType, and rule-based description vector ruleVector generated by the large language model. Historical bidding case nodes (Case) have the following attributes: case ID, decision, focusText (original text of the disputed point), semantic embedding vector caseVector, and other structured metadata; Expert knowledge nodes are used to store validated review rules or industry practices. Establish the following directed weighted relationships between nodes: (1) Interpretive Relationship (INTERPRETS): Pointing from the historical bidding case node to the legal provisions node, it indicates that the case constitutes a specific practical interpretation of the legal provisions, and the relationship weight is set to 0.9; (2) Support Relationships: Points from expert experience knowledge nodes to legal provisions nodes, indicating that expert experience supports the applicability of the legal provisions. The relationship weight is set to 0.7. (3) Conflict relationship (CONFLICTS_WITH): Established between two legal provisions nodes, indicating that there is a direct legal conflict between them, and the relationship weight is set to 1.0; (4) Applicability relationship (APPLIES_TO): The relationship points from the expert experience knowledge node to the historical bidding case node, indicating that the expert experience is applicable to the case context. The relationship weight is set to 0.8. (5) Refutation Relationship (REFUTES): Pointing from the historical bidding case node to the expert experience knowledge node, it indicates that the authoritative case refutes the expert experience, and the relationship weight is set to 0.85; In this embodiment, after receiving the electronic bidding requirements document submitted by the procurement requester, the document receiving and logical rule transformation are performed: the uploaded electronic bidding requirements document is parsed into structured text data. Since the bidding requirements document is in a standard template format, all mandatory and binding clauses in the document can be accurately identified and extracted.

[0035] Extraction and contextualization of tender terms: The system submits all identified terms to the deepseek big model, and extracts the issues and contexts from the terms through prompt word engineering, outputting a set containing all the atomic arguments in the tender document.

[0036] Discovering the internal relationships within the tender terms: By traversing all possible argument pairs in the set of atomic arguments and using the DeepSeek large model in conjunction with cue word engineering, natural language inference is performed to discover "attack" or "support" relationships between arguments, and an internal evidence set containing all supporting / attack relationships between internal arguments is output. Examples of cue words are as follows: # ROLE: You are a logically meticulous system architect who excels at identifying the inherent connections and contradictions in requirements documents.

[0037] # TASK: Please analyze the following two atomic arguments from the same tender document and determine whether there is a logical "support" or "attack" relationship between them.

[0038] # RELATION_TYPES_TO_CONSIDER: # - Attack Relationship: # - 'Logical Contradiction Attack': Statements A and B cannot both be true at the same time.

[0039] # - 'Irrational Attack': Under the project objective, A makes B seem unreasonable or unnecessary.

[0040] # -Support Relationship: # - 'Prerequisite Dependency Support': B is a technical or business prerequisite for achieving A.

[0041] # - 'Target Alignment Support': A and B both serve a specific higher-level target.

[0042] # CONTEXT: # - Argument A: { "id": "Internal-5.1-1", "statement": "The system must support 24 / 7 uninterrupted operation."} # - Argument B: { "id": "Internal-20.3-2", "statement": "The system allows for maintenance downtime between 2 AM and 4 AM daily."} # - Core Project Objective: "Highly Available Citizen Service Platform" # OUTPUT_FORMAT: Output in JSON format, including 'relationType' (choose from the above types, or 'irrelevant'), 'justification' (brief explanation of the reason), and 'confidence' (between 0 and 1).

[0043] # OUTPUT: External evidence retrieval for tender terms: A contextualized retrieval query is performed on each atomic argument in the atomic argument set. Using the deepseek large model in conjunction with a prompt word library, relevant external evidence is retrieved from TKON, ultimately outputting a set of internal and external relations, Relations_Cross. An example prompt word template is shown below: # ROLE: You are a seasoned compliance lawyer.

[0044] # TASK: Please determine whether the following clauses (internal arguments) from the tender documents are “supporting” or “attacking” to the given external knowledge (law / case / experience).

