A tender document automatic generation method, system, terminal and medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INSPUR GENERSOFT CO LTD
- Filing Date
- 2025-08-26
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies lack the ability to understand multimodal information in the generation of tender documents, cannot effectively process technical drawings and data tables, and have simple matching mechanisms that lack reasoning ability, resulting in tender documents that lack personalization and adaptability.
Multimodal parsing is used to process project requirement information, generating a unified project feature vector. This vector is then matched and reasoned using a knowledge graph of the bidding domain. A dynamic relevance threshold strategy is used to filter entity nodes, and a first draft of the bidding document is generated by combining the domain-wide model.
It achieves a comprehensive understanding of project requirements, improves the accuracy and applicability of generated tender documents, ensures precise matching of core clauses and intelligent recommendation of auxiliary clauses, provides in-depth knowledge support, and enhances the reliability of documents.
Smart Images
Figure CN121031791B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, specifically to a method, system, terminal, and medium for automatically generating tender documents. Background Technology
[0002] In the field of tender document preparation, the traditional method relies entirely on manual processing. Preparers need to manually review documents, draft each tender clause according to project requirements, and then format and proofread the content. This manual approach is not only time-consuming and labor-intensive, but also prone to inconsistent document quality due to differences in the professional skills of the personnel.
[0003] Current automation technologies mainly employ simple template-filling methods. While these methods can improve efficiency to some extent, they still have significant drawbacks: First, existing technologies lack the ability to understand multimodal information and cannot effectively process non-textual information such as technical drawings and data tables in project requirements, resulting in insufficient information utilization. Second, the matching mechanisms of existing methods are too simplistic, typically using fixed thresholds or simple rules for clause matching, and cannot be dynamically adjusted according to project characteristics and clause importance. Finally, existing systems lack good reasoning capabilities and cannot discover potential relationships from massive amounts of historical data, resulting in generated tender documents that lack personalization and adaptability. Summary of the Invention
[0004] To address the aforementioned issues, this invention provides a method, system, terminal, and medium for automatically generating tender documents, thereby improving the quality and efficiency of tender document generation.
[0005] In a first aspect, the technical solution of the present invention provides a method for automatically generating tender documents, comprising the following steps: Receive project requirement information input by the user, wherein the project requirement information includes one or more of structured text, unstructured text, technical drawings, and data tables; The project requirement information is processed using multimodal parsing to generate a unified project feature vector; The project feature vector is matched and reasoned with a knowledge graph in the bidding domain. The knowledge graph is constructed by extracting entities, relationships, and attributes from a rule base, a historical bidding document database, and a technical standard database. The matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes that are relevant to the current project requirements. The domain big model is invoked to receive the project feature vector and the reasoning results from the knowledge graph to generate a draft tender document, including matching the tender document template, drafting the tender terms, and attaching traceability information to each generated target term.
[0006] In an optional implementation, the project feature vector is matched and reasoned with a knowledge graph of the bidding domain, specifically including: The entity nodes and relation edges in the knowledge graph of the bidding field are vectorized to construct a graph embedding vector space; the entity nodes include relevant clauses, technical standard parameters, historical bidding projects, and qualification requirements; the relation edges are used to connect entity nodes with logical relationships, and their relation types include "belong to", "constraint", "conflict", "reference", and "contain". Projecting the project feature vectors into the graph embedding vector space yields their vector representation in that space. In the graph embedding vector space, the cosine similarity between the vector representation of the project feature vector and the vector of each entity node is calculated. Based on a preset similarity threshold, the Top-K entity nodes most relevant to the current project requirements are selected as the initial matching node set; Starting from the initial matching node set, perform multi-hop traversal and reasoning on the knowledge graph to discover new node sets; The initial set of matched nodes is integrated with the new set of nodes discovered through association reasoning to form the final set of matching and reasoning results.
