Bid invitation file knowledge graph construction and correlation analysis method based on large model

Abstract

The invention relates to the technical field of intelligent bid invitation file analysis, and discloses a bid invitation file knowledge graph construction and association analysis method based on a large model. Comprising the steps of obtaining multi-source heterogeneous data related to a bid invitation file, and performing data cleaning and standardization on the multi-source heterogeneous data to generate to-be-analyzed data; inputting the to-be-analyzed data into a pre-trained domain large model, and constructing an initial knowledge graph; newly-added bid invitation data and bidding data are obtained regularly, the newly-added bid invitation data and bidding data are preprocessed and input to the field large model, and a newly-added result is generated; the newly added result is fused to the initial knowledge graph, and a bidding and tendering knowledge graph is generated; and performing multi-modal correlation analysis on the bidding and tendering knowledge graph, outputting a risk mode set, and generating and outputting a risk report according to the risk mode set. Automatic construction, dynamic updating and intelligent risk analysis of bidding and tendering knowledge are realized, and bidding management efficiency and risk control capability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing and risk analysis of bidding documents, and in particular to a method for constructing and analyzing a knowledge graph of bidding documents based on a large model. Background Technology

[0002] With the increasing digitalization and complexity of bidding activities, bidding documents and related data are characterized by multiple sources, heterogeneity, and massive volume. Traditional risk analysis methods mainly rely on expert manual review or keyword matching based on simple rules, which has obvious shortcomings. They are difficult to extract entities and their complex relationships accurately and efficiently from unstructured text, resulting in incomplete knowledge construction and low efficiency. Existing methods lack the ability to dynamically integrate and resolve conflicts with new data, making it difficult to update the knowledge base in real time and easily forming information silos.

[0003] Therefore, there is an urgent need in this field for a method that can automatically and accurately construct a knowledge system for bidding and tendering, and can dynamically integrate new data and conduct in-depth correlation analysis to identify complex risk patterns. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing and analyzing knowledge graphs of tender documents based on large models, in order to solve the problems of low knowledge construction efficiency, weak dynamic update capability and insufficient risk analysis in the existing technology.

[0005] In some embodiments of this application, a method for constructing and analyzing a knowledge graph of tender documents based on a large model is provided, including:

[0006] Acquire multi-source heterogeneous data related to the tender documents, and perform data cleaning and standardization on the multi-source heterogeneous data to generate data to be analyzed;

[0007] The data to be analyzed is input into a pre-trained domain-specific large model to construct an initial knowledge graph;

[0008] Regularly acquire newly added bidding and tendering data, preprocess the newly added bidding and tendering data and input them into the domain-wide model to generate new results;

[0009] The newly added results are integrated into the initial knowledge graph to generate a bidding knowledge graph.

[0010] The bidding knowledge graph is subjected to multimodal association analysis to output a risk pattern set. Based on the risk pattern set, a risk report is generated and output.

[0011] In some embodiments of this application, when acquiring multi-source heterogeneous data related to the tender documents, and performing data cleaning and standardization on the multi-source heterogeneous data to generate data to be analyzed, the process includes:

[0012] The acquired multi-source heterogeneous data is cleaned and standardized to generate intermediate data;

[0013] The new entities and relationships generated during the entity recognition process of the large domain model are periodically reviewed and updated to the preset list of bidding domain keywords and entities.

[0014] Based on the updated list of keywords and entities in the bidding field, extract key information fragments from the intermediate data;

[0015] The key information fragments are restructured to generate data to be analyzed.

[0016] In some embodiments of this application, when inputting the data to be analyzed into a pre-trained domain-wide model to construct an initial knowledge graph, the following steps are included:

[0017] The data to be analyzed is input into a pre-trained domain-specific large model;

[0018] The domain-wide model is used for entity recognition and relation extraction, and the output is a standardized triple consisting of entity, relation and attribute.

[0019] The standardized triples are mapped to a graph database, and the initial knowledge graph is constructed using entities as nodes, relations as edges, and attributes as features.

[0020] In some embodiments of this application, when using the aforementioned domain-wide model for entity recognition and relation extraction, and outputting standardized triples consisting of entities, relations, and attributes, the following methods are included:

[0021] The domain-wide model performs entity recognition and relation extraction through multi-step reasoning to generate an initial set of triples.

[0022] Based on the bidding process logic and legal knowledge, the initial set of triples is validated to identify abnormal triples with conflicts.

[0023] For the identified abnormal triples, the abnormal triples are corrected through a conflict resolution subprocess to generate normalized triples.

