Knowledge graph-based bidding supplier risk intelligent detection method and device

Abstract

The application discloses a kind of based on knowledge graph's bidding supplier risk intelligent detection method and device, comprising: constructing bidding supplier risk knowledge graph, for the graph entity of knowledge graph, embedding vector is generated for each graph entity using graph entity embedding model;Original query sentence input by user is obtained, and is converted into structured query;Query intent in structured query, query entity and extended query association dimension are sequentially based on, in knowledge graph, obtain target subgraph associated with query sentence matching;Risk path is obtained by carrying out multi-hop risk conduction detection based on reinforcement learning algorithm in target subgraph, and risk degree analysis is carried out to the risk path obtained, and the whole process automation from natural language query to risk result visualization is realized by the application.

Description

Technical Field

[0001] This invention relates to the field of bidding and tendering technology, specifically to a knowledge graph-based intelligent risk detection method and device for bidding and tendering suppliers. Background Technology

[0002] Bidding supervision and supplier qualification review are crucial for maintaining market fairness and ensuring the rational allocation of public resources. However, traditional risk detection methods are limited by technology and cannot meet the current requirements for refined and real-time supervision. Specific pain points are as follows:

[0003] 1. The query expression capabilities are mismatched with complex regulatory requirements.

[0004] In practice, regulators (users) often express multi-dimensional detection needs using natural language, such as "finding companies associated with Construction Company A that participated in the bidding for a new hospital construction project in a certain city in 2024" or "investigating the shareholding companies of suppliers with bid-rigging records in the past six months." These queries involve complex conditions such as "entity association," "project constraints," and "time range," and are characterized by ambiguity and insufficient completeness. However, traditional systems only support preset fixed query templates (such as "query a company's bidding records" or "query bidders for a certain project"), and cannot parse complex needs. This forces regulators to split queries multiple times, manually merge results, and even fails to accurately pinpoint target risks.

[0005] 2. Semantic understanding has a high barrier to entry, making it difficult for non-professionals to operate.

[0006] Traditional systems rely on specialized query syntax such as SQL, requiring regulators to master complex statements like "SELECT * FROM supplier WHERE legal person = 'Zhang Mou' AND bidding project IN (SELECT project ID FROM project WHERE bidding time > '2024-01-01')". However, most regulators lack technical backgrounds and struggle to use this specialized syntax, leading to a contradiction of "knowing how to regulate but not how to query". Frontline regulators, unable to write SQL, cannot independently complete complex risk queries and must rely on technical personnel, significantly reducing regulatory efficiency.

[0007] 3. Weak ability to identify complex and interconnected risks.

[0008] The core characteristic of bid rigging and collusion is "indirect cross-entity connections," such as "Supplier A → Legal Entity B → Supplier C → Jointly Bidding on Project X" or "Supplier D → Equity Partner Supplier E → Supplier F with the Same Registered Address → Bid Price Deviation Rate <5%." Traditional detection technologies can only identify single-layer direct relationships such as "supplier-project" or "supplier-legal entity," and cannot penetrate entity links of two hops or more to discover hidden connections. In bid rigging cases missed by traditional methods, 85% involve indirect connections of three hops or more, allowing violations to remain hidden for extended periods and severely damaging market fairness. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides a knowledge graph-based intelligent risk detection method and device for bidding suppliers, which transforms fuzzy requirements into structured queries, lowers the barrier to entry, and supports multi-hop risk transmission detection. The technical solution is as follows:

[0010] Firstly, a knowledge graph-based intelligent risk detection method for bidding suppliers is provided, which includes the following steps:

[0011] A risk knowledge graph for bidding suppliers is constructed. The knowledge graph is constructed with graph entities as nodes and graph entity relationships as edges. Graph entities include: supplier entities, personnel entities, project entities, risk event entities, and document entities. For the graph entities of the knowledge graph, an embedding vector is generated for each graph entity using a graph entity embedding model based on graph attention network GAT.

[0012] The system obtains the original query statement input by the user, extracts the query entity and query intent through semantic parsing, expands the query content according to the query intent, and converts it into a structured query. The structured query includes query entity information, expanded query association dimensions, constraints, and intent type.

[0013] Based on the query intent, query entity, and extended query association dimensions in the structured query, the target subgraph that matches and is associated with the query statement is obtained in the knowledge graph.

[0014] Risk paths are obtained by performing multi-hop risk propagation detection based on reinforcement learning algorithm in the target subgraph. The risk level of the obtained risk paths is analyzed to obtain risk detection results for responding to user query statements.

[0015] Secondly, a knowledge graph-based intelligent risk detection device for bidding suppliers is provided, comprising:

[0016] The knowledge graph unit is used to construct a risk knowledge graph for bidding suppliers. The knowledge graph is constructed with graph entities as nodes and graph entity relationships as edges. Graph entities include: supplier entities, personnel entities, project entities, risk event entities, and document entities. For the graph entities of the knowledge graph, an embedding vector is generated for each graph entity using a graph entity embedding model based on graph attention network GAT.

[0017] The query statement parsing unit is used to obtain the original query statement input by the user, obtain the query entity and query intent through semantic parsing of the original query statement, expand the query content according to the query intent, and convert it into a structured query. The structured query includes query entity information, expanded query association dimensions, constraints, and intent type.

[0018] The subgraph acquisition unit is used to obtain the target subgraph that matches and is associated with the query statement in the knowledge graph based on the query intent, query entity and extended query association dimension in the structured query.

[0019] The risk detection unit is used to perform multi-hop risk propagation detection in the target subgraph based on reinforcement learning algorithm to obtain risk paths, analyze the risk level of the obtained risk paths, and obtain risk detection results for responding to user query statements.

[0020] Thirdly, a computer device is provided, the computer device comprising:

[0021] Memory, used to store executable instructions;

[0022] The processor, when executing executable instructions stored in the memory, implements the first aspect of the knowledge graph-based intelligent risk detection method for bidding suppliers.

[0023] Fourthly, a computer-readable storage medium is provided, storing executable instructions, which, when executed by a processor, implement the knowledge graph-based intelligent risk detection method for bidding suppliers described in the first aspect.

[0024] The present invention provides a knowledge graph-based intelligent risk detection method and apparatus for bidding suppliers, which has the following beneficial effects:

[0025] This invention obtains query entities and query intents from the original query statement through semantic parsing, expands the query content according to the query intent, and converts it into a structured query. It achieves accurate understanding and intelligent rewriting of natural language queries, supports fully automatic processing of unstructured natural language queries, requires no professional syntax, and transforms ambiguous requirements into structured queries through semantic understanding, entity recognition, and query expansion, thereby lowering the usage threshold and ensuring query accuracy.

[0026] This invention integrates structured and unstructured data to construct a multimodal knowledge graph, covering six core dimensions including "equity, personnel, address, business, documents, and historical risks". It supports multi-hop risk transmission detection with three or more hops and accurately identifies complex association paths such as "supplier → legal person → supplier → project" and "supplier → equity participation → supplier → document similarity → supplier", ensuring that no risk is missed. Attached Figure Description

[0027] Figure 1 This is a flowchart of the intelligent risk detection method for bidding suppliers based on knowledge graphs in the embodiments of this application;

[0028] Figure 2 This is a schematic diagram of the method for constructing a risk knowledge graph and embedding entities for bidding suppliers in this application embodiment;

[0029] Figure 3 This is a schematic diagram of the process of parsing, expanding and converting the user's original query statement into a structured query in the embodiments of this application;

[0030] Figure 4 This is a schematic diagram of the query parsing model based on the basic BERT model in an embodiment of this application;

[0031] Figure 5 This is a schematic diagram of the model structure for identifying bid-rigging evasion behavior in an embodiment of this application;

[0032] Figure 6 This is a schematic diagram of the intelligent risk detection device for bidding suppliers based on knowledge graphs in this application embodiment. Detailed Implementation

[0033] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0034] See Figure 1 This application provides a knowledge graph-based intelligent risk detection method for bidding suppliers, which includes the following steps:

[0035] Step 1: Construct a risk knowledge graph for bidding suppliers. The knowledge graph is constructed with graph entities as nodes and graph entity relationships as edges. Graph entities include: supplier entities, personnel entities, project entities, risk event entities, and document entities. For the graph entities of the knowledge graph, an embedding vector is generated for each graph entity using a graph entity embedding model based on graph attention network GAT.

[0036] Step 2: Obtain the original query statement input by the user, extract the query entity and query intent through semantic parsing, expand the query content according to the query intent, and convert it into a structured query. The structured query includes query entity information, expanded query association dimensions, constraints, and intent type.

[0037] Step 3: Based on the query intent, query entity, and extended query association dimensions in the structured query, obtain the target subgraph that matches and is associated with the query statement in the knowledge graph.

[0038] Step 4: In the target subgraph, perform multi-hop risk propagation detection based on reinforcement learning algorithm to obtain risk paths, analyze the risk level of the obtained risk paths, and obtain risk detection results for responding to user query statements.

[0039] This application deeply integrates NLP, knowledge graphs, and reinforcement learning technologies with the business rules, risk characteristics, and regulatory processes of bidding and tendering supervision scenarios, forming a closed-loop solution of "precise query understanding → multimodal data support → efficient risk reasoning → regulatory-friendly output." This addresses the four major pain points of existing technologies: "generalization and incompatibility, single data leading to missed risks, low efficiency and lack of real-time updates, and black-box results making supervision difficult." This application focuses on the bidding and tendering field, designing technical components such as knowledge graphs, intent expansion, multi-hop graph reasoning, and reinforcement learning path selection. These multiple technical components are interconnected, using the pain points of bidding and tendering supervision as a blueprint for scenario-based transformation and a closed-loop connection throughout the entire process.

[0040] In this embodiment, the three modules of "GraphRAG knowledge graph construction", "intelligent query rewriting" and "multi-hop risk transmission detection algorithm" realize the full-process automation from natural language query to risk result visualization.