[0007] In an optional implementation, starting from the initial matched node set, multi-hop traversal and reasoning are performed on the knowledge graph to discover new node sets, specifically including: Starting from the initial set of matched nodes, explore its first-order neighbors to obtain the set of first-order neighbor nodes; The second graph neural network is used to learn an embedding representation containing graph structure information for each neighbor node in the first-order neighbor node set. This includes obtaining the embedding vector of the neighbor node in the l-th layer through the message passing mechanism of the L-layer second graph neural network, and obtaining the final embedding vector of the neighbor node after L-layer iteration. The overall relevance score between the neighboring nodes and the initial matching node set is calculated using the final embedding vectors of the neighboring nodes. The dynamic relevance threshold of neighboring nodes is calculated based on the complexity of the user's initial query and the importance of neighboring nodes in the knowledge graph. If the overall correlation score is greater than the calculated dynamic correlation threshold, then the corresponding neighbor node is a newly discovered node, and all the newly discovered nodes constitute a new node set.
[0008] In an optional implementation, a dynamic relevance threshold is calculated based on the complexity of the user's initial query and the importance of the node in the knowledge graph, specifically including: The embedding vectors of all nodes in the initial matching node set are extracted using the second graph neural network. The average semantic dispersion of the initial matching node set is then calculated as a query complexity score. ,
[0009] in, To initially match the number of nodes in the node set. , These are the nodes in the initial matching node set. , The embedding vector; Query complexity score It is converted into a basic dynamic threshold through a linear mapping function.
[0010] in, , The upper and lower limits are preset thresholds. and This is the normalized boundary for the query complexity score obtained based on historical data statistics; Read neighbor nodes from the pre-computed node authority table. Authority score ; Calculate neighbor nodes using the following formula. Authority adjustment factor
[0011] in, This is the preset authority influence coefficient; Calculate neighbor nodes using the following formula. The final dynamic correlation threshold , .
[0012] In one optional implementation, the project requirement information is subjected to multimodal parsing to generate a unified project feature vector, including: For unstructured text, a BERT model finely tuned with bidding domain corpus is used for semantic analysis and key information extraction to obtain text feature vectors; Visual features are extracted from the technical drawings using a visual encoder, and the first image neural network is used to identify the components and relationships in the drawings to obtain the drawing feature vector; The data table is analyzed using a table recognition model, and its content is converted into structured data in key-value pair form to obtain table feature vectors. The text feature vector, drawing feature vector, and table feature vector are fused together to generate a unified project feature vector.
[0013] In an optional implementation, the text feature vector, drawing feature vector, and table feature vector are fused to generate the project feature vector, specifically including: The extracted text feature vectors, drawing feature vectors, and table feature vectors are respectively input into a modality-specific encoder for high-dimensional mapping and feature alignment; The pre-aligned feature vectors are input into a fusion module based on gated cross-attention to achieve intermodal interaction and information complementarity, resulting in enhanced text features. The enhanced text features, the feature-aligned drawing feature vectors, and the feature-aligned table feature vectors are concatenated to generate a joint representation; The joint representation is passed through a residual projection layer and normalized via a LayerNorm layer to output the final unified project feature vector.
[0014] In an optional implementation, the method further includes the following steps: The knowledge graph for the bidding field is dynamically updated based on relevant change information.
[0015] Secondly, the technical solution of the present invention provides an automatic tender document generation system, comprising: The project requirement information receiving module is used to receive project requirement information input by the user. The project requirement information includes one or more of the following: structured text, unstructured text, technical drawings, and data tables. The project feature vector generation module is used to perform multimodal parsing processing on the project requirement information to generate a unified project feature vector; The knowledge graph reasoning module is used to match and reason with the project feature vector and the bidding domain knowledge graph. The bidding domain knowledge graph is constructed by extracting entities, relationships and attributes from the rule base, historical bidding document base and technical standard base. The matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes that are relevant to the current project requirements. The initial draft tender document generation module is used to call the domain big model to receive the project feature vector and the reasoning results from the knowledge graph to generate an initial draft tender document, including matching the tender document template, writing the tender terms, and attaching traceability information to each generated target term.