[0045] # RELATION_TYPES_TO_CONSIDER: # - Attack Relationships (ATTACKS): # - 'Direct conflict attack': The internal argument's requirements explicitly violate the prohibitive provisions regarding external knowledge.

[0046] # - 'Avoiding risk attacks': The internal argument is set up to be highly similar to the risk points revealed in external knowledge (such as historical risk cases).

[0047] # - 'Irrational attack': External knowledge (such as expert experience) indicates that the requirements of the internal argument are unreasonable or unnecessary in the current context.

[0048] # - Supports: # - 'Compliance Support': The internal arguments clearly comply with or respond to mandatory or recommended requirements of external knowledge.

[0049] # - 'Best Practice Support': Internal arguments adopt best practices recommended from external knowledge, such as expert experience.

[0050] # CONTEXT: # - Internal Argument: { "id": "Internal-7.2.1-3", "statement": "The CPU brand must be Intel Xeon Platinum series."} # - External Knowledge: { "id": "Regulations-2.20", "type": "LawClause", "text": "Article 20 of the Bidding Law stipulates that suppliers shall not be subjected to unreasonable conditions that discriminate against them, nor shall specific brands be designated."} # OUTPUT_FORMAT: Outputs in JSON format and must include the following fields: # - 'relationType': Select one from ['Direct Conflict Attack', 'Risk Avoidance Attack', 'Unreasonable Attack', 'Compliance Support', 'Best Practice Support', 'Irrelevant'].

[0051] # - 'justification': (string) Briefly explain the logical reasons for your judgment.

[0052] # - 'confidence': (floating-point number, 0.0-1.0) How certain are you about this judgment?

[0053] # - 'source_id': (string) The ID of the originator of the attack or support relationship (id from 'external knowledge').

[0054] # - 'target_id': (string) The ID of the target party in the attack or support relationship (id from 'internal argument').

[0055] # OUTPUT: In this embodiment, as Figure 3 As shown, S9 specifically includes: S901: Initialize an empty graph data structure to store nodes and edges; S902: Traverse the set of atomic arguments, add each atomic argument as an independent internal argument node to the node set of the graph data structure, and carry all its attribute information, including but not limited to: argument_id (argument identifier), atomic_statement (atomic statement content), original_text (original text fragment), etc. S903: Traverse the external evidence set, add each external evidence as an independent external evidence node to the node set of the graph data structure, and carry all its attributes in the bidding knowledge ontology network, including: id (unique node identifier), text (text content), sourceDoc (source document), clauseType (clause type), and node type (such as legal and regulatory clause nodes, historical bidding case nodes, or expert experience knowledge nodes), etc. S904: Traverse the internal evidence set, and for each internal relation, create a directed edge between the corresponding internal argument nodes in the graph data structure. The edge type is either support or attack, and set its weight value. S905: Traverse the merged internal and external evidence sets, and for each internal-external cross relationship, create a directed edge between the corresponding internal argument node and external evidence node in the graph data structure. The edge type is either support or attack, and set its weight value. S906: Traverse all pairs of external evidence nodes in the external evidence set, and determine whether there is a predefined weighted relationship between the two nodes (such as conflict relationship, substitution relationship, interpretation relationship, application relationship or support relationship) by querying the graph database. If it exists, add the relationship as an edge to the graph data structure, keeping the original relationship type and weight unchanged. S907: Once all nodes and relationships have been successfully added, return the complete bipolar weighted argument graph, which includes internal arguments, external evidence, and their multidimensional interactions.

[0056] In this embodiment, the "Aggregate Influence Model (AIM)" is run to calculate the stance strength: on the complete bipolar weighted argument graph, the AIM solver is run to obtain the final stance strength of each argument and output the final stance strength mapping table, which contains the final "Stance Strength" score of all internal clauses and external knowledge, which is in the range of [-1, 1].