[0041] In step 3 above, the structured query output is rewritten based on the intelligent query, and combined with the GAT node embedding of GraphRAG, a subgraph related to the query is retrieved from the complete graph, reducing the amount of data for subsequent multi-hop detection. It should be noted that in this embodiment, when obtaining the target subgraph associated with the query statement, a cross-type comparison of the "intent vector" and the "graph entity vector" is first performed, followed by further filtering based on the query entity and extended query association dimensions. First, a cross-type comparison is performed, filtering "entities with this type of risk characteristic" according to the "risk semantics of the query intent," adapting to the "semantic retrieval" of the bidding risk detection scenario. This not only meets the core requirement of "finding associated risk entities" but also quickly narrows down the candidate range. Using cross-type comparison, entities unrelated to "associated risk" (such as suppliers without any associated bidding records) are quickly excluded, reducing the number of candidate entities from over 100,000 to 50, avoiding subsequent traversal of the entire graph by relationship and improving efficiency. First, an initial screening (cross-type comparison of intent vectors) reduces the data volume from the entire graph to 5%. Then, a fine screening (core entities + extended query association dimensions) further identifies key associations. Finally, the subgraph only contains entities strongly related to the query, reducing multi-hop detection computation by more than 90%. This application supplements the initial intent screening, avoiding the slow efficiency problem of directly searching the entire graph using "core entities + extended query association dimensions".

[0042] Specifically, the steps for obtaining the target subgraph that matches and is associated with the query statement include: based on the query intent in the structured query, calculating the cosine similarity between the BERT model encoded vector of the query intent and the embedding vectors of all entities in the graph, to obtain preliminary candidate entities associated with the structured query; further filtering from the preliminary candidate entities based on the query entities in the structured query and the extended query association dimensions to obtain the target subgraph that matches and is associated with the query statement. The intent in the structured query (such as "association risk query") is encoded into a 64-dimensional vector using a fine-tuned BERT model, and the cosine similarity is calculated between this vector and the embedding vectors of all entities in the graph. It should be noted that the entity IDs in the structured query and the entity IDs in the knowledge graph are linked. The intent BERT encoding vector and the GAT entity embedding vector are both semantic mappings of the same ID entity, only with different encoding perspectives (BERT focuses on query intent association, while GAT focuses on graph relationship association). However, the anchor points are consistent. Both BERT and GAT are trained based on bidding and tendering domain data. The core semantics of the same entity (such as ID=C123) ("supplier + high correlation risk potential") are emphasized in both vectors, causing their semantic directions to converge. When calculating cosine similarity, it can accurately identify "semantic associations of the same entity".

[0043] In some implementations, in step 1 above, the entity relationships in the bidding supplier risk knowledge graph include: equity relationships between suppliers, business relationships between suppliers, personnel relationships between suppliers and individuals, business relationships between suppliers and projects related to winning bids, address relationships between suppliers, document relationships between suppliers and documents, document relationships between similar documents, and risk relationships between suppliers and risk events. Specifically, in this application embodiment, the eight categories of graph entity relationships can be shown in Table 1.

[0044] Table 1. Types and Illustrations of Knowledge Graph Relationships

[0045] Serial Number Knowledge Graph Relationship Types Relationship diagram 1 Equity Relationship Supplier A → Holding Company → Supplier B (including holding percentage); Supplier A → Equity Participation → Supplier B; 2 Business relationships (between suppliers) Supplier A → Subcontractor → Supplier B 3 personnel relations Supplier A → Legal Representative → Personnel B; Supplier A → Senior Management → Personnel B; Personnel B → Appointment → Supplier C 4 Business relationships (between suppliers and projects) Supplier A → Bidding → Project P, Supplier A → Winning Bid → Project P 5 Address Relationships Supplier A → Same registered address → Supplier B 6 Document relationships (between suppliers and documents) Supplier A → Submission → Document D 7 Document relationships (between documents) Document D → Similarity > 80% → Document E (corresponding to tender documents submitted by different suppliers) 8 Risk Relationship Supplier A → Involved in → Risk Event E, Risk Event E → Related to → Supplier B

[0046] See Figure 2 In step 1 above, the construction of the risk knowledge graph of bidding suppliers includes:

[0047] Step 11: Collect multi-source heterogeneous data on enterprise operation information and bidding information, including structured business registration data, bidding transaction data, supply chain data, court judgment data, regulatory history data, and unstructured bid documents, bidding announcements, and enterprise annual reports;

[0048] Step 12: Based on multi-source heterogeneous data, extract graph entities and graph entity relationships for constructing the knowledge graph using a hybrid extraction scheme combining a rule engine and the BERT-NER model. Construct the knowledge graph based on the extracted graph entities and graph entity relationships, using graph entities as graph nodes and graph entity relationships as graph edges.

[0049] The construction of the knowledge graph in this application also includes updating the knowledge graph based on real-time updated data to ensure its real-time availability. For example, for newly added completed bidding projects, the graph can be incrementally updated based on the supplier business information and bidding information involved in the bidding project. Real-time incremental updates of the graph and supplier risk detection (on-site verification at the bid opening) can also be performed on projects currently under evaluation. In this case, the real-time acquisition channel for bidding data at the bid opening site can be utilized. After the bid opening process is completed (i.e., after all bidding supplier lists, bid documents, quotations, and other information are confirmed to be correct), the data is synchronized in real time. New bidding data at the bid opening site will immediately trigger an incremental update of the graph, and the entire process does not require a full reconstruction of the graph.

[0050] Specifically, the process of constructing the risk knowledge graph of bidding suppliers in this application embodiment includes the following steps:

[0051] (1) Multi-source data integration, integrating 8 types of core data sources, covering structured and unstructured data: 1) Structured data: business registration data (company name, legal person, equity structure, registered address, registered capital), bidding transaction data (project information, list of bidders, quotation, winning bid results), supply chain data (upstream and downstream relationships of suppliers, subcontractor information), court documents data (company litigation records, bid rigging judgment results), regulatory history data (past violation records, risk level); 2) Unstructured data: tender documents (technical solutions, quotation details, OCR text of scanned copies of qualification certificates), bidding announcements (project requirements, qualification requirements), company annual reports (business scope and related companies in the text description). Data integration process: 1) Data cleaning: Deduplication (e.g., duplicate bidding records), completion (e.g., missing corporate legal person information), format standardization (e.g., date format standardization to "YYYY-MM-DD"), and conflict resolution (if the legal person information of the same enterprise is inconsistent, the latest publicly disclosed data from the Administration for Industry and Commerce shall prevail); 2) Unstructured data processing: For text data such as bidding documents and bidding announcements, OCR technology (for scanned documents) is used to extract the text content, and then preprocessing is performed through word segmentation, stop word removal, and keyword extraction (based on TF-IDF) to prepare for subsequent entity and relation extraction; 3) Data alignment: Based on "unique entity identifiers" (e.g., unified social credit code of enterprises, project number), the same entity in different data sources is associated. For example, "Company A" in the industrial and commercial data (unified social credit code "91110000XXXX") is aligned with "Company A" (the enterprise name in the bidding record) in the bidding data to ensure data consistency ≥99%.

[0052] (2) Entity and Relationship Extraction: Entity extraction adopts a hybrid extraction scheme of "rule engine + BERT-NER model" to extract 5 types of core entities in the graph: 1) Supplier entity: company name, unified social credit code, registered address, legal person, registered capital, establishment time; 2) Personnel entity: name, position, affiliated company, ID number (de-sensitized); 3) Project entity: project number, project name, bidding unit, bidding time, winning bid amount, project location; 4) Risk event entity: event number, event type (bid rigging, qualification fraud, etc.), involved companies, processing result, occurrence time; 5) Document entity: document number, document type (tender document, tender announcement), associated project / company, upload time, text content. The rule engine is used to extract entities in structured data (such as directly extracting company name and legal person from business registration data tables), and the BERT-NER model is used to extract entities in unstructured text (such as extracting project name and subcontractor information from tender documents). The model was trained with 50,000 labeled text samples, achieving an entity extraction accuracy of 96% and a recall of 95%. For relationship extraction, a hybrid "rule + model" approach was also used to extract eight categories of graph entity relationships, covering key dimensions of bid-rigging and collusion. The rule engine was used to extract explicit relationships in structured data (e.g., extracting "Supplier A holds a controlling stake in Supplier B" from an equity table), while the BERT-based relationship classification model was used to extract implicit relationships in unstructured text (e.g., extracting "Supplier A subcontracts to Supplier B" from a tender document). The model was trained with 30,000 labeled relationship samples, achieving a relationship extraction accuracy of 93% and a recall of 92%.

[0053] (3) Incremental update map

[0054] Incremental update trigger conditions: The graph is triggered when the following three types of data changes occur: 1) New data: new suppliers, new projects, new bidding records, new risk events; 2) Data changes: supplier equity structure adjustment, legal person change, supplementary uploading of bidding documents; 3) Data expiration: supplier cancellation, project termination, historical risk records exceeding the validity period (if there are no violations for 3 years, the risk level can be reduced).

[0055] Incremental update process: 1) Change data identification: Identify new / changed / invalid data entries through data change logs (such as database binlog) or regular inspections (compare data differences every morning) to avoid full scan; 2) Local processing: Only the entities and relationships corresponding to the changed data are processed—new data is processed by “entity extraction → relationship extraction → node embedding update”, changed data is processed by “attribute update → relationship adjustment → embedding vector recalculation”, and invalid data is processed by “entity mark invalid → relationship deletion → embedding vector removal”; 3) Graph fusion: Integrate the processed local data into the existing graph, update the association relationships and embedding vectors of related entities, without rebuilding the entire graph; 4) Consistency verification: Check whether the updated data conflicts with the existing graph (such as new equity relationships contradicting historical records). If a conflict exists, manual review is triggered to ensure graph data consistency.

[0056] In this embodiment, incremental updates only process changed data. Compared with traditional full reconstruction (which takes 2 hours for 100,000 suppliers), the update efficiency is improved by more than 80%. For example, the update time for adding 100 suppliers and 20 projects is only 3 minutes, and the updated data can be used immediately for risk detection without delay.

[0057] In one implementation, see Figure 2 In step 1 above, for the graph entities of the knowledge graph, a graph entity embedding model based on the graph attention network GAT is used, including:

[0058] Step 13: For edges on the knowledge graph, use the preset weights of different relationships between graph entities as the initial values ​​of the relationship weights carried on the edges.