[0016] Thirdly, the technical solution of the present invention provides a terminal, comprising: The memory is used to store the automatic tender document generation program; A processor is configured to implement the steps of the automatic tender document generation method as described in any of the preceding embodiments when executing the automatic tender document generation program.
[0017] Fourthly, the present invention provides a computer-readable storage medium storing a tender document automatic generation program, wherein the tender document automatic generation program, when executed by a processor, implements the steps of the tender document automatic generation method as described in any of the above claims.
[0018] As can be seen from the above technical solutions, this application has the following advantages: 1. This application utilizes multimodal information parsing and fusion technology to simultaneously understand project requirement information in text, drawings, and tables, generating a unified project feature vector. This achieves a comprehensive and accurate understanding of project requirements, avoiding cognitive biases caused by information fragmentation. 2. This application adopts a dynamic relevance threshold strategy based on query complexity and node importance, which overcomes the limitations of existing technologies that use fixed thresholds or simple rules for matching. It can intelligently perceive the complexity of project requirements, automatically distinguish the importance level of knowledge clauses, and dynamically adjust the recommendation criteria. This not only ensures the accurate matching of core clauses, but also realizes the intelligent recommendation of auxiliary clauses, thereby improving the accuracy and applicability of the recommendations. 3. This application constructs a knowledge graph in the bidding field and adopts a multi-hop traversal reasoning mechanism, which can mine potential relationships from massive historical bidding documents and technical standards, and discover those clauses that are indirectly related to project requirements but are important, thus providing in-depth knowledge support for the generation of bidding documents and improving the reliability of the generated bidding documents. Attached Figure Description
[0019] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a schematic diagram of a method for automatically generating tender documents according to an embodiment of the present invention.
[0021] Figure 2 This is a schematic block diagram of an automatic tender document generation system provided in an embodiment of the present invention.
[0022] Figure 3 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present invention. Detailed Implementation
[0023] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0024] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.
[0025] Figure 1 This is a schematic flowchart illustrating a method for automatically generating tender documents according to an embodiment of the present invention. Figure 1 The executing entity can be an automated tender document generation system. The automated tender document generation method provided in this embodiment of the invention is executed by a computer device; correspondingly, the automated tender document generation system runs on the computer device. Depending on different needs, the order of the steps in this flowchart can be changed, and some steps can be omitted.
[0026] like Figure 1 As shown, the method includes the following steps.
[0027] S1, receive project requirement information input by the user, the project requirement information including one or more of structured text, unstructured text, technical drawings and data tables.
[0028] S2, perform multimodal parsing processing on the project requirement information to generate a unified project feature vector.
[0029] S3, matching and reasoning the project feature vector with the bidding domain knowledge graph, which is constructed by extracting entities, relationships and attributes from the rule base, historical bidding document base and technical standard base; the matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes related to the current project requirements.
[0030] S4, invoke the domain big model to receive the project feature vector and the reasoning results from the knowledge graph to generate a draft tender document, including matching the tender document template, writing the tender terms, and attaching traceability information to each generated target term.
[0031] As a refinement and extension of the specific implementation of the above embodiments, in order to fully explain the specific implementation process of this embodiment, the following will provide possible embodiments to describe the specific implementation of the above steps in a non-limiting manner.
[0032] In some optional implementations, step S2 performs multimodal parsing processing on the project requirement information to generate a unified project feature vector, specifically including the following steps.
[0033] S2.1. The BERT model, finely tuned to the bidding domain corpus, is used to perform semantic analysis and extract key information from unstructured text to obtain text feature vectors.
[0034] S2.2, Visual features are extracted from the technical drawings using a visual encoder, and the first neural network is used to identify the components and relationships in the drawings to obtain the drawing feature vector.
[0035] S2.3, use a table recognition model to parse the structure of the data table and convert its content into structured data in key-value pair form to obtain the table feature vector.
[0036] S2.4, the text feature vector, drawing feature vector and table feature vector are fused to generate a unified project feature vector.
[0037] S2.4.1 The extracted text feature vector, drawing feature vector, and table feature vector are input into the modality-specific encoder for high-dimensional mapping and feature alignment.