[0024] In some embodiments of this application, when correcting identified abnormal triples through a conflict resolution subprocess to generate standardized triples, the process includes:

[0025] Tracing back to the original text corresponding to the abnormal triples, we locate all related evidence sentences;

[0026] The semantic importance score of each evidence sentence is calculated using the domain-wide model.

[0027] All evidence sentences are sorted according to the semantic importance score to generate an ordered set of evidence sentences;

[0028] Based on the ordered evidence sentences, the abnormal triples are corrected to generate multiple corrected candidate triples;

[0029] The revised candidate triples are ranked based on the evaluation scores;

[0030] The candidate triplet with the highest evaluation score is selected as the final correction result and used to update the standardized triplet.

[0031] The evaluation score is calculated using the following formula:

[0032]

[0033] Where E is the evaluation score, C is the model confidence score, S is the semantic consistency score, and λ is the coefficient used to balance the weights of model confidence and semantic consistency, with a value range of [0,1].

[0034] In some embodiments of this application, mapping the standardized triples to a graph database, and constructing the initial knowledge graph using entities as nodes, relations as edges, and attributes as nodes, includes:

[0035] The standardized triples are converted into a graph data structure, where each entity corresponds to a node, each relation corresponds to an edge, and each attribute corresponds to the feature value of the node.

[0036] The graph data structure is imported into a graph database, and an index is built in the graph database based on the entity types and relation attributes in the graph data to generate an indexed graph.

[0037] The indexed graph is identified and corrected to output the initial knowledge graph.

[0038] In some embodiments of this application, newly added bidding data and tender data are periodically acquired, preprocessed, and input into the domain-wide model to generate new results, including:

[0039] The newly added bidding data and tender data are cleaned and standardized to generate new intermediate data;

[0040] Based on the updated list of keywords and entities in the bidding field, key information fragments are extracted from the newly added intermediate data;

[0041] The key information fragments are restructured in a structured manner to generate standardized batches of new data;

[0042] Incremental data splitting is performed on the newly added data batches, and the data is allocated to the corresponding extraction mode according to the data characteristics;

[0043] The split data is input into the domain big model, and entity recognition and relation extraction are performed to match the allocation pattern to generate an initial set of triples;

[0044] The initial set of triples is subjected to logical consistency verification and conflict resolution, and a standardized set of triples consisting of the newly added entities and relations is output as the new result.

[0045] In some embodiments of this application, when performing incremental data diversion processing on the newly added data batch and allocating it to the corresponding extraction mode according to data characteristics, the process includes:

[0046] Calculate the semantic similarity between the newly added data batch and the existing items in the initial knowledge graph to obtain an overall similarity score;

[0047] The overall similarity score is compared with a preset threshold, and the items are divided into high similarity items or low similarity new items based on the comparison result.

[0048] Based on the project type classification results, a first extraction mode is assigned to the highly similar projects, and a second extraction mode is assigned to the low-similarity new projects.

[0049] The first extraction mode is a verification extraction mode based on existing maps, and the second extraction mode is an extraction mode oriented towards unknown domain structures.

[0050] In some embodiments of this application, when the newly added results are integrated into the initial knowledge graph to generate a bidding knowledge graph, the following steps are included:

[0051] Identify new entities in the standardized triple set;

[0052] The newly added entities are aligned with the existing entities in the initial knowledge graph, and entity conflicts are detected and resolved.

[0053] Successfully aligned entities are merged, and newly added unaligned entities are added as new nodes to the initial knowledge graph.

[0054] Based on the newly added relationships, corresponding edge connections are established, and the initial knowledge graph is updated.

[0055] The updated knowledge graph is validated to generate a bidding knowledge graph.

[0056] In some embodiments of this application, when performing multimodal association analysis on the bidding knowledge graph to output a risk pattern set, and generating and outputting a risk report based on the risk pattern set, the process includes:

[0057] Multimodal association analysis was performed on the bidding knowledge graph to identify risk patterns.

[0058] The identified risk patterns are aggregated and deduplicated to generate a risk pattern set.

[0059] Based on the set of risk patterns, a corresponding risk report is generated and output.

[0060] Compared with existing technologies, the method for constructing and analyzing a knowledge graph of tender documents based on a large model in this application has the following advantages:

[0061] This invention employs a pre-trained domain-specific large-scale model to perform deep semantic understanding and multi-step reasoning on multi-source heterogeneous bidding document data, achieving high-precision entity recognition and relation extraction. Furthermore, based on bidding process logic and regulatory knowledge, the generated triples are validated and conflict resolved, effectively improving the accuracy and reliability of knowledge graph construction. Simultaneously, by periodically acquiring new bidding data and performing incremental fusion, the knowledge graph possesses the ability to continuously evolve and self-update, dynamically reflecting the latest entities, relationships, and risk patterns in the bidding market, providing a solid and up-to-date data foundation for subsequent analysis.