[0057] like Figure 4 As shown, it specifically includes: S1001: Input a bipolar weighted argument graph, a basic strength mapping table (stores the "initial combat power" or "innate authority" of each node, legal provisions = 1.0, historical bidding cases = 0.8, expert experience = 0.6, and the initial value of internal arguments from the bidding documents is set to 0.5), an impact factor α, a maximum number of iterations, and a convergence threshold (a very small number, such as 0.0001). S1002: Initialize the position strength mapping table, and set the initial position strength of each node to its corresponding base strength; S1003: For each node in the bipolar weighted debate graph, iterate through all its supporters and calculate the total support influence. ; Traverse all its attackers and calculate the total opposing influence. ; S1004: Calculate the node's position strength in the next round based on its base strength, total support influence, and total opposition influence;

[0058] in, Let the position strength of node e be in the next round. The strength of the position in the current round, BS(e) Based on the basic strength, α is the influence factor. and These are the weights for supportive and attacking relationships, respectively. and These represent the strength of the positions of supporters and attackers in the current round, respectively.

[0059] S1005: The calculated field strength is normalized using the Tanh activation function, compressing it to the interval [-1, 1]. S1006: Store the normalized position strength into the position strength mapping table as the position strength of the node after this round of update; S1007: Compare the absolute difference in the strength of the position before and after the update. If the change of any node is greater than the preset convergence threshold, then set the convergence flag to false. After one round of iteration is completed, check whether the convergence flag is still true. If it is true, it is determined that the entire bipolar weighted argument graph has reached a stable state, the iteration is terminated early, and the final position strength mapping table is output. If it is false, repeat S1003 to S1007 until the convergence condition is met or the maximum number of iterations is reached, and the final position strength mapping table is output.

[0060] Based on the final position strength of each internal atomic argument calculated using the aggregated impact model, the system performs automated compliance assessment and risk classification. The specific judgment rules are as follows: High Risk: If the final stance strength < If so, the atomic argument is deemed high-risk.

[0061] This indicates that the tender terms have been strongly refuted by highly authoritative external evidence (such as mandatory legal provisions or guiding judicial precedents), and the supporting arguments cannot effectively offset the impact of the refutation. This indicates serious compliance or reasonableness deficiencies and is very likely to trigger complaints, challenges, or legal disputes.

[0062] Medium Risk: like ≤Final Stance Strength< If so, the atomic argument is judged to be of medium risk.

[0063] This indicates that the forces refuting (attacking) this clause are generally significantly stronger than those supporting it, but the sources of rebuttal may not be decisive legal provisions, or the clause itself may still have some reasonable justification. Such clauses pose potential compliance risks and require close attention during the expert review process.

[0064] Disputed items / requiring manual review: like If the final position strength is less than or equal to 0.2, then the atomic argument is considered a controversial item.

[0065] This indicates a basic balance between support and opposition, with a near-neutral stance. It suggests the clause may be in a gray area of ​​legal application, or its wording may be ambiguous or open to interpretation. Such a situation itself constitutes a risk signal, and it is strongly recommended that it be subject to manual review and discretion by experts.

[0066] Low risk / Approved: If the final position strength is greater than 0.2, then the atomic argument is considered low-risk / passable.

[0067] This indicates that the clause has significant support (such as being consistent with current regulations, verified by historical cases, or supported by expert experience) and has not been substantially refuted, meeting the compliance and reasonableness requirements of current bidding and procurement, and can enter the automated approval channel.

[0068] For each atomic argument, after completing the above judgment, the system assigns it a corresponding risk level label and attaches its final position strength score, which serves as the core basis for generating the structured audit opinion text.

[0069] In this embodiment, based on the position strength score output by the AIM solver, the key logical path of each risk item in the argument diagram is automatically traced, and the Deepseek large model is called to generate structured, interpretable, and interactive audit opinion text.

[0070] Constructing an influence tracing algorithm: This algorithm adopts a "greedy strategy", which only finds the node with the greatest influence on the current node when backtracking at each level, thereby constructing the core causal evidence chain.