[0038] The modality-specific encoder consists of three parallel sub-networks: The text encoding subnetwork consists of a fully connected layer and a LayerNorm layer, which will... Mapped to ; The drawing encoding subnetwork consists of a fully connected layer and a LayerNorm layer, which will... Mapped to ; The table encoding subnetwork consists of a fully connected layer and a LayerNorm layer, which will... Mapped to .
[0039] The three subnetworks have the same output dimension, achieving initial alignment of different modal features in the same vector space.
[0040] S2.4.2, the pre-aligned feature vectors are input into the fusion module based on gated cross-attention to achieve intermodal interaction and information complementarity to obtain enhanced text features.
[0041] First, a cross-attention graph is constructed, with the text modality as the query and the drawing and table modalities as the key and value, respectively. Then, the text-drawing attention is calculated. With text-table attention :
[0042]
[0043] in, , , For a trainable projection matrix, This is the scaling factor.
[0044] Then, gating information filtering is performed, equipping each cross-attention output with a learnable gating unit to adaptively filter complementary information from other modalities:
[0045]
[0046]
[0047] in, The sigmoid activation function is used, and [;] indicates vector concatenation. This represents element-wise multiplication. Text features enhanced by drawing and table information.
[0048] S2.4.3 concatenates the enhanced text features, the feature-aligned drawing feature vector, and the feature-aligned table feature vector to generate a joint representation.
[0049] The compressor consists of a bidirectional LSTM network or a single Transformer Encoder layer, used to capture global dependencies across modalities and generate a compact, modality-independent joint representation.
[0050] S2.4.4 passes the joint representation through the residual projection layer and normalizes it through the LayerNorm layer to output the final unified item feature vector.
[0051] By fusing multimodal features, high-dimensional mapping and alignment of heterogeneous modal features are achieved through parallel sub-networks, laying the foundation for subsequent deep interaction. This solves the problem of inconsistent original feature distributions. Combining text-based cross-attention with gating units empowers the model to dynamically perceive and filter complementary information: gating coefficient g. d and g bThe size of the data intuitively reflects the contribution of drawing and table information to the text description, achieving selective and emphasized fusion rather than blindly piling up all information. Further extraction of global contextual information from the concatenated features using LSTM or Transformer generates a highly abstract joint representation that transcends a single modality, making it more conducive to semantic matching with knowledge graphs. Finally, residual connections and LayerNorm are used to ensure training stability and the normalization of output features.
[0052] In some optional implementations, step S3 involves matching and reasoning the project feature vector with the bidding domain knowledge graph, specifically including the following steps.
[0053] S3.1, vectorize the entity nodes and relation edges in the knowledge graph of the bidding domain, and construct the graph embedding vector space.
[0054] Entity nodes include relevant terms, technical standard parameters, historical bidding projects, and qualification requirements; relationship edges are used to connect logically related entity nodes, and their relationship types include "belongs to", "constraint", "conflict", "reference", and "contains". S3.2 Project the project feature vector into the graph embedding vector space to obtain its vector representation in the space.
[0055] S3.3, In the graph embedding vector space, calculate the cosine similarity between the vector representation of the project feature vector and the vector of each entity node.
[0056] S3.4, Based on the preset similarity threshold, select the Top-K entity nodes that are most relevant to the current project requirements as the initial matching node set.
[0057] The initial matching node set includes: usage clause nodes, referenceable historical tender document template nodes, and relevant technical standard parameter nodes.
[0058] S3.5, Starting from the initial matching node set, perform multi-hop traversal and reasoning on the knowledge graph to discover a new node set.
[0059] S3.5.1, starting from the initial matching node set, explore its first-order neighbors to obtain the first-order neighbor node set.
[0060] S3.5.2, learns an embedding representation containing graph structure information for each neighbor node in the first-order neighbor node set through a second graph neural network, including obtaining the embedding vector of the neighbor node in the l-th layer through the message passing mechanism of the L-layer second graph neural network, and obtaining the final embedding vector of the neighbor node after L-layer iteration.