[0059] Step 14: For the graph attention network GAT to be trained, with entity risk classification and whether there is a relationship between two entities as training objectives, train the graph attention network GAT to aggregate the neighbor information for each graph node by combining the relationship types and relationship weights on different edges, and generate the embedding vector of each graph entity; it can be understood that the embedding vector contains the neighbor node information of the current entity node in the graph, the different types of relationships with the neighbor nodes, and the relationship weight information.

[0060] Step 15: Based on the trained Graph Attention Network (GAT) as the graph entity embedding model, generate an embedding vector for each graph entity.

[0061] In this embodiment, GAT is used to adaptively control the information aggregation intensity by learning the attention weights of neighboring nodes, effectively improving the model's ability to represent node entities. The GAT model learns the vector representation (node ​​embedding) of entities, with a dimension of 64. Different weights are assigned to different relationships through an attention mechanism, highlighting key associations (e.g., "holding > 50%" has a higher weight than "participation < 10%", and "legal entity association" has a higher weight than "same address association"), thus improving the accuracy of subsequent retrieval and reasoning.

[0062] Training Process: 1) Data Preparation: The constructed graph is converted into a format that can be input into the GAT model. Each entity is a node, and relationships are edges. The weight of the edges is preset based on the importance of the relationships (e.g., the weight of a controlling relationship is set to 1.0, a shareholding relationship to 0.6, and a relationship at the same address to 0.4); 2) Model Training: The training objectives are "entity risk classification" (used to characterize whether a supplier is a high-risk enterprise, i.e., whether a supplier is a high-risk or low-risk enterprise) and "relationship prediction" (e.g., predicting whether there is an unextracted equity relationship between supplier A and B). The cross-entropy loss function is used, the number of training epochs is set to 50, and the learning rate is set to 0.001; 3) Embedding Vector Generation: After training, each entity outputs a 64-dimensional embedding vector. The higher the vector similarity, the stronger the correlation between the entities in terms of risk association (e.g., the embedding vector similarity between two related suppliers is higher than that between unrelated suppliers). In this embodiment, through GAT node embedding, the accuracy of subgraph retrieval is improved by 25%, and the misjudgment rate of risk path reasoning is reduced by 30%, providing accurate vector support for subsequent multi-hop risk detection.

[0063] Specifically, during training, the core of the GAT model is learning the 64-dimensional embedding vectors of entities, while simultaneously dynamically optimizing relationship weights. First, relationship weights are preset for different relationships according to business rules to help the model converge quickly. For example, weights are set based on common sense regarding bidding and tendering regulations (e.g., controlling stakes have stronger control over a company than minority stakes, and legal entity affiliations are more likely to lead to bid-rigging than address affiliations). Then, the relationship weights are optimized through model training. During training, GAT indirectly optimizes relationship weights by calculating the "contribution of each neighbor node to the current node" through an attention mechanism (a core feature).

[0064] When training the GAT model, the "triples" in the graph are used as the basic unit, namely (head entity, relationship, tail entity), such as (supplier A, holding company, supplier B) and (supplier C, legal entity related, personnel D). Each entity is given an "entity risk classification label", such as supplier A being labeled "high risk" and supplier E being labeled "low risk". The labels are derived from regulatory historical data and court records. The court records contain data such as corporate litigation records and judgments on bid rigging and collusion. The regulatory historical data contains data such as past violations and risk levels. Suppliers with records of bid rigging, violations, or qualification fraud are labeled "high risk". A "relationship existence label" is also added to the triples, such as (supplier A, equity related, supplier F) being labeled "exists" and (supplier B, same address, supplier G) being labeled "does not exist". After converting the graph into the GAT input format (node ​​= entity, edge = relation + initial weight), the neighboring nodes and corresponding relations of each entity constitute its training samples. For example, the sample of supplier A includes neighbor information such as (supplier B, holding, initial weight 1.0) and (supplier C, same address, initial weight 0.4). The training loss function is a weighted sum of "entity risk classification loss + relationship prediction loss + attention weight regularization loss". The entity risk classification loss, for the objective of "determining whether a supplier is high-risk", calculates the difference between the model's predicted entity risk label and the actual label. Cross-entropy (CE) can be used for this calculation, and it has the highest weight (approximately 0.5), ensuring that the entity embedding reflects risk attributes. The relationship prediction loss, for the objective of "predicting whether a certain type of relationship exists between entities", calculates the difference between the model's predicted relationship probability and the actual label (existence / non-existence). Cross-entropy loss can also be used, and it has a weight of approximately 0.3, ensuring that the embedding vector encodes relationship features. The attention weight regularization loss, a new constraint, prevents the model from over-weighting certain relationships (e.g., excessively amplifying the weight of "holding" while ignoring other associations). It is calculated as the L2 norm of the relationship attention weight, and its weight is approximately 0.2, ensuring a balanced weight distribution. During training, backpropagation is used to simultaneously optimize the entity embedding vector and the relationship attention weight.

[0065] In this embodiment, "entity risk classification" is used as the training objective, so that the entity embedding vectors generated by the trained GAT model have "risk attribute labels" by default. "Relationship prediction" is used as the training objective, so that the entity embedding vectors generated by the trained GAT model encode "association type and strength". The finally learned vectors satisfy both "the stronger the risk association, the more similar the vectors" (e.g., Company A → Holding Company → Company B, vectors are highly similar) and avoid "vectors with no risk association are overly similar" (e.g., Company A → Same address → Company C, vectors are low similarity), perfectly adapting to the subsequent needs of "subgraph retrieval + multi-hop risk inference".

[0066] The knowledge graph constructed in this application addresses the "static" and "shallow" problems of traditional knowledge graphs by integrating multi-source heterogeneous data, constructing a multimodal graph, and implementing incremental updates and GNN node embedding learning, thus providing a data foundation for deep relational reasoning. The knowledge graph construction in this application has the following characteristics:

[0067] 1) Multimodal integration: For the first time, unstructured bidding documents (technical solutions, quotation details) are included in the graph. Implicit relationships such as "document similarity" are extracted through text processing and similarity calculation, filling the gap that traditional graphs only cover structured data, and improving the risk detection recall rate by 31%.

[0068] 2) Incremental update mechanism: Avoid full map reconstruction, only process new / changed data, improve update efficiency by 80%, solve the data lag problem caused by the "static" nature of traditional maps, and meet the data timeliness requirements of real-time supervision.

[0069] 3) GAT Embedding Learning: By highlighting key associations through the attention mechanism, the discriminativeness of entity embedding vectors is improved by 40%, and the accuracy of subgraph retrieval and risk inference is improved by 25% respectively, providing accurate data support for subsequent multi-hop risk detection.

[0070] In one implementation, see Figure 3 In step 2 above, the original query statement input by the user is obtained, the query entity and query intent are obtained through semantic parsing, the query content is expanded according to the query intent, and converted into a structured query, including:

[0071] Step 21: For the original query statement, use the query statement parsing model based on the basic BERT model to parse and obtain the query entity information and query intent type in the query statement;

[0072] Step 22: Invoke the entity linker and associate the queried entity with the entity ID in the knowledge graph based on the "name + attribute" matching strategy;

[0073] Step 23: Obtain extended query content based on query intent type;

[0074] Step 24: Based on the query entity information, query intent type, and query extended content, transform the query into a first structured query; the first structured query includes: query entity information, extended query association dimensions, constraints (time, project, anomaly threshold, etc.), and intent type, wherein the query entity information includes entity ID, entity name, type, and attributes.

[0075] In this embodiment, the entire process of natural language querying for bidding supplier risks—"semantic understanding → entity recognition → intent classification → query expansion → logic optimization → structured output"—is implemented to ensure query accuracy and subsequent retrieval efficiency. In step 21 above, a query statement parsing model is pre-trained based on "bidding domain knowledge text + corresponding different types of labeled data." This allows the model to learn the terminology, entity relationships, and semantic logic of the bidding domain, enabling a more comprehensive understanding of domain semantics and avoiding the limitation of only understanding queries without understanding risk scenarios and rules. The query entity-related information in the query statement includes: entity name, entity type, entity semantic role, and entity attributes. In step 22 above, based on the "name + attribute" matching strategy, the query entity is associated with the entity ID in the knowledge graph. In this embodiment, the entity in the query statement (such as "Company A") is bound to the unique ID of the graph (such as C123), and accurately associated with the 64-dimensional embedding vector trained by GAT. The entity link completes the mapping of "entity name → ID". When the query statement is matched with the knowledge graph to obtain the target subgraph, there is no need to deal with entity ambiguity. The vector can be called directly by ID to ensure accurate matching. This "name + attribute" matching strategy first matches the entity name in the query statement against the knowledge graph. If multiple entities are matched, it further filters the knowledge graph based on the entity attributes in the query statement, ensuring an entity link accuracy rate of ≥98% and resolving ambiguity issues related to entities with the same name. For example, if the entity in the query statement is "Company A", it first filters using "entity name + type" (e.g., finding all suppliers named "Company A"); then it uses core attributes for precise filtering (e.g., combining "project location = a certain city" in the query to match Company A with "registered address = a certain city"); finally, it associates the entity with a unique entity ID in the knowledge graph. In steps 23 and 24 above, the query content is expanded according to preset expansion rules based on the query intent type. Then, the information in the user's original query statement is transformed into a structured format recognizable by the knowledge graph retrieval, including query entity information (entity ID, entity name, entity type, entity attributes), expanded query association dimensions, constraints (time, project, anomaly threshold, etc.), and intent type.

[0076] In one implementation, the method for obtaining query entity-related information in the query statement in step 21 above includes:

[0077] Step 2101: Construct and train the first branch of the query parsing model: The first branch includes a semantic understanding module based on the basic BERT model and a first parsing submodule connected in series. The first parsing submodule is a named entity recognition output layer (see the schematic diagram of the first branch structure of the query parsing model). Figure 4 );

[0078] Step 2102: Create the first training data based on historical query statements, risk reports, and business rule documents. Label the first training data with entity types, semantic roles, query intents, and risk logical relationships. The entity types include suppliers, projects, personnel, time, risk indicators, and risk events; the semantic roles include subjects, objects, constraints, results, and evidence.