[0071] like Figure 5 As shown, it specifically includes: S1101: Input the set of atomic arguments to be explained, the bipolar weighted argument graph, the final position strength mapping table, and the maximum backtracking depth; use the atomic arguments as the current target risk nodes, initialize the attack chain and support chain, which are used to store the key nodes in the attack path and support path, respectively, and execute S1102 and S1104 respectively. S1102: Among all the input edges of the current target risk node, filter out the relationship edges of the type of attack, calculate the actual attack influence of the source node of each edge, select the node with the largest actual attack influence as the strongest attacker in this layer, and add it to the attack chain. S1103: Take the strongest attacker as the new current target risk node, repeat S1102, and continue to trace back to its upstream attacker until there are no effective attackers or the maximum backtracking depth is reached, thus completing the construction of the attack chain. S1104: Among all the input edges of the current target risk node, filter out the relationship edges of type support, calculate the actual support influence of the source node of each edge, select the node with the largest actual support influence as the strongest supporter of this layer, and add it to the support chain. S1105: Take the strongest supporter as the new current target risk node, repeat S1104, and continue to trace back to its upstream supporters until there are no effective supporters or the maximum backtracking depth is reached, thus completing the construction of the support chain. S1106: Once the attack chain and support chain have been traced, the target risk nodes that need to be explained, the final position strength score, support chain, and attack chain are encapsulated into a structured argument summary.

[0072] The structured argument summary is used in conjunction with the Deepseek large-scale model and a cue word library to generate a high-quality, structured audit opinion text that can be directly presented to users. Every sentence in the audit opinion text can be traced back to its original source in the argument summary, ensuring the report's traceability and eliminating illusions. A cue word template example is shown below: # ROLE: You are a world-class expert in bidding and auditing and a master of report writing. Your writing style is: professional, rigorous, logically clear, and well-supported by evidence.

[0073] # TASK: Please generate a complete audit opinion for the risk item based on the following structured "Argument Summary". Please strictly adhere to the given reporting structure, and all your analysis must be directly derived from the facts provided in the "Argument Summary".

[0074] # CONTEXT: Argument Summary # { #"risk_item_id": "internal-3.1.2-1", #"final_stance_strength": -0.85, #"risk_level": "high risk", #"strongest_attack_chain": [ #{ #"node_id": "Regulations-2.20", #"node_text": "Article 20 of the Bidding Law stipulates that suppliers shall not be subjected to unreasonable conditions or discriminatory treatment." #"node_type": "LawClause", #"node_AS": 0.95 #}, #{ #"relation_type": "REFUTES", #"relation_weight": 1.0 #} #], #"strongest_support_chain": [] #} # REPORT_STRUCTURE: # 1. **Risk Level and Summary**: [Based on risk_level and final_stance_strength, summarized in one sentence] # 2. **Original Parameter Text**: [Quoting the original text of the clause under review] # 3. **Core Conflict Analysis (Logical Origin)**: [Describe the strongest_attack_chain in detail. It must clearly explain which evidence (referencing node_text and node_type), and with what strength (referencing relation_weight), attacked the current clause. The credibility of the attacker must also be emphasized (referencing node_AS).] # 4. **Supporting and Counter-arguments**: [Describe the strongest support chain. If empty, explicitly state "No valid supporting arguments found."] # 5. **Suggested Modifications**: [Based on risk analysis, provide 1-2 specific and actionable modification plans.] # START OF REPORT: In this embodiment, the present invention further includes S13, submitting the audit opinion text to an expert review platform for manual review; specifically including: Create two empty task lists: a high-priority task list and a low-priority task list, to store review tasks that require in-depth expert involvement and review tasks that can be approved quickly, respectively. Iterate through each audit opinion in the structured audit opinion text, and based on the position strength score corresponding to the audit opinion, for medium-risk audit opinions ( ≤Final Stance Strength< ), Dispute Review Opinions ( ≤Final Stance Strength< If the audit opinion is deemed highly controversial, it will be added to the high-priority task list; otherwise, it will be deemed highly credible and added to the low-priority task list. The audit opinions in the high-priority task list are sent to the expert in-depth review workbench for manual review and intervention; Audit comments from low-priority task lists can be directly pushed to the automated approval channel for rapid processing.

[0075] The adopted expert opinions are stored in the bidding knowledge ontology network in the form of expert experience knowledge nodes, so as to realize the self-evolution and updating of the knowledge graph.