[0061] The neighbor nodes obtained by direct traversal may be numerous and contain a lot of noise. Graph neural networks (such as GraphSAGE) are used to learn embedded representations for nodes that contain graph structure information, and relevance scores are calculated accordingly.
[0062] S3.5.3 Calculate the overall relevance score between the neighboring nodes and the initial matching node set using the final embedding vectors of the neighboring nodes.
[0063] Specifically, the maximum cosine similarity between the neighboring node and the embeddings of all nodes in the initial matching node set is calculated as the overall relevance score, expressed as:
[0064] in, For the initial set of matching nodes, For neighboring nodes, For the initial matching of nodes in the node set Embedded vector, Neighboring nodes The final embedding vector.
[0065] S3.5.4 Calculate the dynamic relevance threshold of neighboring nodes based on the complexity of the user's initial query and the importance of neighboring nodes in the knowledge graph.
[0066] In some optional implementations, the threshold is dynamically adjusted based on the importance of neighboring nodes and the complexity of user queries to improve the accuracy of inference results, specifically including the following steps.
[0067] S3.5.41, the embedding vectors of all nodes in the initial matching node set are extracted using the second graph neural network. The average semantic dispersion of the initial matching node set is then calculated as the query complexity score. ,
[0068] in, To initially match the number of nodes in the node set. , These are the nodes in the initial matching node set. , The embedding vector.
[0069] S3.5.42, query complexity score It is converted into a basic dynamic threshold through a linear mapping function.
[0070] in, , The upper and lower limits are preset thresholds. and This is the normalized boundary of the query complexity score obtained based on historical data statistics.
[0071] S3.5.43, Read neighbor nodes from the pre-calculated node authority table. Authority score ; Calculate neighbor nodes using the following formula. Authority adjustment factor
[0072] in, This is the preset authority influence coefficient.
[0073] Authority scores are pre-calculated and stored using the PageRank algorithm or manual annotation.
[0074] S3.5.44, calculate neighboring nodes using the following formula. The final dynamic correlation threshold , .
[0075] S3.5.5 If the overall correlation score is greater than the calculated dynamic correlation threshold, then the corresponding neighbor node is a newly discovered node, and all the newly discovered nodes constitute a new node set.
[0076] S3.6 Integrate the initial matching node set with the new node set discovered through association reasoning to form the final matching and reasoning result set.
[0077] The results set is structured to include: a recommended tender document template, essential clauses that must be included, suggested technical specifications, and conflicting or high-risk clauses that should be noted and their avoidance suggestions.
[0078] The knowledge graph of the bidding field is dynamically updated based on relevant change information. Specifically, the rule change announcements and industry standard update information of external data sources are captured by the real-time data monitoring engine, and the information extraction process is automatically triggered to update, associate or mark the corresponding nodes in the initial knowledge graph as invalid, thereby forming the dynamically updated knowledge graph of the bidding field.
[0079] A dedicated domain-specific large-scale model service interface is configured. This interface receives project feature vectors and knowledge graph inference result sets from the preprocessing module, initiating the initial draft generation process for the tender document. The domain-specific large-scale model employs a generative pre-trained model based on the Transformer architecture and is obtained through deep fine-tuning training using tender domain corpora, possessing the capability to generate professional tender text. Specifically, based on a selected template framework, the domain-specific large-scale model generates clause content according to the knowledge graph inference results. The generation process uses conditional text generation technology, with the large-scale model as the main generator and project feature vectors and knowledge graph inference results as conditional inputs and control signals. The model automatically replaces general clauses with specific clauses matching the project features. Source information is automatically attached to each generated key clause, embedded in the clause content in the form of metadata. The source information includes: the clause number on which the clause is based, the project number of the referenced historical tender document, the technical standard code and version number used, and the corresponding knowledge graph node ID. The final generated initial draft of the tender document is output in a structured document format, preserving the document hierarchy while embedding semantic tags.
[0080] The above text provides a detailed description of an embodiment of a method for automatically generating tender documents. Based on the above embodiment, this invention also provides a corresponding system for automatically generating tender documents.