[0079] Step 2103: Based on the first training data and its labeled data, train the first branch. The first branch is used to perform semantic understanding, parsing keywords and entity recognition on the input query statement. The semantic understanding module outputs a semantic vector, and the first parsing submodule outputs the entity recognition result based on the semantic vector.

[0080] Step 2104: Based on the user query statement to be analyzed, use the first branch of the trained query statement parsing model to output the entity recognition result, including entity name, entity type, entity semantic role and entity attributes.

[0081] In this embodiment, the basic BERT model (bert-base-chinese) is used as the foundation, and 100,000 labeled data points (including historical query statements, risk reports, and business rule documents) from the bidding and tendering field are used for secondary training. The semantic understanding module based on the basic BERT model (bert-base-chinese) outputs semantic vectors ([CLS] vectors / Token-level vectors). The named entity recognition output layer uses the CRF decoding layer to output entity type labeled data from the Token vectors, and uses the linear layer to output the semantic roles of the entities from the [CLS] vectors (converting their corresponding hidden states into the dimension of the label space through linear transformation).

[0082] The three data sources have different purposes (covering "query understanding, risk description, and rule definition"), and the specific forms of the samples may vary slightly, but the essence of the sample forms is the same: a piece of text in the bidding field + structured annotation. The core of the annotation is to help the model master: domain terminology (bid collusion, bid deviation rate, relationship, etc.), entity types (supplier, project, time, threshold, etc.), and semantic logic (who - what is it targeting - what conditions are met - what results are produced).

[0083] For Data Source 1: Historical Query Statements (core adaptation "Natural Language Query Parsing"), the samples are real query statements from regulatory personnel. The core annotation is "entity + entity type + semantic role," helping the model learn to extract key information from fuzzy queries. Example:

[0084] Sample text (natural user query): "Query the bidding risks of A Construction Company's affiliated companies in the construction project of a hospital in a certain city in 2024".

[0085] Annotated data (structured tags):

[0086] 1. Entity List:

[0087] Name: Construction Company A, Type: Supplier, Semantic Role: Subject

[0088] Name: 2024, Type: Time, Semantic Role: Constraint

[0089] Name: New Construction Project of a Municipal Hospital, Type: Project, Semantic Role: Constraint

[0090] Name: Bidding Risk, Type: Risk Indicator, Semantic Role: Object

[0091] 2. Intent Type: Related Risk Query

[0092] In this embodiment, the sample text is the user's original query, retaining ambiguity and natural language features (such as "related enterprises" not being explicitly related to equity / personnel); the labeled data not only includes entities and types, but also supplements "semantic roles", allowing the model to understand "who (subject) - what (object) - what restrictions (constraints)", laying the foundation for subsequent query expansion and logic optimization.

[0093] For data source 2: risk reports, we adapt to "risk event semantic understanding." Risk reports describe investigated and prosecuted violations. Samples are key sentences / paragraphs from these reports, and the core annotation is "risk-related entities + relationships + semantic roles," helping the model understand the expression logic of scenarios such as bid-rigging and collusion. Example:

[0094] Sample text (risk report excerpt): "In March 2024, Supplier B (Unified Credit Code 91110000XXXX) and Supplier C were identified as bid rigging for a 'municipal road renovation project' because they jointly bid and their bid documents had a similarity of 85%, involving a winning bid amount of 12 million yuan."

[0095] Annotated data (structured tags):

[0096] 1. Entity List:

[0097] Name: Supplier B, Type: Supplier, Semantic Role: Risk Subject

[0098] Name: Supplier C, Type: Supplier, Semantic Role: Risk Subject

[0099] Name: March 2024, Type: Time, Semantic Role: Constraint

[0100] Name: Municipal Road Renovation Project; Type: Project; Semantic Role: Associated Object

[0101] Name: Bid rigging, Type: Risk Event, Semantic Role: Result

[0102] Name: 85%, Type: Risk Indicator (Document Similarity), Semantic Role: Evidence

[0103] 2. Relationship labeling: Supplier B → Joint Bidding → Municipal Road Renovation Project; Supplier B → Document Similarity → Supplier C

[0104] In this embodiment, the sample text is a factual description in the report, which includes entities, relationships, and risk outcomes; the labeled data supplements the semantic roles and entity relationships of "risk events" and "evidence", enabling the model to learn to identify the core associations in risk scenarios (such as "joint bidding + document similarity" = collusion related statements).

[0105] For data source 3: business rule documents (adapted to "industry rule semantic parsing"), business rules are regulatory judgment standards, samples are rule clauses, and the core annotation is "rule-related entities + conditions + results," helping the model understand professional expressions such as "abnormal thresholds and judgment logic." Example:

[0106] Sample text (business rule clause): "If the same supplier or its affiliated companies (equity / personnel / address association) submit more than 2 bids for the same project, or if the bid price deviation rate is lower than the industry average of 5%, it is considered an abnormal bid."

[0107] Labeled data (structured tags)

[0108] 1. Entity List:

[0109] Name: Same Project, Type: Project, Semantic Role: Constraint

[0110] Name: 2 times, Type: Numerical threshold, Semantic role: Conditional threshold

[0111] Name: 5%, Type: Numerical Threshold (Quote Deviation Rate), Semantic Role: Conditional Threshold

[0112] Name: Bidding Anomaly, Type: Risk Indicator, Semantic Role: Result

[0113] Name: Equity / Personnel / Address Association, Type: Relationship Type, Semantic Role: Constraints

[0114] 2. Logical annotation: Condition 1 (Number of bids > 2) ∨ Condition 2 (Price deviation rate < 5%) → Result (Bidding anomaly)

[0115] In this embodiment, the sample text is the original rule text, which includes the judgment conditions, thresholds and results; the labeled data clarifies the "condition thresholds" and "logical relationships", enabling the model to understand the professional terms (such as "quotation deviation rate") and judgment logic (or / and) in the industry rules, and providing support for subsequent query expansion (such as automatically adding "quotation deviation rate < 5%" as an abnormal condition).

[0116] In one implementation, step 2103 above, when training the first branch, adopts a contrastive learning strategy, using fuzzy queries and corresponding standard structured queries as paired samples, using the minimization of semantic difference loss as the first loss function, constructing a second loss function based on the error between the labeled data of the samples and the actual output of the first branch, and training the first branch based on the first loss function and the second loss function.

[0117] In this embodiment, a contrastive learning strategy is employed to teach the model the semantic equivalence between "user fuzzy expressions" and "system standard queries." After using contrastive learning, the accuracy of fuzzy query parsing increases from 75% in traditional supervised training to 92%. The core difference lies in the model's ability to map "fuzzy expressions → precise semantics" and its generalization capabilities. Each training sample consists of "fuzzy natural language query + corresponding standard structured query." For example, the user-input fuzzy query "find bidding units related to Company A" corresponds to the standard structured query {"entities": [{"name": "Company A", "id":"C123", "type": "supplier"}], "relations": ["equity holding", "legal entity association"], "constraints": ["Bidding Project ≥ 1"]}, input "fuzzy query" and "standard structured query" into the basic BERT model respectively to obtain the semantic vectors of the two (such as the vector at the [CLS] position). Minimize the semantic difference between the two vectors (commonly cosine similarity loss, triplet loss, etc.) to make the model recognize that the two sentences express the same query requirement; jointly train the first branch based on contrastive learning loss (first loss function) + supervised task loss (second loss function). For specific tasks such as "entity recognition", "semantic role labeling", and "intent classification", construct the first loss function (commonly cross-entropy loss) based on manually labeled data (such as entity type, semantic role label); loss calculation during training, supervised loss is the main method and contrastive learning loss is the auxiliary method. The total loss = α × supervised loss + β × contrastive learning loss (α and β are weights, α is set to be greater than β).

[0118] In one implementation, step 21 above, the method for obtaining the query intent type in the query statement, includes:

[0119] Step 2105: Construct and train the second branch of the query statement parsing model. The second branch includes a pre-trained semantic understanding module and a second parsing submodule connected in series. The second parsing submodule is a query intent classification layer. This intent classification layer sequentially inputs the [CLS] vector output by the semantic understanding module into a linear layer and a softmax classification layer. The softmax classification layer outputs the intent type (see the schematic diagram of the second branch structure of the query statement parsing model). Figure 4 );

[0120] Step 2106: Using historical query statements as the second training data, label the query intent types on the second training data, and use the second training data as input to the second branch to train the second parsing sub-model; the second branch is used to identify the query intent types in the user's query statement; the query intent types include: related risk query, bidding anomaly detection, historical violation investigation, qualification compliance review, and batch risk screening;

[0121] Step 2107: Based on the user query statement to be analyzed, the second branch of the trained query statement parsing model is used to output the user's query intent type.

[0122] In this embodiment, the semantic understanding module and named entity recognition output layer in the first branch are first trained using 100,000 training data points in the bidding domain, including historical query statements, risk reports, and business rule documents. Then, 10,000 historical query statements are selected from these 100,000 data points as samples for further annotation, ensuring that the training data for the intent classification task is semantically consistent with the training data of the first branch, thereby improving classification accuracy. The training focus of the first branch is "semantic parsing, entity recognition, and mapping between fuzzy queries and structured queries," while the intent classification task is based on this, with an additional 10,000 annotated query samples for "classification task fine-tuning," specifically strengthening "the ability to distinguish multiple types of intents." The first and second branches reuse the domain semantic understanding capabilities of the same semantic understanding module (such as bidding terminology recognition and semantic understanding of association relationships), allowing intent classification to be trained from scratch and ultimately achieving a high accuracy of 94%, avoiding semantic inconsistencies caused by model fragmentation. In this embodiment of the application, the entire query statement parsing model includes a first branch and a second branch. The entire query statement parsing model is responsible for "understanding the query and extracting key information" (semantic understanding + entity recognition), and through "specialized training", it has the ability to "determine the purpose of the query" (intent classification). The two achieve task collaboration of "two sides of one body".