[0062] Furthermore, by performing multimodal association analysis on the constructed bidding knowledge graph, this invention can deeply uncover potential complex risk patterns and generate corresponding risk reports, significantly improving the intelligence level and efficiency of bidding risk identification. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the method for constructing and analyzing a knowledge graph of tender documents based on a large model in a preferred embodiment of this application. Detailed Implementation

[0064] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0065] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0066] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0067] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0068] In some embodiments of this application, a method for constructing and analyzing a knowledge graph of tender documents based on a large model is provided, including:

[0069] Acquire multi-source heterogeneous data related to the tender documents, and perform data cleaning and standardization on the multi-source heterogeneous data to generate data to be analyzed;

[0070] The data to be analyzed is input into a pre-trained domain-specific large model to construct an initial knowledge graph;

[0071] Regularly acquire newly added bidding and tendering data, preprocess the newly added bidding and tendering data and input them into the domain-wide model to generate new results;

[0072] The newly added results are integrated into the initial knowledge graph to generate a bidding knowledge graph.

[0073] The bidding knowledge graph is subjected to multimodal association analysis to output a risk pattern set. Based on the risk pattern set, a risk report is generated and output.

[0074] In this embodiment, the domain big model refers to a pre-trained language model trained with bidding domain data.

[0075] In some embodiments of this application, when acquiring multi-source heterogeneous data related to the tender documents, and performing data cleaning and standardization on the multi-source heterogeneous data to generate data to be analyzed, the process includes:

[0076] The acquired multi-source heterogeneous data is cleaned and standardized to generate intermediate data;

[0077] The new entities and relationships generated during the entity recognition process of the large domain model are periodically reviewed and updated to the preset list of bidding domain keywords and entities.

[0078] Based on the updated list of keywords and entities in the bidding field, extract key information fragments from the intermediate data;

[0079] The key information fragments are restructured to generate data to be analyzed.

[0080] In this embodiment, data cleaning and standardization aim to eliminate noise, inconsistencies, and redundant information in the data. Specifically, this includes, but is not limited to: removing irrelevant characters, correcting spelling errors, handling missing values, identifying and removing duplicate documents, and uniformly converting the data into a preset standardized format. For example, standardizing date formats, currency units, and company names (both full and abbreviated), and performing preliminary structuring processes such as sentence segmentation and word segmentation on unstructured text paragraphs.

[0081] In this embodiment, the intermediate data has eliminated most of the surface errors and format differences present in the original data, but its information dimensions are still relatively broad.

[0082] In this embodiment, structured reorganization transforms the described text information into quasi-structured data with fixed format and clearly defined fields. A context-aware fragment boundary recognition algorithm is used, which uses the domain big model to calculate the semantic coherence score of the text fragment. Only when the score exceeds a threshold is the fragment identified as a key information fragment.

[0083] In some embodiments of this application, when inputting the data to be analyzed into a pre-trained domain-wide model to construct an initial knowledge graph, the following steps are included:

[0084] The data to be analyzed is input into a pre-trained domain-specific large model;

[0085] The domain-wide model is used for entity recognition and relation extraction, and the output is a standardized triple consisting of entity, relation and attribute.

[0086] The standardized triples are mapped to a graph database, and the initial knowledge graph is constructed using entities as nodes, relations as edges, and attributes as features.

[0087] In this embodiment, the graph database is used to store, manage, and query graph structure data. Each node corresponds to an entity extracted from the tender documents. Each edge corresponds to a relationship between entities.

[0088] In some embodiments of this application, when using the aforementioned domain-wide model for entity recognition and relation extraction, and outputting standardized triples consisting of entities, relations, and attributes, the following methods are included:

[0089] The domain-wide model performs entity recognition and relation extraction through multi-step reasoning to generate an initial set of triples.

[0090] Based on the bidding process logic and legal knowledge, the initial set of triples is validated to identify abnormal triples with conflicts.

[0091] For the identified abnormal triples, the abnormal triples are corrected through a conflict resolution subprocess to generate normalized triples.