[0081] Figure 2 This is a schematic block diagram of an automatic tender document generation system provided by an embodiment of the present invention. In this embodiment, the automatic tender document generation system 200 can be divided into multiple functional modules according to the functions it performs. A module, as referred to in this invention, is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory.
[0082] The project requirement information receiving module 210 is used to receive project requirement information input by the user, which includes one or more of structured text, unstructured text, technical drawings, and data tables.
[0083] The project feature vector generation module 220 is used to perform multimodal parsing processing on the project requirement information to generate a unified project feature vector.
[0084] The knowledge graph reasoning module 230 is used to match and reason with the project feature vector and the bidding domain knowledge graph. The bidding domain knowledge graph is constructed by extracting entities, relationships and attributes from the rule base, historical bidding document base and technical standard base. The matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes related to the current project requirements.
[0085] The initial draft tender document generation module 240 is used to call the domain big model to receive the project feature vector and the reasoning results from the knowledge graph to generate an initial draft tender document, including matching the tender document template, writing the tender terms, and attaching traceability information to each generated target term.
[0086] The automatic tender document generation system of this embodiment is used to implement the aforementioned automatic tender document generation method. Therefore, the specific implementation of this system can be found in the embodiment section of the automatic tender document generation method above. Thus, the specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.
[0087] Furthermore, since the automatic tender document generation system in this embodiment is used to implement the aforementioned automatic tender document generation method, its function corresponds to the function of the above method, and will not be repeated here.
[0088] Figure 3 A schematic diagram of a terminal 300 provided in an embodiment of the present invention includes: a processor 310, a memory 320, and a communication unit 330. The processor 310 is used to implement the following steps when executing the automatic tender document generation program stored in the memory 320: Receive project requirement information input by the user, wherein the project requirement information includes one or more of structured text, unstructured text, technical drawings, and data tables; The project requirement information is processed using multimodal parsing to generate a unified project feature vector; The project feature vector is matched and reasoned with a knowledge graph in the bidding domain. The knowledge graph is constructed by extracting entities, relationships, and attributes from a rule base, a historical bidding document database, and a technical standard database. The matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes that are relevant to the current project requirements. The domain big model is invoked to receive the project feature vector and the reasoning results from the knowledge graph to generate a draft tender document, including matching the tender document template, drafting the tender terms, and attaching traceability information to each generated target term.
[0089] The terminal 300 includes a processor 310, a memory 320, and a communication unit 330. These components communicate via one or more buses. Those skilled in the art will understand that the server structure shown in the figure does not constitute a limitation of the present invention. It can be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.
[0090] The memory 320 can be used to store the execution instructions of the processor 310. The memory 320 can be implemented by any type of volatile or non-volatile memory terminal or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. When the execution instructions in the memory 320 are executed by the processor 310, the terminal 300 is able to perform some or all of the steps in the above method embodiments.
[0091] The processor 310 serves as the control center of the storage terminal, connecting various parts of the electronic terminal via various interfaces and lines. It executes software programs and / or modules stored in the memory 320, and calls data stored in the memory to perform various functions of the electronic terminal and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 310 may consist only of a central processing unit (CPU). In this embodiment of the invention, the CPU may have a single processing core or include multiple processing cores.
[0092] The communication unit 330 is used to establish a communication channel, enabling the storage terminal to communicate with other terminals. It can receive user data sent by other terminals or send user data to other terminals.
[0093] The present invention also provides a computer storage medium, which may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.
[0094] The present invention also provides a computer storage medium, which may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.
[0095] The computer storage medium stores a tender document automatic generation program, which, when executed by the processor, performs the following steps: Receive project requirement information input by the user, wherein the project requirement information includes one or more of structured text, unstructured text, technical drawings, and data tables; The project requirement information is processed using multimodal parsing to generate a unified project feature vector; The project feature vector is matched and reasoned with a knowledge graph in the bidding domain. The knowledge graph is constructed by extracting entities, relationships, and attributes from a rule base, a historical bidding document database, and a technical standard database. The matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes that are relevant to the current project requirements. The domain big model is invoked to receive the project feature vector and the reasoning results from the knowledge graph to generate a draft tender document, including matching the tender document template, drafting the tender terms, and attaching traceability information to each generated target term.