[0123] In one implementation, step 23 above, obtaining extended query content based on the query intent type, includes:

[0124] If the intent type is "Association Risk Query," then "association" is expanded to five core relationships: equity holding, legal representative association, senior management association, same registered address, and supply chain cooperation. If the intent type is "Bidding Anomaly Detection," then three anomaly conditions are added: number of bids for the same project > preset bid number threshold, bid price deviation rate < preset deviation rate threshold, and bid document similarity > preset document similarity threshold. If the intent type is "Batch Risk Screening," then batch processing parameters are automatically added. In this embodiment, based on the bidding domain knowledge base (including business rules and risk dimension definitions), and combined with the categorized intents, query conditions and relationship dimensions are automatically added to avoid users missing key risk points. If the intent is "related risk query", then "related" will be expanded to 5 core relationships: equity holding (including controlling and participating shares), legal representative association, senior management association, same registered address, and supply chain cooperation (upstream suppliers and downstream subcontractors); if the intent is "bid anomaly detection", then 3 anomaly conditions will be added: number of bids for the same project > 2 (same supplier or related supplier), bid price deviation rate < 5% (compared with industry average or other bidders' bids), and bid document similarity > 80% (calculated based on text similarity algorithm); if the intent is "batch risk screening", then batch processing parameters will be automatically added (e.g., 50 suppliers will be tested per batch, and the concurrency will be controlled to 10 batches).

[0125] During the expansion process, the system records the expansion conditions and relationship dimensions, and displays "Expansion Description" in the final result (such as "Automatically expanded 'association' to equity, personnel, and address associations"), which supports manual adjustment by users, balancing automation and flexibility.

[0126] In one implementation, see Figure 3 After obtaining the first structured query in step 24 above, the process also includes:

[0127] Step 25: Analyze the execution priority based on the constraints in the first structured query, adjust the execution order of the first structured query to obtain an optimized execution order, and form a second structured query based on the first structured query and the optimized execution order.

[0128] In step 25, the methods for adjusting the query execution order include:

[0129] Step 251: Obtain the first structured query, which includes: query entity information, extended query association dimensions, constraints, and intent type;

[0130] Step 252: Extract multiple types of key information from the first structured query. The key information includes: core entities, filter condition set, and relationship set. Then, classify the conditions in the filter condition set into strong filtering conditions and weak filtering conditions according to the fields.

[0131] Step 253: Based on the query dataset size and preset rules corresponding to each filtering condition, determine the priority execution order of the filtering conditions; the preset rules include: taking filtering conditions based on globally unique identifiers or fixed ranges as the first priority; the globally unique identifiers or fixed ranges include: project ID, time interval;

[0132] Step 254: Following the fixed logic of filtering first, then association, and finally detection, and combining the priority execution order of the filtering conditions, optimize the logic execution order to obtain the adjusted query execution order.

[0133] In this embodiment of the application, the user's original query statement is further optimized by adjusting the query execution order through the strategy of "filtering first and then associating", thereby reducing the amount of data for subsequent graph retrieval and improving efficiency.

[0134] Specifically, the methods for adjusting the query execution order include:

[0135] Assume that the first structured query has been obtained (including entities, constraints, extended query relational dimensions, and intent).

[0136] (1) First, extract three types of key information from the structured query:

[0137] -Entity Core: Extracts core entities (such as "Company A", ID: C123, type: Supplier) from entities, which serve as anchors for related operations;

[0138] - Filter set (filters): Extracts "independently filterable conditions" from constraints, categorized by field type: Strong filter conditions (high_priority_filters): Conditions that can significantly narrow the data range, such as project (e.g., "P001"), time (e.g., "last year"), and region (usually the amount of data after filtering is ≤30% of the original data); Weak filter conditions (low_priority_filters): In contrast to strong filter conditions, weak filter conditions are conditions that cannot significantly narrow the data range, and their effect on reducing the data range is relatively limited. The amount of data in the result set after applying weak filter conditions is usually higher than 30% of the original total data, for example, the amount of data after filtering is ≥50% of the original data, and may even retain most of the original data. After applying weak filter conditions, it is necessary to further combine them with other filter conditions for data filtering. For example, the number of bidding projects (e.g., ">1") and the price deviation rate threshold (e.g., "<5%)" are weak filter conditions and need to be further filtered based on the relationship;

[0139] - Relationship set: Extract the related operations to be performed (such as equity and legal entity relationships) from the relationships.

[0140] (2) Evaluate the "screening strength" of the filtering conditions, determine the priority order, and judge which filtering conditions are suitable for "priority execution" by using the preset rules and the size of the query dataset corresponding to the filtering conditions:

[0141] Preset rules: Prioritize conditions "based on globally unique identifiers / fixed ranges" (such as project ID, time interval). These conditions do not depend on association results and can directly filter subsets from the graph.

[0142] Evaluation of the dataset size corresponding to the filtering conditions: The statistical interface of the Neo4j database is called to quickly query the dataset size corresponding to each filtering condition. For example: query the total base number of "all suppliers": 100,000+; query the base number of "suppliers participating in the bidding for project P001": 500+; if the base number of the data after filtering by a certain condition is ≤ 10% of the original base number, it is marked as a "strong filtering condition" and included in the priority execution queue.

[0143] Final output: A list of filter criteria sorted by "screening strength from largest to smallest" (e.g., [Project Filtering → Time Filtering → Number of Bids Filtering]).

[0144] (3) Generate an optimized execution plan and rearrange the operation sequence.

[0145] The code follows a fixed logic of "filter first → associate then check finally", and generates an execution plan based on the filtering priority in step 2:

[0146] For example, for the query "the bidding risk of Company A's related companies in the construction project of a hospital in a certain city", the optimized logic is as follows: First, filter out "all suppliers participating in the bidding of the construction project of a hospital in a certain city (P001)", second, search within this subset for "suppliers that are related to Company A (C123)", and third, detect the bidding anomalies of these suppliers. Compared with the traditional "associate first and then filter", it can reduce the amount of data processing by 70% and improve the query efficiency by 77%.

[0147] For example, the second structured output of the query "Find bidding entities associated with Company A" is:

[0148] {

[0149] "entities": [{"name": "Company A", "id": "C123", "type": "Supplier"}],

[0150] "relations": ["equity holding", "legal representative", "senior management", "same registered address", "supply chain cooperation"],

[0151] "constraints": [{"key": "Number of Bidding Projects", "value": ">1"}, {"key": "Bidding Time", "value": "Recent 1 Year"}],

[0152] "intent": "Related risk query",

[0153] "execution_order": ["Filter suppliers for bidding items > 1", "Retrieve associations with C123", "Detect bidding anomalies"]

[0154] }

[0155] Structured queries support JSON format output, which can be directly called by the GraphRAG knowledge graph module.

[0156] In summary, the original user query statements in this embodiment, after parsing, entity recognition, query rewriting, and structured output, have the following characteristics:

[0157] (1) Domain-specific fine-tuning of BERT model: For the first time, the pre-trained BERT model is adapted to the bidding scenario. Through secondary training with bidding domain data, the problem of insufficient understanding of professional terms by the basic BERT model is solved. The query understanding accuracy is increased from 65% of the traditional method to 92%, ensuring accurate parsing of fuzzy queries.

[0158] (2) Multi-dimensional query expansion: Combining business rules and intent classification, it automatically supplements risk dimensions and constraints to prevent users from missing key risk points due to insufficient professional knowledge (such as only mentioning "related" and ignoring supply chain cooperation relationships), covering more than 95% of high-frequency risk points of bid rigging and collusion.

[0159] (3) Logic optimization mechanism: By adjusting the execution order of "filter first and then associate", invalid data traversal is reduced, query efficiency is improved by 77%, laying the foundation for real-time response of subsequent map retrieval.

[0160] In one implementation, in step 4 above, the reinforcement learning algorithm includes an agent, a state, an action, and a reward function. The agent is responsible for selecting the next entity node to be detected. The state includes the embedding vector of the current node, the set of visited nodes, and the number of risky paths discovered. The action represents selecting a node from the neighboring nodes of the current node as the next hop. The reward function represents the risk relevance of the selected node.

[0161] In the process of determining risky paths using reinforcement learning algorithms, the entities in the structured query are used as starting nodes. In each hop, the agent selects the next hop node based on the current state through the policy network and analyzes whether the edge between the current node and the next hop node is a risky relationship. If so, the edge between the current node and the next hop node is added to the risky path. When the maximum number of hops is reached or there are no unvisited neighbor nodes, the path selection stops.

[0162] Specifically, in this embodiment, the reinforcement learning model is designed as follows: 1) Agent: responsible for selecting the next entity node to be detected; 2) State: includes the embedding vector of the current node, the set of visited nodes, and the number of risky paths discovered; 3) Action: selects a node from the neighboring nodes of the current node as the next hop; 4) Reward: calculates the reward value based on the risk relevance of the selected node. For example, if the next hop node is a high-risk enterprise (with historical violations), the reward value is +1.0; if the next hop node has a bidding relationship with the project in the query, the reward value is +0.8; if the next hop node has been visited or is not related to risk, the reward value is -0.5.

[0163] The steps for using reinforcement learning algorithms to determine risky paths, for example:

[0164] 1) Initialization: Starting with the core entity in the structured query (e.g., "Company A", ID: C123), the set of visited nodes is initialized to {C123}, and the risk path list is empty;

[0165] 2) Iterative selection: In each hop, the agent selects the next hop node based on the current state (C123 embedding vector, visited nodes) through the policy network (an Actor-Critic model based on a fully connected neural network);

[0166] 3) Risk assessment: Check whether the edge between the current node and the next hop node is a risk relationship (such as "equity association + joint bidding project"). If so, add the path (such as C123→Company B→Project P001) to the risk path list.

[0167] 4) Termination condition: Stop path selection when the maximum number of hops is reached (set to 3 hops, covering 95% of bid-rigging scenarios) or there are no unvisited neighbor nodes.

[0168] By dynamically avoiding low-risk paths (such as related companies unrelated to the query project) through reinforcement learning, path selection efficiency is improved by 60%, and invalid computation is reduced by 70% compared to traditional full path traversal.

[0169] In one implementation, step 4 above involves analyzing the risk level of the obtained risk path, specifically including:

[0170] Step 41: Obtain the scores of different risk indicators in the risk path. The risk indicators include relationship strength, historical risk, bidding anomaly, and evidence completeness. The score of relationship strength is determined based on the entity relationship weights obtained during the process of generating embedding vectors for each graph entity using the Graph Attention Network (GAT). The score of historical risk is determined based on the historical violation records of entities in the path. The score of bidding anomaly is determined based on the anomalies in the supplier's current bidding data. The score of evidence completeness is determined based on the quantity and credibility of evidence supporting the risk path.