[0092] In this embodiment, the domain-wide model performs entity identification and relationship extraction through multi-step reasoning, including: Step 1: The model first reads through the data to be analyzed, performing preliminary entity identification and relationship hypotheses. It identifies all company names, personnel, qualifications, project names, etc., appearing in the text as candidate entities and preliminarily determines which entities may have potential relationships. Step 2: Based on the candidate entities and relationship hypotheses generated in Step 1, the model performs deeper contextual analysis. For example, for the candidate relationship "Company A - Winning Bid - XX Project," the model will backtrack to the original text to find evidence supporting or opposing the relationship. The correctness of the relationship is verified by analyzing the context sentences. Step 3: Before finally outputting the triples, the model performs a logical consistency check between all extracted information fragments. For example, it may find that the same bidding manager is simultaneously associated with two competing companies, which is logically conflicting. Based on its understanding of the conflict of interest rules in the bidding domain, the model will prioritize relationships with higher confidence, or mark them as items to be verified if the conflict cannot be resolved.

[0093] In this embodiment, the bidding process logic and regulatory knowledge include: the uniqueness rule, meaning that for a specific bid section / package number, there can only be one successful bidder at any given time; the timeliness rule, meaning that the date of submission of bid documents must be before the bid opening date; the compliance rule, meaning that bidders must possess the specific qualifications required by the bidding documents; and the logical consistency rule, meaning that a company cannot simultaneously act as a bidder and a bidding agent.

[0094] In this embodiment, the verification process involves the system applying the aforementioned rules to the initial set of triples, performing logical reasoning and conflict detection between triples and between triples and rules. For example, the initial set of triples may contain (Company A, Winning Bid, Project P) and (Company B, Winning Bid, Project P), and both triples point to the same bid segment within the same time period. In this case, the system will trigger a uniqueness rule, marking both conflicting triples as anomalous triples. In some embodiments of this application, when generating standardized triples by correcting the identified anomalous triples through a conflict resolution subprocess, the process includes:

[0095] Tracing back to the original text corresponding to the abnormal triples, we locate all related evidence sentences;

[0096] The semantic importance score of each evidence sentence is calculated using the domain-wide model.

[0097] All evidence sentences are sorted according to the semantic importance score to generate an ordered set of evidence sentences;

[0098] Based on the ordered evidence sentences, the abnormal triples are corrected to generate multiple corrected candidate triples;

[0099] The revised candidate triples are ranked based on the evaluation scores;

[0100] The candidate triplet with the highest evaluation score is selected as the final correction result and used to update the standardized triplet.

[0101] The evaluation score is calculated using the following formula:

[0102]

[0103] Where E is the evaluation score, C is the model confidence score, S is the semantic consistency score, and λ is the coefficient used to balance the weights of model confidence and semantic consistency, with a value range of [0,1].

[0104] In this embodiment, when calculating the semantic importance score, the model uses the evidence sentence to be evaluated and the anomalous triples as joint inputs to quantitatively analyze the performance of the evidence sentence in four key dimensions: first, relevance, i.e., whether the evidence sentence explicitly mentions all the core entities and relationships in the triples; second, certainty, i.e., whether the context used in the sentence is decisive or speculative; third, information completeness, i.e. whether the sentence itself constitutes a logically complete statement rather than a fragmented mention; and fourth, contextual authority, i.e. whether the sentence comes from a chapter with decisive force or a non-core narrative part. The model ultimately synthesizes the analysis of these four dimensions to produce a normalized semantic importance score between 0 and 1.

[0105] In this embodiment, when generating the corrected candidate triplet, the following steps are taken: when evidence indicates that project P is divided into bid section one and bid section two, the model will decompose and refine the original triplet (company A, winning bid, project P) into two candidates: (company A, winning bid, project P-bid section one) and (company B, winning bid, project P-bid section two); ambiguous cooperative relationships will be corrected to more precise "consortium winning bid" or "subcontracting" based on the evidence; and logical conflicts will be resolved by supplementing attributes such as time and status.

[0106] In this embodiment, the C-model confidence score refers to the deterministic probability assessment of the output result by the pre-trained model that generated the corrected candidate triplet. This value is directly derived from the model's internal calculations and is typically a value between 0 and 1. It reflects the model's inherent grasp of the current correction scheme in terms of syntax, semantics, and domain common sense, based on the knowledge and parameters it has acquired during training. A higher C-value indicates a more accurate correction result from the model.

[0107] In this embodiment, the S semantic consistency score refers to the degree of semantic matching and agreement between the revised candidate triple and the associated evidence sentence located by backtracking. This score is obtained by calculating the similarity between the text embedding vectors of the candidate triple and the evidence sentence. It quantifies the strength of external evidence support between the revised scheme and the original text context. A higher S value indicates that the revised scheme is supported by more original text evidence.