[0096] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or other media capable of storing program code. It includes several instructions to cause a computer terminal (which may be a personal computer, server, or a second terminal, network terminal, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.
[0097] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.
[0098] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0099] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0100] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for automatically generating tender documents, characterized in that, Includes the following steps: Receive project requirement information input by the user, which includes unstructured text, technical drawings, and data tables; The project requirement information is processed using multimodal parsing to generate a unified project feature vector; The project feature vectors are matched and reasoned with a knowledge graph in the bidding domain. The knowledge graph in the bidding domain is constructed by extracting entities, relationships and attributes from a rule base, a historical bidding document base and a technical standard base. The matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes that are relevant to the current project requirements. This includes: selecting the Top-K entity nodes that are most relevant to the current project requirements as the initial matching node set, and calculating the dynamic relevance threshold based on the complexity of the user's initial query and the importance of the nodes in the knowledge graph. The domain big model is invoked to receive the project feature vector and the reasoning results from the knowledge graph to generate a draft tender document, including matching the tender document template, writing the tender terms, and attaching traceability information to each generated target term; The dynamic relevance threshold is calculated based on the complexity of the user's initial query and the importance of the node in the knowledge graph. Specifically, this includes: The embedding vectors of all nodes in the initial matching node set are extracted using the second graph neural network. The average semantic dispersion of the initial matching node set is calculated as the query complexity score. , in, To initially match the number of nodes in the node set. , These are the nodes in the initial matching node set. , The embedding vector; Query complexity score It is converted into a basic dynamic threshold through a linear mapping function. in, , The upper and lower limits are preset thresholds. and This is the normalized boundary for the query complexity score obtained based on historical data statistics; Read neighbor nodes from the pre-computed node authority table. Authority score ; Calculate neighbor nodes using the following formula. Authority adjustment factor in, This is the preset authority influence coefficient; Calculate neighbor nodes using the following formula. The final dynamic correlation threshold , 。 2. The method for automatically generating tender documents according to claim 1, characterized in that, The project feature vectors are matched and reasoned with the knowledge graph of the bidding domain, specifically including: The entity nodes and relation edges in the knowledge graph of the bidding domain are vectorized to construct a graph embedding vector space; the entity nodes include relevant clauses, technical standard parameters, historical bidding projects, and qualification requirements; the relation edges are used to connect entity nodes with logical relationships, and their relation types include belonging, constraint, conflict, reference, and inclusion; Projecting the project feature vectors into the graph embedding vector space yields their vector representation in that space. In the graph embedding vector space, the cosine similarity between the vector representation of the project feature vector and the vector of each entity node is calculated. Based on a preset similarity threshold, the Top-K entity nodes most relevant to the current project requirements are selected as the initial matching node set; Starting from the initial matching node set, perform multi-hop traversal and reasoning on the knowledge graph to discover new node sets; The initial set of matched nodes is integrated with the new set of nodes discovered through association reasoning to form the final set of matching and reasoning results.
3. The method for automatically generating tender documents according to claim 2, characterized in that, Starting with the initial matched node set, multi-hop traversal and reasoning are performed on the knowledge graph to discover new node sets, specifically including: Starting from the initial set of matched nodes, explore its first-order neighbors to obtain the set of first-order neighbor nodes; The second graph neural network is used to learn an embedding representation containing graph structure information for each neighbor node in the first-order neighbor node set. This includes obtaining the embedding vector of the neighbor node in the l-th layer through the message passing mechanism of the L-layer second graph neural network, and obtaining the final embedding vector of the neighbor node after L-layer iteration. The overall relevance score between the neighboring nodes and the initial matching node set is calculated using the final embedding vectors of the neighboring nodes. The dynamic relevance threshold of neighboring nodes is calculated based on the complexity of the user's initial query and the importance of neighboring nodes in the knowledge graph. If the overall correlation score is greater than the calculated dynamic correlation threshold, then the corresponding neighbor node is a newly discovered node, and all the newly discovered nodes constitute a new node set.