[0171] Step 42: Based on the scores and weights of different indicators, the risk level analysis results of the risk path are determined by weighted summation.

[0172] In this embodiment, four core indicators are constructed to assess the severity of risk paths from different dimensions, avoiding misjudgment based on a single dimension. Multi-dimensional assessment reduces the risk misjudgment rate from 22% using traditional methods to 7%, ensuring the accuracy of risk level classification.

[0173] 1) Relationship strength: Based on the attention weight of the GAT model, the importance of the relationship is quantified, such as holding > 50% (weight 1.0) > participating 10%-50% (weight 0.6) > participating < 10% (weight 0.3), legal person association (weight 0.9) > senior management association (weight 0.7) > same address (weight 0.4).

[0174] 2) Historical risks: The past violations of entities in the statistical path, such as the number of times of bid rigging and collusion in the past 3 years (+0.3 points each time), the number of times of qualification fraud (+0.2 points each time), and no violation records (0 points).

[0175] 3) Bidding anomalies: Based on the bidding data, calculate the degree of anomaly, such as price deviation rate <3% (+0.5 points), 3%-5% (+0.3 points), >5% (0 points), and bid document similarity >90% (+0.4 points), 80%-90% (+0.2 points), <80% (0 points);

[0176] 4) Completeness of evidence: Assess the quantity and credibility of evidence supporting the risk path, such as having both business registration and shareholding certificates and bidding records (+0.3 points), having only a single piece of evidence (+0.1 points), or having no direct evidence (0 points).

[0177] Risk level calculation: The scores of the four categories of indicators are weighted and summed (the weights of relationship strength, historical risk, bidding anomalies, and evidence completeness decrease successively, with weights of 0.4, 0.3, 0.2, and 0.1 respectively), with a total score range of 0-1.0. Risk is divided into three levels based on the score:

[0178] 1) High risk: Score ≥ 0.7 (e.g., “controlling relationship + two instances of bid rigging in the past 3 years + bid deviation rate of 2% + dual evidence”);

[0179] 2) Medium risk: Score 0.3-0.7 (e.g., "equity participation + no historical violations + quotation deviation rate of 4% + single piece of evidence").

[0180] 3) Low risk: Score < 0.3 (e.g., "same address + no historical violations + quote deviation rate of 6% + single evidence").

[0181] In this embodiment of the application, an interpretable evidence report is also generated for each risk path. The report contains four core parts:

[0182] 1) Risk Path Diagram: Visually displays the links between entities and relationships, such as "Company A (C123) → Legal Representative Zhang (P005) → Company B (C124) → Bidding → New Construction Project of a Municipal Hospital (P001)";

[0183] 2) Risk Level and Score: Clearly define the risk level (high / medium / low) and the specific score for each assessment indicator (e.g., "relationship strength 0.9, historical risk 0.6, bidding anomaly 0.4, evidence completeness 0.3, total score 0.71").

[0184] 3) Evidence List: List the evidence supporting the risk item by item, including the type of evidence (business registration data / tender documents / court documents), the content of the evidence (e.g., "business registration data shows that C123 holds C124 with a holding ratio of 60%, and the announcement date is 2024-01-15"), the source of the evidence (e.g., "National Enterprise Credit Information Publicity System" or "Bidding and Tendering Transaction Platform"), and the link to the evidence (click to view the original document).

[0185] 4) Risk Description: Explain the causes of risks based on business rules, such as "Company B and Company A have a legal relationship and jointly participated in the bidding for Project P001. The bid deviation rate is only 3%, which meets the characteristics of bid rigging and collusion."

[0186] Report output formats: Supports three formats: HTML (for web interface display), PDF (for regulatory document archiving), and JSON (for system interface calls) to meet the needs of different scenarios.

[0187] In summary, the embodiments of this application, through reinforcement learning algorithms and multi-dimensional risk assessment indicators, address the risks of bidding suppliers and have the following characteristics:

[0188] (1) Enhanced learning path selection: dynamically avoids low-risk paths and focuses on high-risk associated links, improving multi-hop detection efficiency by 60%, solving the efficiency bottleneck of traditional full path traversal, and meeting the needs of real-time supervision (response time < 1 second).

[0189] (2) Multi-dimensional risk assessment: Combining four indicators, namely relationship strength, historical risk, bidding anomaly, and evidence integrity, to avoid misjudgment caused by a single dimension (such as only looking at the price). The risk detection recall rate increased from 58% to 89%, and the misjudgment rate decreased from 22% to 7%.

[0190] (3) Explainable chain of evidence: Provides a complete link and original evidence for each type of risk, with 100% evidence integrity. Regulatory personnel do not need to manually verify, reducing manual review costs by 60%.

[0191] In one implementation, after obtaining the risk path and performing a risk level analysis on the obtained risk path, step 4 above further includes:

[0192] Step 51: Using bid-rigging avoidance cases as training data, label entity types, entity relationships, text matching tags, avoidance risk coefficients, and avoidance behavior types;

[0193] Step 52: Based on the training data and the corresponding labeled data, train a bid-rigging avoidance detection model. (See the bid-rigging avoidance detection model...) Figure 5The system comprises a BERT model, a first-stage model connected to the output of the BERT model, and a second-stage model connected to the output of the first-stage model. The first-stage model includes two branches connected in parallel with the BERT model output: latent entity recognition and text matching analysis. The latent entity recognition branch includes a CRF decoding layer to output the latent entity recognition results, and the text matching analysis branch includes a pooling layer and a similarity calculation head connected in sequence to calculate text similarity based on the BERT model output vector. The second-stage model includes an attention layer, two fully connected layers, a BatchNorm layer, a Dropout layer, and an output layer (the sigmoid function outputs the risk avoidance probability value). The second-stage model is used to analyze the supplier's bid-rigging risk coefficient and avoidance behavior type based on the output data of the first-stage model, combined with supplier bidding behavior data and supplier business registration data in the knowledge graph.

[0194] Step 53: Based on the risk coefficient of the supplier's bid-rigging avoidance and the risk level analysis results of the risk path, determine the risk detection results used to respond to the user's query.

[0195] Specifically, the training process for the BERT model and the first-stage model of the BERT model's output connections is as follows:

[0196] Based on 10,000 real-world circumvention case texts, we extracted various types of circumvention behaviors, such as "split equity to below 50%", "rewriting the text of the tender document while maintaining the same logic", and "changing the nominal legal representative". We created sample data based on the real-world circumvention case texts and labeled the samples with the corresponding entity type, entity relationship, and text matching tags to train the BERT model and the first-stage model. For example, using "the business registration announcement regarding the equity split to 49% and the technical solution after rewriting the tender document" as samples, the samples are labeled with "entity tags: actual controller / nominee shareholder; relationship tags: implicit control / relative nominee shareholding; text matching tags: semantically consistent / semantically irrelevant" for training the BERT model and the first-stage model. The first-stage model outputs: ① entity / relationship recognition results (e.g., "supplier A → implicit control → supplier B"); ② text similarity scores (e.g., the semantic similarity between tender document A and tender document B is 0.85). The training loss function of the BERT model and the first-stage model includes the error between the entity recognition results and the real entity labeled data, and the error between the text similarity detection and the real labeled data. Specifically, the error loss of the text similarity detection task can be determined based on the contrastive loss, treating samples that are "semantically consistent but rewritten" as positive sample pairs and "semantically irrelevant" text as negative sample pairs, minimizing the vector distance of positive sample pairs and maximizing the distance of negative sample pairs; the error loss of the entity recognition results can be calculated based on the cross-entropy loss.

[0197] Furthermore, after the BERT model and the first-stage model are trained, new bidding texts / business change documents (such as a supplier's latest equity disclosure or submitted bidding technical solutions) are used as input data. The BERT model and the first-stage model output: ① Robust entities / relationships (such as identifying "nominal legal person C → actual controller D"); ② Text semantic / format similarity scores.

[0198] The training process for the second-stage model includes: using the data output from the first-stage model, the calculation results of related business data, and the supplier bidding behavior data from the related graph to form 12-dimensional feature data as input to the second-stage model. This 12-dimensional feature data includes:

[0199] 1) Equity characteristics (3 dimensions): ① The magnitude of sudden changes in equity ratio over the past 6 months; ② The proportion of equity held by relatives on behalf of others; ③ The number of implicit control levels (such as the number of indirect shareholding levels of the actual controller).

[0200] 2) Text features (3 dimensions): ④ Semantic similarity score of bid documents; ⑤ Format similarity score of bid documents; ⑥ Semantic-format similarity difference;

[0201] 3) Entity Relationship Characteristics (3D): ⑦ Strength of kinship between the nominal legal person and the actual controller (e.g., direct / collateral relatives); ⑧ Overlap of social security payment units for personnel across enterprises; ⑨ Physical distance of registered address (e.g., whether they are in the same industrial park or on the same floor);

[0202] 4) Behavioral characteristics (3 dimensions): ⑩ Time difference between multiple suppliers' bid registrations (e.g., whether they are concentrated in the last hour before the deadline); Price adjustment trajectory synchronization rate; The number of times a tender document has been submitted for the same IP address / device;

[0203] Among them, ①②③⑦⑧: Entity / relationship results from the first-stage BERT-NER module (such as the identified "family nominee shareholding" and "implicit shareholding"), combined with business registration data to calculate specific values; ④⑤⑥: Output scores directly from the first-stage text similarity detection module; : Extracted from bidding behavior data in the knowledge graph (combined with entity associations identified in the first stage).

[0204] In the second stage of model training, the aforementioned 12-dimensional features are used as input data, labeled with: ① manually labeled risk aversion coefficients (e.g., a coefficient of 0.8 corresponding to "equity split + text rewriting"); ② types of risk aversion behaviors (6 categories: equity split / text rewriting / nominal legal person replacement / IP sharing / pricing synchronization / qualification fraud). The second-stage model is trained using the 12-dimensional feature input data and the corresponding labeled data. The output of the second-stage model is: ① a predicted risk aversion coefficient within the range of 0-1; ② the probability distribution of multiple preset types of risk aversion behaviors (6 types of risk aversion behaviors) (internal and external output: the risk aversion behavior type corresponding to the highest probability value among the preset multiple types of risk aversion behaviors). The training loss function of the second-stage model includes the mean squared error (MSE) loss between the model-output risk aversion coefficient and the manually labeled true risk aversion coefficient, and the cross-entropy loss between the risk aversion behavior type and the true labeled type.