[0076] This invention applies Abstract Argumentation Theory to the field of bidding and tendering review. It goes beyond simply "retrieving" knowledge; instead, it transforms each clause under review and all relevant regulations, cases, and experiences into equal "arguments." A dynamic AI model is then used to construct an "evidence conflict and support network" encompassing "support" and "attack" relationships. Through a Convergence Impact Model (AIM) solver and rigorous iterative calculations, a final "stance strength" score (between -1 and 1) is derived for each clause. This score is not a simple "yes" or "no," but a quantitative expression of the clause's logical stance after intense debate among all opposing evidence. This enables the handling of complex scenarios that are completely unattainable with existing technologies, such as rulings on conflicts of multiple legal provisions, legal exemptions in special circumstances, and the compliance of internal regulations with higher-level laws—all deep-seated logical problems. Furthermore, each review conclusion is accompanied by a quantified "stance strength" score and a clear and visible "argumentation path." Experts can clearly see how a review conclusion is achieved through a defined causal reasoning chain, as well as the quantitative calculation process of the conclusion, thereby greatly enhancing the trust level of human-machine collaboration.

[0077] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bidding and procurement compliance and risk decision-making method based on bipolar argumentation theory, characterized in that, Includes the following steps: S1: Obtain electronic documents of relevant laws, regulations, and historical bidding cases related to bidding and procurement; S2: Perform structured preprocessing on the electronic document to generate structured text with clear hierarchy; S3: Input structured text into a large language model, break it down into atomic logical units, and convert it into structured JSON data; S4: Based on the structured JSON data, construct a bidding knowledge ontology network in the graph database; S5: Receive electronic documents of bidding requirements submitted by the procurement requesting party; S6: Use a large language model to extract the set of atomic arguments from the electronic document of the tender requirements; S7: Analyze the support and attack relationships within the set of atomic arguments to generate an internal evidence set; S8: Based on each atomic argument in the atomic argument set, retrieve relevant external evidence in the bidding knowledge ontology network and generate an external evidence set; S9: Merge the internal evidence set and the external evidence set to construct a bipolar weighted argument diagram that includes internal and external connections; S10: Run the aggregation influence model on the bipolar weighted argument graph to calculate the position strength of each argument and obtain the final position strength mapping table; S11: In the bipolar weighted argument diagram, combined with the final position strength mapping table, the risk level is assessed based on the position strength of each argument, an influence tracing algorithm is constructed, and the core causal evidence chain is traced. S12: Based on the aforementioned causal evidence chain, generate a structured audit opinion text.

2. The method as described in claim 1, characterized in that, Based on the structured JSON data, a bidding knowledge ontology network is constructed in the graph database, including: S401: For each historical bidding case node, the key-value pairs in the structured JSON data are used as attributes of the historical bidding case node, and the original text of the disputed focus is converted into a semantic embedding vector through a large language model and saved as a node vector attribute; S402: Traverse all historical bidding case nodes, find the corresponding legal and regulatory provisions nodes based on the legal basis attributes of each historical bidding case node, and create an interpretation relationship between the two nodes; S403: Iterate through the risk label array of all historical bidding case nodes. For each label in the array, create a risk relationship between the historical bidding case node and the preset risk label node in the graph, so that all relevant cases can be quickly retrieved by risk type in the future. S404: Repeat the above processing flow until all historical bidding case nodes have completed the graph mapping; S405: For disputed clauses that have been reviewed and ultimately adopted by past experts, create an expert experience knowledge node and embed the original text of the expert opinion and the text of the core opinion into a vector, along with the expert ID, project ID, and timestamp as attributes of the expert experience knowledge node; S406: For each expert experience knowledge node, traverse every legal and regulatory provision node in the graph, calculate the cosine similarity between the text embedding vector of the expert experience knowledge node and the text vector of the legal and regulatory provision node, and when the similarity is greater than a preset threshold, create a support relationship between the expert experience knowledge node and the legal and regulatory provision node. S407: For each expert experience knowledge node, traverse every historical bidding case node in the graph, and use LLM analysis to determine whether there is an applicable relationship, a contradictory relationship, or no relation between the expert experience knowledge node and the historical bidding case node. If it is an applicable relationship, create an applicable relationship between the two nodes; if it is a contradictory relationship, create a contradictory relationship between the two nodes. S408: Traverse all legal and regulatory clause nodes in the graph database and establish conflict relationships between two legal and regulatory clause nodes.