4. The method for automatically generating tender documents according to claim 1, characterized in that, The project requirement information is processed using multimodal parsing to generate a unified project feature vector, including: For unstructured text, a BERT model finely tuned with bidding domain corpus is used for semantic analysis and key information extraction to obtain text feature vectors; Visual features are extracted from the technical drawings using a visual encoder, and the first image neural network is used to identify the components and relationships in the drawings to obtain the drawing feature vector; The data table is analyzed using a table recognition model, and its content is converted into structured data in key-value pair form to obtain table feature vectors. The text feature vector, drawing feature vector, and table feature vector are fused together to generate a unified project feature vector.
5. The method for automatically generating tender documents according to claim 4, characterized in that, The text feature vector, drawing feature vector, and table feature vector are fused to generate the project feature vector, specifically including: The extracted text feature vectors, drawing feature vectors, and table feature vectors are respectively input into a modality-specific encoder for high-dimensional mapping and feature alignment; The pre-aligned text feature vectors are input into a fusion module based on gated cross-attention to achieve intermodal interaction and information complementarity, resulting in enhanced text features. The enhanced text features, the feature-aligned drawing feature vectors, and the feature-aligned table feature vectors are concatenated to generate a joint representation; The joint representation is passed through a residual projection layer and normalized via a LayerNorm layer to output the final unified project feature vector.
6. The method for automatically generating tender documents according to claim 1, characterized in that, The method also includes the following steps: The knowledge graph for the bidding field is dynamically updated based on relevant change information.
7. A system for automatically generating tender documents, characterized in that, include: The project requirement information receiving module is used to receive project requirement information input by the user, which includes unstructured text, technical drawings, and data tables. The project feature vector generation module is used to perform multimodal parsing processing on the project requirement information to generate a unified project feature vector; The knowledge graph reasoning module is used to match and reason with the project feature vector and the bidding domain knowledge graph. The bidding domain knowledge graph is constructed by extracting entities, relationships and attributes from the rule base, historical bidding document base and technical standard base. The matching and reasoning process uses a dynamic relevance threshold strategy based on query complexity and node importance to filter out entity nodes that are relevant to the current project requirements. This includes: selecting the Top-K entity nodes that are most relevant to the current project requirements as the initial matching node set, and calculating the dynamic relevance threshold based on the complexity of the user's initial query and the importance of the nodes in the knowledge graph. The initial draft tender document generation module is used to call the domain big model to receive the project feature vector and the reasoning results from the knowledge graph to generate an initial draft tender document, including matching the tender document template, writing tender clauses, and attaching traceability information to each generated target clause; The dynamic relevance threshold is calculated based on the complexity of the user's initial query and the importance of the node in the knowledge graph. Specifically, this includes: The embedding vectors of all nodes in the initial matching node set are extracted using the second graph neural network. The average semantic dispersion of the initial matching node set is calculated as the query complexity score. , in, To initially match the number of nodes in the node set. , These are the nodes in the initial matching node set. , The embedding vector; Query complexity score It is converted into a basic dynamic threshold through a linear mapping function. in, , The upper and lower limits are preset thresholds. and This is the normalized boundary for the query complexity score obtained based on historical data statistics; Read neighbor nodes from the pre-computed node authority table. Authority score ; Calculate neighbor nodes using the following formula. Authority adjustment factor in, This is the preset authority influence coefficient; Calculate neighbor nodes using the following formula. The final dynamic correlation threshold , 。 8. A terminal, characterized in that, include: The memory is used to store the automatic tender document generation program; A processor is configured to implement the steps of the automatic tender document generation method as described in any one of claims 1 to 6 when executing the automatic tender document generation program.
9. A computer-readable storage medium, characterized in that, The readable storage medium stores an automatic tender document generation program, which, when executed by a processor, implements the steps of the automatic tender document generation method as described in any one of claims 1 to 6.