[0108] In this embodiment, the weight coefficient λ ranges from [0,1]. It determines the relative importance of the model's internal confidence C and external semantic evidence S in the comprehensive evaluation. When λ approaches 1, the evaluation score E mainly depends on C. This indicates that the system relies more on the judgment ability of the large domain model itself, suitable for scenarios where the model is well-trained and the domain knowledge coverage is comprehensive. When λ approaches 0, the evaluation score E mainly depends on S, indicating that the system emphasizes the original text, suitable for scenarios with high data noise or where strict adherence to the original text is required. By adjusting λ, system administrators can flexibly customize conflict resolution strategies to adapt to the characteristics and quality requirements of different bidding data sources, achieving configurability and adaptability of the strategy.

[0109] In some embodiments of this application, mapping the standardized triples to a graph database, and constructing the initial knowledge graph using entities as nodes, relations as edges, and attributes as nodes, includes:

[0110] The standardized triples are converted into a graph data structure, where each entity corresponds to a node, each relation corresponds to an edge, and each attribute corresponds to the feature value of the node.

[0111] The graph data structure is imported into a graph database, and an index is built in the graph database based on the entity types and relation attributes in the graph data to generate an indexed graph.

[0112] The indexed graph is identified and corrected to output the initial knowledge graph.

[0113] In this embodiment, building an index in the graph database to generate an indexed graph includes: the system creating various types of indexes in the graph database based on the entity type of nodes and the relational attributes of edges in the graph data, forming an indexed graph. This mainly includes: Entity type index: creating indexes for different types of entity nodes; Relationship type index: creating indexes for different types of relational edges; Node attribute index: creating indexes for frequently queried node attributes; Full-text index: establishing full-text indexes for text-based attributes, supporting keyword-based fuzzy search and semantic search. The indexed graph enables the retrieval of massive amounts of bidding and tendering knowledge, providing important support for subsequent complex multi-hop association analysis and real-time risk detection.

[0114] In this embodiment, the identification and correction of the indexed graph includes: the system scans the indexed graph to identify potential structural and logical defects. The main identification targets include: isolated nodes: nodes without any edge connections; redundant nodes: multiple nodes representing the same real-world entity; relationship loops: abnormal logical cycles; and attribute anomalies: detecting attribute values ​​that do not conform to business logic. For the identified problems, the system executes an automatic correction strategy. For isolated nodes: the system associates them with core entities in the graph. If no association can be found within a certain threshold, it is marked as low-confidence data and submitted for manual review or temporarily stored for subsequent data supplementation. For redundant nodes: an entity alignment process is initiated, comparing attributes such as name similarity and unified social credit code to confirm that Company A and Company A (head office) are the same entity, then merging them into one node and integrating all associated relationships and attributes. For relationship loops: the loop is automatically interrupted, and based on metadata such as relationship confidence and timestamps, the most reliable relationship is retained, while potentially erroneous relationships are marked as abnormal and alerts are issued. For attribute anomalies: automatically correct them according to predefined business rules or set them to null values ​​and mark them for verification.

[0115] In some embodiments of this application, newly added bidding data and tender data are periodically acquired, preprocessed, and input into the domain-wide model to generate new results, including:

[0116] The newly added bidding data and tender data are cleaned and standardized to generate new intermediate data;

[0117] Based on the updated list of keywords and entities in the bidding field, key information fragments are extracted from the newly added intermediate data;

[0118] The key information fragments are restructured in a structured manner to generate standardized batches of new data;

[0119] Incremental data splitting is performed on the newly added data batches, and the data is allocated to the corresponding extraction mode according to the data characteristics;

[0120] The split data is input into the domain big model, and entity recognition and relation extraction are performed to match the allocation pattern to generate an initial set of triples;

[0121] The initial set of triples is subjected to logical consistency verification and conflict resolution, and a standardized set of triples consisting of the newly added entities and relations is output as the new result.

[0122] In this embodiment, the key information fragments extracted from the newly added intermediate data include, but are not limited to, the following categories: Subject declaration fragments: information about the parties involved in the bidding process. Object definition fragments: core elements of the bidding project itself. Qualification and requirements fragments: clearly defined entry conditions and standards. Process and event fragments: descriptions of key activities and results in the bidding process. Relationship statement fragments: descriptions that directly or indirectly indicate the relationships between entities.

[0123] In this embodiment, after generating the initial set of triples, a rigorous logical consistency check and conflict resolution are performed. The logical consistency check includes: internal consistency check: checking for contradictions within the newly added triple set; external consistency check: comparing the new triples with the existing initial knowledge graph to check for conflicts; and compliance check: reviewing the data according to bidding regulations. Conflict resolution is a corrective measure taken for abnormal triples identified by the logical consistency check.