[0205] It should be noted that the attention layer of the second-stage model adopts Scaled Dot-Product Attention, which automatically amplifies the weight of key features such as "equity fluctuation magnitude" and "text semantic similarity" by calculating the similarity weights between 12-dimensional features, while weakening the influence of secondary features.

[0206] In one implementation, in this embodiment of the application, when constructing the knowledge graph in step 1 above, the method further includes: adding multiple preset causal chains and corresponding causal strengths (intervention effect data) to the already constructed knowledge graph containing graph entities and graph entity relationships. Furthermore, in step 4 above, when using reinforcement learning algorithms to detect risk paths, the method further includes: performing risk path selection and analysis based on the knowledge graph and the causal chains and corresponding causal strengths (intervention effect data) carried on the knowledge graph.

[0207] Specifically, in step 1, the relationship causal chains and intervention effects added to the knowledge graph are described in Table 2 below. Above the original graph's "entity-relationship layer," eight additional core causal chains (such as "same actual controller → joint bidding → bid rigging") and their corresponding intervention effects are labeled, upgrading the original graph from the "association relationship layer" to the "association + causal relationship layer." The causal strength of each causal chain is quantified using the Do-calculus algorithm. Causal strength = P(outcome | intervention reason) - P(outcome | non-intervention reason), with a value range of 0-1. Causal strength ≥ 0.7 is judged as a high-risk causal chain.

[0208] Table 2. Different types of causal chains and intervention effects

[0209] Serial Number Core causal chain (cause → intermediate behavior → risky outcome) Intervention effect description (change in outcome probability after removing the cause) 1 Same ultimate controller → joint bidding → bid rigging Removing the association with the "same actual controller" reduces the probability of joint bidding by 80% (document example). 2 Falsifying qualifications (such as forging qualification certificates) → Passing the qualification review → Bid invalid Removing "fraudulent qualifications" reduces the probability of passing the qualification review by 95%. 3 Bid document similarity > preset similarity threshold (e.g., 80%) → Collusive bidding → Bid rigging Removing "document similarity > 80%" reduces the probability of collaborative bidding by 75%. 4 Cross-appointment of legal representatives / senior executives → Simultaneous bidding → Related-party bidding violations Removing "cross-appointment of personnel" reduces the probability of simultaneous bidding by 70%. 5 Holding / participating related parties (shareholding >30%) → joint bidding → bid rigging Removing "controlling / participating stake" reduces the probability of joint bidding by 65%. 6 Historical record of bid rigging (≥1 time in the last 3 years) → Participation in the same project bidding again → Recurrence of bid rigging Removing historical bid-rigging records reduces the probability of bid-rigging in subsequent bids for the same project by 60%. 7 Submitting tender documents from the same IP address / device → collaborative operation → bid rigging Removing the "submission from the same IP / device" option reduces the probability of collaborative operations by 90%. 8 Equity relationship between subcontractor and general contractor → Subcontracting winning bid → Transfer of benefits Removing "equity affiliation" reduces the probability of subcontractors winning bids by 55%.

[0210] In step 4, during the reinforcement learning path selection phase, when the agent selects the next hop node, it will prioritize the edge corresponding to the high-risk causal chain with a causal strength ≥ 0.7 (for example, the edge "same actual controller → joint bidding" has a higher weight, and the agent is more inclined to choose the node corresponding to this edge); in the risk judgment phase, "false associations" are excluded by causal strength - for example, "bidding in the same region" is an association but the causal strength is < 0.3, and even if it is selected by the agent, it will not be judged as a risk path.

[0211] Intervention effect data can be obtained through three steps: "historical data statistics + causal simulation calculation + domain rule calibration", as detailed below:

[0212] 1. Step 1: Historical risk case data statistics (core data source)

[0213] Extract verified cases of bid rigging and collusion from the "judgment documents data" and "regulatory history data" of the GraphRAG knowledge graph, and statistically analyze the probability of the outcome variable when there is a "causal variable" and a "no causal variable". For example, if 1,000 cases of bid rigging are analyzed, 800 of them are "related to the same actual controller", and the probability of "joint bidding" in these cases is 90%; at the same time, 1,000 normal bidding cases "without the same actual controller" are analyzed, and the probability of "joint bidding" is only 10%; the preliminary calculation of the intervention effect is: 90% (with causal) - 10% (without causal) = 80% (consistent with the example in the document).

[0214] 2. Second step: Do-calculus algorithm to simulate intervention (data completion)

[0215] If some case data is incomplete (e.g., insufficient observation data for "uncausal variables"), the Do-calculus algorithm is used to simulate intervention scenarios based on the observation data: using the structured data of "supplier-relationship-project" in GraphRAG, a probability model P(outcome|cause, confounding variables) is constructed (confounding variables such as project type and region); based on the "backdoor adjustment formula" of Do-calculus (refer to Oxford causal graph theory), the influence of confounding variables is removed, and the probability of the outcome after "intervention cause" is calculated.

[0216] 3. Third step: Domain expert rule calibration (accuracy optimization)

[0217] By combining the experience of bidding and tendering supervision experts (such as the "business rules knowledge base" mentioned in the document), the intervention effect obtained from statistics / simulation is calibrated.

[0218] See Figure 6 This application provides a knowledge graph-based intelligent risk detection device for bidding suppliers, comprising:

[0219] The knowledge graph unit is used to construct a risk knowledge graph for bidding suppliers. The knowledge graph is constructed with graph entities as nodes and graph entity relationships as edges. Graph entities include: supplier entities, personnel entities, project entities, risk event entities, and document entities. For the graph entities of the knowledge graph, an embedding vector is generated for each graph entity using a graph entity embedding model based on graph attention network GAT.

[0220] The query statement parsing unit is used to obtain the original query statement input by the user, obtain the query entity and query intent through semantic parsing of the original query statement, expand the query content according to the query intent, and convert it into a structured query. The structured query includes query entity information, expanded query association dimensions, constraints, and intent type.

[0221] The subgraph acquisition unit is used to obtain the target subgraph that matches and is associated with the query statement in the knowledge graph based on the query intent, query entity and extended query association dimension in the structured query.

[0222] The risk detection unit is used to perform multi-hop risk propagation detection in the target subgraph based on reinforcement learning algorithm to obtain risk paths, analyze the risk level of the obtained risk paths, and obtain risk detection results for responding to user query statements.

[0223] For specific limitations on the intelligent risk detection device for bidding suppliers, please refer to the limitations on the intelligent risk detection method for bidding suppliers mentioned above, which will not be repeated here.

[0224] This application provides a computer device comprising: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the knowledge graph-based intelligent risk detection method for bidding suppliers as described in the preceding method embodiments. The computer device may include a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions in the memory to execute a knowledge graph-based intelligent risk detection method for bidding suppliers.

[0225] This application provides a computer-readable storage medium storing executable instructions. When executed by a processor, these executable instructions implement the knowledge graph-based intelligent risk detection method for bidding suppliers described in the foregoing method embodiments. The computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage node, etc.

[0226] This invention is not limited to the specific embodiments described above. Any modifications made by those skilled in the art based on the above concept without creative effort are within the scope of protection of this invention.

Claims

1. A knowledge graph-based intelligent risk detection method for bidding suppliers, characterized in that, Includes the following steps: A risk knowledge graph for bidding suppliers is constructed. The knowledge graph is constructed with graph entities as nodes and graph entity relationships as edges. Graph entities include: supplier entities, personnel entities, project entities, risk event entities, and document entities. For the graph entities of the knowledge graph, an embedding vector is generated for each graph entity using a graph entity embedding model based on graph attention network GAT. The system obtains the original query statement input by the user, extracts the query entity and query intent through semantic parsing, expands the query content according to the query intent, and converts it into a structured query. The structured query includes query entity information, expanded query association dimensions, constraints, and intent type. Based on the query intent, query entity, and extended query association dimensions in the structured query, the target subgraph that matches and is associated with the query statement is obtained in the knowledge graph. Risk paths are obtained by performing multi-hop risk propagation detection based on reinforcement learning algorithms in the target subgraph. Risk severity analysis is then performed on the obtained risk paths to obtain risk detection results for responding to user queries. Specifically, the risk severity analysis includes: obtaining scores for different risk indicators within the risk path, including relationship strength, historical risk, bidding anomaly, and evidence completeness. The relationship strength score is determined based on entity relationship weights obtained during the generation of embedding vectors for each graph entity using the Graph Attention Network (GAT); the historical risk score is determined based on historical violation records of entities in the path; the bidding anomaly score is determined based on anomalies in the supplier's current bidding data; and the evidence completeness score is determined based on the quantity and credibility of evidence supporting the risk path. The risk severity analysis results for the risk path are determined by weighted summation of the scores and weights of different indicators.

2. The knowledge graph-based intelligent risk detection method for bidding suppliers according to claim 1, characterized in that, In the risk knowledge graph of bidding suppliers, the entity relationships include: equity relationships between suppliers, business relationships between suppliers, personnel relationships between suppliers and individuals, business relationships between suppliers and projects in terms of winning bids, address relationships between suppliers, document relationships between suppliers and documents, document relationships between similar documents, and risk relationships between suppliers and risk events.

3. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 1, characterized in that, The construction of a risk knowledge graph for bidding suppliers includes: Collect multi-source heterogeneous data on enterprise operation information and bidding information, including structured business registration data, bidding transaction data, supply chain data, court judgment data, regulatory history data, and unstructured bid documents, bidding announcements, and enterprise annual reports; Based on multi-source heterogeneous data, a hybrid extraction scheme combining a rule engine and the BERT-NER model is used to extract graph entities and graph entity relationships for constructing a knowledge graph. The knowledge graph is constructed based on the extracted graph entities and graph entity relationships, with graph entities as graph nodes and graph entity relationships as graph edges.

4. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 1, characterized in that, For graph entities in a knowledge graph, an embedding vector is generated for each entity using a graph entity embedding model based on the Graph Attention Network (GAT), including: For edges on a knowledge graph, the preset weights of different relationships between graph entities are used as the initial values ​​of the relationship weights carried on the edges. For the graph attention network GAT to be trained, the training objectives are entity risk classification and whether there is a relationship between two entities. The training of the graph attention network GAT is to aggregate the neighbor information of each graph node by combining the relationship type and relationship weight on different edges to generate the embedding vector of each graph entity. The entity risk classification represents whether the supplier is a high-risk enterprise. The trained Graph Attention Network (GAT) serves as the graph entity embedding model, generating embedding vectors for each graph entity.

5. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 1, characterized in that, Obtain the user's original query statement, extract the query entities and query intent through semantic parsing, expand the query content based on the query intent, and convert it into a structured query, including: The original query statement is parsed using a query statement parsing model based on the basic BERT model to obtain information related to the query entity and the query intent type in the query statement; The entity linker is invoked to associate the queried entity with the entity ID in the knowledge graph based on the "name + attribute" matching strategy; Obtain extended query content based on query intent type; Based on the query entity information, query intent type, and query extended content, a first structured query is generated; the first structured query includes: query entity information, extended query association dimensions, constraints, and intent type, wherein the query entity information includes entity ID, entity name, type, and attributes.

6. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 5, characterized in that, Methods for obtaining information related to the query entity in a query statement include: The first branch of the query parsing model is constructed and trained: The first branch includes a semantic understanding module based on the basic BERT model and a first parsing submodule, which are connected in series. The first parsing submodule is a named entity recognition output layer. The first training data was created based on historical query statements, risk reports, and business rule documents. The first training data was labeled with entity types, semantic roles, query intents, and risk logical relationships. The entity types include suppliers, projects, personnel, time, risk indicators, and risk events; the semantic roles include subjects, objects, constraints, results, and evidence. Based on the first training data and its labeled data, the first branch is trained. The first branch is used to perform semantic understanding, parsing keywords and entity recognition on the input query statement. The semantic understanding module outputs semantic vectors, and the first parsing submodule outputs entity recognition results based on the semantic vectors. Based on the user query statement to be analyzed, the first branch of the trained query statement parsing model is used to output the entity recognition results, including entity name, entity type, entity semantic role and entity attributes.

7. The knowledge graph-based intelligent risk detection method for bidding suppliers according to claim 6, characterized in that, When training the first branch, a contrastive learning strategy is adopted, using fuzzy queries and corresponding standard structured queries as paired samples. The first loss function is based on minimizing semantic difference loss, and the second loss function is constructed based on the error between the labeled data of the samples and the actual output of the first branch. The first branch is trained based on the first loss function and the second loss function.

8. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 6, characterized in that, Methods for obtaining the query intent type in a query statement include: The second branch of constructing and training the query statement parsing model includes a pre-trained semantic understanding module and a second parsing submodule, which is a query intent classification layer. Using historical query statements as the second training data, the query intent types are labeled on the second training data. The second training data is then used as input to the second branch to train the second parsing sub-model. The second branch is used to identify the query intent types in user query statements. The query intent types include: related risk query, bidding anomaly detection, historical violation investigation, qualification compliance review, and batch risk screening. Based on the user query statement to be analyzed, the second branch of the trained query statement parsing model is used to output the user's query intent type.

9. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 8, characterized in that, Retrieve extended query content based on query intent type, including: If the intent type is "related risk query", then "related" will be expanded to 5 core relationships: equity holding, legal representative association, senior management association, same registered address, and supply chain cooperation; If the intent type is "Bidding Anomaly Detection", then add 3 types of anomaly conditions: number of bids for the same project > preset bid number threshold, bid price deviation rate < preset deviation rate threshold, and bid document similarity > preset document similarity threshold; If the intent type is "batch risk screening", then the batch processing parameters will be automatically added.

10. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 5, characterized in that, After obtaining the first structured query, the process also includes: Based on the constraints in the first structured query, an execution priority analysis is performed, and the execution order of the first structured query is adjusted to obtain an optimized execution order. Based on the first structured query and the optimized execution order, a second structured query is formed. Methods for adjusting the query execution order include: Obtain the first structured query, which includes: query entity information, extended query association dimensions, constraints, and intent type; Extract multiple types of key information from the first structured query. The key information includes: core entities, filter condition sets, and relationship sets. Then, classify the conditions in the filter condition sets into strong filtering conditions and weak filtering conditions according to the fields. Based on the query dataset size and preset rules corresponding to each filtering condition, the priority execution order of the filtering conditions is determined; the preset rules include: taking filtering conditions based on globally unique identifiers or fixed ranges as the first priority; the globally unique identifiers or fixed ranges include: project ID, time interval; Following the fixed logic of filtering first, then association, and finally detection, and combining the priority execution order of the filtering conditions, the logic execution order is optimized to obtain the adjusted query execution order.

11. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 1, characterized in that, The process involves obtaining a target subgraph in the knowledge graph that matches and is associated with the query statement, based on the query intent, query entity, and extended query association dimensions in the structured query, including: Based on the query intent in the structured query, the cosine similarity is calculated between the BERT model encoding vector of the query intent and the embedding vectors of all graph entities in the graph to obtain preliminary candidate entities associated with the structured query. Based on the query entities in the structured query and the extended query association dimensions, the preliminary candidate entities are further filtered to obtain the target subgraph that matches and is associated with the query statement.

12. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 1, characterized in that, The reinforcement learning algorithm includes an agent, a state, an action, and a reward function. The agent is responsible for selecting the next entity node to be detected. The state includes the embedding vector of the current node, the set of visited nodes, and the number of risky paths discovered. The action represents selecting a node from the current node's neighboring nodes as the next hop. The reward function represents the risk relevance of the selected node.

13. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 1, characterized in that, After obtaining risk paths and performing risk level analysis on the obtained risk paths, the process also includes: Based on bid-rigging avoidance cases as training data, entity types, entity relationships, text matching tags, avoidance risk coefficients, and avoidance behavior types are labeled; Based on training data and corresponding labeled data, a bid-rigging avoidance detection model is trained. This model includes a BERT model, a first-stage model connected to the output of the BERT model, and a second-stage model connected to the output of the first-stage model. The first-stage model includes two branches connected in parallel with the BERT model output: implicit entity recognition and text matching analysis. The implicit entity recognition branch includes a CRF decoding layer to output the implicit entity recognition results, while the text matching analysis branch includes a pooling layer and a similarity calculation head connected sequentially to calculate text similarity based on the BERT model output vector. The second-stage model sequentially includes an attention layer, two fully connected layers, a BatchNorm layer, a Dropout layer, and an output layer. The second-stage model is used to analyze the bid-rigging avoidance risk coefficient and avoidance behavior type of suppliers based on the output data of the first-stage model, combined with supplier bidding behavior data and supplier business registration data from a knowledge graph. The risk detection results used to respond to user queries are determined jointly based on the risk coefficient of the supplier's bid-rigging avoidance and the risk level analysis results of the risk path.

14. The intelligent risk detection method for bidding suppliers based on knowledge graphs according to claim 1, characterized in that, When constructing a knowledge graph, the method further includes: adding multiple preset causal chains and corresponding causal strengths to the already constructed knowledge graph containing graph entities and graph entity relationships; when using reinforcement learning algorithms for risk path detection, the method further includes: performing risk path selection and analysis based on the knowledge graph and the causal chains and corresponding causal strengths carried on the knowledge graph; the preset causal chains include: Same ultimate controller → joint bidding → bid rigging and collusion; Falsifying qualifications → Passing qualification review → Bid invalid; Bid document similarity > preset similarity threshold → coordinated bidding → bid rigging; Cross-appointment of legal representatives / senior executives → simultaneous bidding → related-party bidding violations; Holding / participating related parties → joint bidding → bid rigging; Historical bid-rigging records → Participation in the same project bidding again → Bid-rigging recurrence; Submitting tender documents from the same IP address / device → collaborative operation → bid rigging; Equity relationship between subcontractors and general contractors → winning subcontracting bids → transfer of benefits.

15. A knowledge graph-based intelligent risk detection device for bidding suppliers, characterized in that, include: The knowledge graph unit is used to construct a risk knowledge graph for bidding suppliers. The knowledge graph is constructed with graph entities as nodes and graph entity relationships as edges. Graph entities include: supplier entities, personnel entities, project entities, risk event entities, and document entities. For the graph entities of the knowledge graph, an embedding vector is generated for each graph entity using a graph entity embedding model based on graph attention network GAT. The query statement parsing unit is used to obtain the original query statement input by the user, obtain the query entity and query intent through semantic parsing of the original query statement, expand the query content according to the query intent, and convert it into a structured query. The structured query includes query entity information, expanded query association dimensions, constraints, and intent type. The subgraph acquisition unit is used to obtain the target subgraph that matches and is associated with the query statement in the knowledge graph based on the query intent, query entity and extended query association dimension in the structured query. The risk detection unit is used to detect risk paths in the target subgraph using a reinforcement learning algorithm through multi-hop risk propagation detection, analyze the risk level of the obtained risk paths, and obtain risk detection results for responding to user queries. Specifically, the risk level analysis of the obtained risk paths includes: obtaining scores for different risk indicators in the risk path, including relationship strength, historical risk, bidding anomaly, and evidence completeness. The relationship strength score is determined based on the entity relationship weights obtained during the generation of embedding vectors for each graph entity using the Graph Attention Network (GAT); the historical risk score is determined based on the historical violation records of entities in the path; the bidding anomaly score is determined based on anomalies in the supplier's current bidding data; and the evidence completeness score is determined based on the quantity and credibility of evidence supporting the risk path. The risk level analysis result of the risk path is determined by weighted summation of the scores and weights of different indicators.

16. A computer device, characterized in that, The computer device includes: Memory, used to store executable instructions; When a processor is used to execute executable instructions stored in the memory, it implements the knowledge graph-based intelligent risk detection method for bidding suppliers as described in any one of claims 1 to 14.

17. A computer-readable storage medium storing executable instructions, characterized in that, When the executable instructions are executed by the processor, they implement the knowledge graph-based intelligent risk detection method for bidding suppliers as described in any one of claims 1 to 14.

Images