3. The method as described in claim 2, characterized in that, The bidding knowledge ontology network constructed in S4 includes the following node types and weighted relationships between nodes: Node types include: legal and regulatory provisions nodes, historical bidding case nodes, and expert experience and knowledge nodes; The weighted relationships between nodes include: (1) Interpretive Relationship: Pointing from the historical bidding case node to the legal provisions node indicates that the case constitutes a specific interpretation of the legal provisions. The weight of the interpretation relationship is set to 0.

9. (2) Support relationship: The relationship between the expert experience knowledge node and the legal provisions node indicates that the expert experience supports the applicability of the legal provisions. The weight of the support relationship is set to 0.

7. (3) Conflict relationship: Established between two legal provisions, indicating that there is a direct legal conflict between them, and the weight of the conflict relationship is set to 1.0; (4) Applicability Relationship: The expert experience knowledge node points to the historical bidding case node, indicating that the expert experience is applicable to the case context. The weight of the applicability relationship is set to 0.

8. (5) Refutation relationship: The historical bidding case node points to the expert experience knowledge node, indicating that the case refutes the expert experience. The weight of the refutation relationship is set to 0.

85.

4. The method as described in claim 2, characterized in that, The specific process of establishing a conflict relationship between two legal or regulatory provisions includes the following steps: S4081: Traverse all legal and regulatory clause nodes in the graph database, generate text vectors based on the rule-based descriptions corresponding to each legal and regulatory clause node, calculate the semantic similarity between any two legal and regulatory clause nodes, and select all semantically similar or topic-related node pairs as candidate processing objects. S4082: For the currently processed node pair, determine whether their respective spatial jurisdictions overlap; if the two jurisdictions do not overlap, it is determined that there is no conflict and the node pair is skipped, and the next node pair is processed; if there is an overlap, proceed to S4083. S4083: Compare the validity levels of the node pair to determine if they are at the same validity level: If the levels of effectiveness are different, the law with the higher level of effectiveness is determined as the superior law and the law with the lower level of effectiveness is determined as the inferior law according to the preset rules of effectiveness priority. When the superior law and the inferior law are semantically similar but have potential conflicts, a conflict relationship from the inferior law to the superior law is created; otherwise, a conflict relationship from the superior law to the inferior law is created. When the validity levels are the same, compare whether the effective dates of the node pairs are the same: If the effective dates are different, the legal provisions with the later effective date are regarded as new laws, and the legal provisions with the earlier effective date are regarded as old laws. A relationship edge is created in the graph database from the old law node to the new law node, indicating that the old law has been replaced by the new law. If the effective dates are the same, no relationship will be created; S4084: After processing one node pair and creating the corresponding relationship, continue processing the next node pair until all candidate node pairs have been processed.

5. The method as described in claim 3, characterized in that, S9 specifically includes: S901: Initialize an empty graph data structure to store nodes and edges; S902: Traverse the set of atomic arguments, add each atomic argument as an independent internal argument node to the node set of the graph data structure, and carry all its attribute information; S903: Traverse the external evidence set, add each external evidence as an independent external evidence node to the node set of the graph data structure, and carry all its attributes in the bidding knowledge ontology network. S904: Traverse the internal evidence set, and for each internal relation, create a directed edge between the corresponding internal argument nodes in the graph data structure. The edge type is either support or attack, and set its weight value. S905: Traverse the merged internal and external evidence sets, and for each internal-external intersection relationship, create a directed edge between the corresponding internal argument node and external evidence node in the graph data structure; S906: Traverse all pairs of external evidence nodes in the external evidence set, and determine whether there is a weighted relationship between the nodes by querying the graph database. If there is, add the relationship as an edge to the graph data structure, keeping the original relationship type and weight unchanged. S907: Once all nodes and relationships have been successfully added, return the complete bipolar weighted argument graph.