[0124] In some embodiments of this application, when performing incremental data diversion processing on the newly added data batch and allocating it to the corresponding extraction mode according to data characteristics, the process includes:

[0125] Calculate the semantic similarity between the newly added data batch and the existing items in the initial knowledge graph to obtain an overall similarity score;

[0126] The overall similarity score is compared with a preset threshold, and the items are divided into high similarity items or low similarity new items based on the comparison result.

[0127] Based on the project type classification results, a first extraction mode is assigned to the highly similar projects, and a second extraction mode is assigned to the low-similarity new projects.

[0128] The first extraction mode is a verification extraction mode based on existing maps, and the second extraction mode is an extraction mode oriented towards unknown domain structures.

[0129] In this embodiment, the calculation of semantic similarity to obtain an overall similarity score is implemented as follows: The system represents the key information of newly added data batches and existing projects in the initial knowledge graph. This typically includes textual and semantic information such as project name, core content description, key qualification requirements, and the main entity types involved. The domain big model encodes the above information into semantic vectors. The semantic association strength is quantified by calculating the cosine similarity between the newly added data vectors and the existing project vector set. Finally, the system integrates the comparison results from multiple dimensions and outputs a normalized overall similarity score. This score objectively reflects the degree of conformity between the new projects and historical projects in terms of business domain, technical requirements, and process structure.

[0130] In this embodiment, a preset threshold is used to determine the project type when performing incremental data diversion. Its setting is based on a trade-off between historical data distribution and business needs. If the overall similarity score is greater than or equal to the preset threshold, the project is determined to be a high-similarity project. This indicates that the new data shares many known patterns and entities with the existing knowledge graph in terms of subjects, objects, and processes, and the existing knowledge structure can be reused. If the overall similarity score is less than the preset threshold, the project is determined to be a low-similarity new project. This indicates that the new data may belong to a new field, a new business model, or have an unprecedentedly complex structure, requiring the system to adopt a more exploratory processing strategy.

[0131] In this embodiment, the first extraction mode is a verification extraction mode based on existing knowledge graphs. In this mode, the system prioritizes using existing project structures of the corresponding type in the initial knowledge graph as reference templates. The model focuses on verifying whether the expected entities and relationships appear in the newly added data and quickly fills in specific attribute values. At the same time, it pays special attention to anomalies that are inconsistent with the template; these anomalies may be risks or innovative points that require key verification.

[0132] In this embodiment, the second extraction mode, oriented towards unknown domain structures, aims to discover and define new entity types and relationships, and construct new knowledge subgraph structures. This mode ensures the system's adaptability to new business formats and models in the market, and continuously expands and enriches the boundaries and connotations of the knowledge graph.

[0133] In some embodiments of this application, when the newly added results are integrated into the initial knowledge graph to generate a bidding knowledge graph, the following steps are included:

[0134] Identify new entities in the standardized triple set;

[0135] The newly added entities are aligned with the existing entities in the initial knowledge graph, and entity conflicts are detected and resolved.

[0136] Successfully aligned entities are merged, and newly added unaligned entities are added as new nodes to the initial knowledge graph.

[0137] Based on the newly added relationships, corresponding edge connections are established, and the initial knowledge graph is updated.

[0138] The updated knowledge graph is validated to generate a bidding knowledge graph.

[0139] In this embodiment, the detection and resolution of entity conflicts includes: attribute conflict resolution, where when the same entity has conflicting attribute values ​​in different data sources, the attribute value from the data source with higher authority is prioritized; if the authority is the same, the attribute value with the latest timestamp is used. Relationship conflict resolution, where when there is a logical conflict between "Company A is a subsidiary of Company B" and "Company B is a subsidiary of Company A", the original text evidence sentences are traced back; if the confidence level of evidence sentence A is higher than that of evidence sentence B, then relationship A is adopted and relationship B is discarded. Logical conflict resolution, where when it is identified that individual P is simultaneously a bidder A and a bid evaluation expert, a bid-rigging risk marker is triggered, and a manual review process is initiated.

[0140] In some embodiments of this application, when performing multimodal association analysis on the bidding knowledge graph to output a risk pattern set, and generating and outputting a risk report based on the risk pattern set, the process includes:

[0141] Multimodal association analysis was performed on the bidding knowledge graph to identify risk patterns.

[0142] The identified risk patterns are aggregated and deduplicated to generate a risk pattern set.