6. The method as described in claim 1, characterized in that, S10 includes: S1001: Input the bipolar weighted argument graph, the basic strength mapping table, the influence factor α, the maximum number of iterations, and the convergence threshold; S1002: Initialize the position strength mapping table, and set the initial position strength of each node to its corresponding base strength; S1003: For each node in the bipolar weighted argument graph, traverse all its supporters and calculate the total support influence. Iterate through all its attackers and calculate the total opposing influence; S1004: Calculate the node's position strength in the next round based on its base strength, total support influence, and total opposition influence; S1005: The calculated field strength is normalized by applying the Tanh activation function, compressing it to the interval [-1,1]. S1006: Store the normalized position strength into the position strength mapping table as the position strength of the node after this round of update; S1007: Compare the absolute difference in the strength of the position before and after the update. If the change of any node is greater than the preset convergence threshold, then set the convergence flag to false. After one round of iteration is completed, check whether the convergence flag is still true. If it is true, it is determined that the entire bipolar weighted argument graph has reached a stable state, the iteration is terminated early, and the final position strength mapping table is output. If it is false, repeat S1003 to S1007 until the convergence condition is met or the maximum number of iterations is reached, and the final position strength mapping table is output.

7. The method as described in claim 6, characterized in that, The formula for calculating the position strength of a node in the next round is: in, Let the position strength of node e be in the next round. The strength of the position in the current round, BS(e) Based on the basic strength, α is the influence factor. and These are the weights for supportive and attacking relationships, respectively. and These represent the strength of the positions of supporters and attackers in the current round, respectively.

8. The method as described in claim 1, characterized in that, The chain of evidence for tracing the core causal relationship includes: S1101: Input the set of atomic arguments to be explained, the bipolar weighted argument graph, the final position strength mapping table, and the maximum backtracking depth; use the atomic arguments as the current target risk nodes, initialize the attack chain and support chain, which are used to store the key nodes in the attack path and support path, respectively, and execute S1102 and S1104 respectively. S1102: Among all the input edges of the current target risk node, filter out the relationship edges of the type of attack, calculate the actual attack influence of the source node of each edge, select the node with the largest actual attack influence as the strongest attacker in this layer, and add it to the attack chain. S1103: Take the strongest attacker as the new current target risk node, repeat S1102, and continue to trace back to its upstream attacker until there are no effective attackers or the maximum backtracking depth is reached, thus completing the construction of the attack chain. S1104: Among all the input edges of the current target risk node, filter out the relationship edges of type support, calculate the actual support influence of the source node of each edge, select the node with the largest actual support influence as the strongest supporter of this layer, and add it to the support chain. S1105: Take the strongest supporter as the new current target risk node, repeat S1104, and continue to trace back to its upstream supporters until there are no effective supporters or the maximum backtracking depth is reached, thus completing the construction of the support chain. S1106: Once the attack chain and support chain have been traced, the target risk nodes that need to be explained, the final position strength score, support chain, and attack chain are encapsulated into a structured argument summary.

9. The method as described in claim 1, characterized in that, Also includes: Create two empty task lists: a high-priority task list and a low-priority task list, to store review tasks that require in-depth expert involvement and review tasks that can be approved quickly, respectively. Iterate through each audit opinion in the structured audit opinion text, and based on the position strength score corresponding to the audit opinion, determine that the audit opinion is highly controversial for medium-risk audit opinions and add it to the high-priority task list; otherwise, determine that the audit opinion is highly credible and add it to the low-priority task list. The audit opinions in the high-priority task list are sent to the expert in-depth review workbench for manual review and intervention; Audit comments from low-priority task lists can be directly pushed to the automated approval channel for rapid processing.

10. The method as described in claim 9, characterized in that, Also includes: The adopted expert opinions are stored in the bidding knowledge ontology network in the form of expert experience knowledge nodes, so as to realize the self-evolution and updating of the knowledge graph.

Images