[0143] Based on the set of risk patterns, a corresponding risk report is generated and output.

[0144] In this embodiment, the multimodal association analysis specifically includes the following analysis modalities: Topological structure association analysis: analyzing the network topological characteristics of the knowledge graph; Semantic relationship association analysis: performing deep reasoning based on the semantics of relationships; Temporal dynamic association analysis: using time as a key dimension to analyze the evolution patterns of entities and behaviors over time; Attribute feature association analysis: establishing associations based on the numerical or categorical attributes of nodes. By integrating these multiple analysis modalities, the system can perform multidimensional association queries and reasoning, thereby identifying complex risk patterns that are difficult to discover through single-dimensional analysis.

[0145] In this embodiment, risk patterns include, but are not limited to: **Bid rigging / collusion:** The graph shows hidden equity or personnel connections among multiple bidders, or their bids are generated from the same IP address, or their bids exhibit regular differences. **Qualification affiliation / illegal subcontracting:** The graph reveals that the winning bidder C's qualifications do not meet the project requirements, but it has unusually frequent financial transactions or personnel secondment relationships with the qualified company D, suggesting possible qualification affiliation. Or, after winning the bid, it immediately subcontracts all the main and key work. **Evaluator bias risk:** A certain evaluation committee member E and the winning company F have a stable evaluation-winning association in multiple historical projects, far exceeding statistical randomness. **Historical violation entity recurrence:** Company G, which has been administratively penalized or blacklisted, re-participates in bidding by registering a new affiliated company H, and is identified by the system through entity alignment and association analysis. **Abnormally low bid winning:** The winning bid price is significantly lower than the average budget price and lower than all other bid prices, while the winning bidder's registered capital is severely mismatched with the project scale.

[0146] In this embodiment, the process of generating and outputting a risk report based on a set of risk patterns includes: Risk Title and Type: Clearly identifying the nature of the risk. Risk Level Assessment: Classifying the risk according to its confidence level and potential impact. Core Entities Involved: Listing all key companies, personnel, and projects involved in the pattern. Evidence Chain and Association Path: Displaying the association path derived from the knowledge graph in a combined text and image format, and providing evidence sentences tracing back to the original text. Analysis and Recommendations: Providing a brief analysis based on domain knowledge, and possibly offering preliminary processing suggestions. Multiple Output Formats: The generated risk report can be output in various formats according to requirements.

[0147] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application.

Claims

1. A method for constructing and analyzing a knowledge graph of tender documents based on a large model, characterized in that: include: Acquire multi-source heterogeneous data related to the tender documents, and perform data cleaning and standardization on the multi-source heterogeneous data to generate data to be analyzed; The data to be analyzed is input into a pre-trained domain-specific large model to construct an initial knowledge graph; Regularly acquire newly added bidding and tendering data, preprocess the newly added bidding and tendering data and input them into the domain-wide model to generate new results; The newly added results are integrated into the initial knowledge graph to generate a bidding knowledge graph. The bidding knowledge graph is subjected to multimodal association analysis to output a risk pattern set. Based on the risk pattern set, a risk report is generated and output.

2. The method for constructing and analyzing a knowledge graph of tender documents based on a large model as described in claim 1, characterized in that, When acquiring multi-source heterogeneous data related to the tender documents, and performing data cleaning and standardization on the multi-source heterogeneous data to generate data to be analyzed, the process includes: The acquired multi-source heterogeneous data is cleaned and standardized to generate intermediate data; The new entities and relationships generated during the entity recognition process of the large domain model are periodically reviewed and updated to the preset list of bidding domain keywords and entities. Based on the updated list of keywords and entities in the bidding field, extract key information fragments from the intermediate data; The key information fragments are restructured to generate data to be analyzed.

3. The method for constructing and analyzing a knowledge graph of tender documents based on a large model according to claim 2, characterized in that, When inputting the data to be analyzed into a pre-trained domain-specific large model to construct the initial knowledge graph, the following steps are included: The data to be analyzed is input into a pre-trained domain-specific large model; The domain-wide model is used for entity recognition and relation extraction, and the output is a standardized triple consisting of entity, relation and attribute. The standardized triples are mapped to a graph database, and the initial knowledge graph is constructed using entities as nodes, relations as edges, and attributes as features.

4. The method for constructing and analyzing a knowledge graph of tender documents based on a large model as described in claim 3, characterized in that, When using the aforementioned domain-wide model for entity recognition and relation extraction, and outputting standardized triples consisting of entities, relations, and attributes, the process includes: The domain-wide model performs entity recognition and relation extraction through multi-step reasoning to generate an initial set of triples. Based on the bidding process logic and legal knowledge, the initial set of triples is validated to identify abnormal triples with conflicts. For the identified abnormal triples, the abnormal triples are corrected through a conflict resolution subprocess to generate normalized triples.

5. The method for constructing and analyzing a knowledge graph of tender documents based on a large model according to claim 4, characterized in that, For the identified aberrant triples, the process of correcting the aberrant triples through a conflict resolution subprocess to generate normalized triples includes: Tracing back to the original text corresponding to the abnormal triples, we locate all related evidence sentences; The semantic importance score of each evidence sentence is calculated using the domain-wide model. All evidence sentences are sorted according to the semantic importance score to generate an ordered set of evidence sentences; Based on the ordered evidence sentences, the abnormal triples are corrected to generate multiple corrected candidate triples; The revised candidate triples are ranked based on the evaluation scores; The candidate triplet with the highest evaluation score is selected as the final correction result and used to update the standardized triplet. The evaluation score is calculated using the following formula: ； Where E is the evaluation score, C is the model confidence score, S is the semantic consistency score, and λ is the coefficient used to balance the weights of model confidence and semantic consistency, with a value range of [0,1].

6. The method for constructing and analyzing a knowledge graph of tender documents based on a large model as described in claim 5, characterized in that, Mapping the standardized triples to a graph database, and constructing the initial knowledge graph using entities as nodes, relations as edges, and attributes as features, includes: The standardized triples are converted into a graph data structure, where each entity corresponds to a node, each relation corresponds to an edge, and each attribute corresponds to the feature value of the node. The graph data structure is imported into a graph database, and an index is built in the graph database based on the entity types and relation attributes in the graph data to generate an indexed graph. The indexed graph is identified and corrected to output the initial knowledge graph.

7. The method for constructing and analyzing a knowledge graph of tender documents based on a large model as described in claim 6, characterized in that, Periodically acquire newly added bidding and tendering data, preprocess the newly added bidding and tendering data, and input it into the domain-wide model. When generating new results, the process includes: The newly added bidding data and tender data are cleaned and standardized to generate new intermediate data; Based on the updated list of keywords and entities in the bidding field, key information fragments are extracted from the newly added intermediate data; The key information fragments are restructured in a structured manner to generate standardized batches of new data; Incremental data splitting is performed on the newly added data batches, and the data is allocated to the corresponding extraction mode according to the data characteristics; The split data is input into the domain big model, and entity recognition and relation extraction are performed to match the allocation pattern to generate an initial set of triples; The initial set of triples is subjected to logical consistency verification and conflict resolution, and a standardized set of triples consisting of the newly added entities and relations is output as the new result.

8. The method for constructing and analyzing a knowledge graph of tender documents based on a large model according to claim 7, characterized in that, When performing incremental data diversion processing on the newly added data batches and allocating them to the corresponding extraction modes according to data characteristics, the process includes: Calculate the semantic similarity between the newly added data batch and the existing items in the initial knowledge graph to obtain an overall similarity score; The overall similarity score is compared with a preset threshold, and the items are divided into high similarity items or low similarity new items based on the comparison result. Based on the project type classification results, a first extraction mode is assigned to the highly similar projects, and a second extraction mode is assigned to the low-similarity new projects. The first extraction mode is a verification extraction mode based on existing maps, and the second extraction mode is an extraction mode oriented towards unknown domain structures.

9. The method for constructing and analyzing a knowledge graph of tender documents based on a large model as described in claim 8, characterized in that, When integrating the newly added results into the initial knowledge graph to generate the bidding and tendering knowledge graph, the process includes: Identify new entities in the standardized triple set; The newly added entities are aligned with the existing entities in the initial knowledge graph, and entity conflicts are detected and resolved. Successfully aligned entities are merged, and newly added unaligned entities are added as new nodes to the initial knowledge graph. Based on the newly added relationships, corresponding edge connections are established, and the initial knowledge graph is updated. The updated knowledge graph is validated to generate a bidding knowledge graph.

10. The method for constructing and analyzing a knowledge graph of tender documents based on a large model according to claim 9, characterized in that, When performing multimodal association analysis on the bidding knowledge graph to output a risk pattern set, and generating and outputting a risk report based on the risk pattern set, the process includes: Multimodal association analysis was performed on the bidding knowledge graph to identify risk patterns. The identified risk patterns are aggregated and deduplicated to generate a risk pattern set. Based on the set of risk patterns, a corresponding risk report is generated and output.

Images