A tender document generation method and system, and a storage medium

Abstract

This invention provides a method, system, and storage medium for generating tender documents. The method includes: acquiring raw data from multiple modalities, wherein the raw data includes project requirement text, design model files, legal and regulatory texts, and historical tender case data; performing multi-module semantic deep analysis and structured processing on the raw data to generate a project requirement sub-graph, a spatial logic sub-graph, a structured legal knowledge base, and a historical case strategy feature base; dynamically fusing the project requirement sub-graph and the spatial logic sub-graph across modalities to construct a unified knowledge graph, and performing dynamic knowledge tracing and strategy injection based on the structured legal knowledge base and the historical case strategy feature base during the fusion process; generating a draft tender document based on the unified knowledge graph; performing compliance verification on the draft tender document; and determining the draft tender document that passes the verification as the final tender document. This application can improve the accuracy of tender document generation.

Description

Technical Field

[0001] This application relates to the field of document generation technology, and in particular to a method, system and storage medium for generating tender documents. Background Technology

[0002] In engineering construction procurement and large equipment procurement, tender documents, as core legal documents defining procurement needs, clarifying technical specifications, standardizing evaluation criteria, and stipulating contractual rights and obligations, directly determine the legality and fairness of the tendering process, as well as the efficiency and risk control level of subsequent project execution. With increasing project complexity and the continuous improvement of the tendering legal system, the traditional model relying on manual tender document preparation has gradually revealed problems such as low efficiency, long cycles, and significant impact on quality due to differences in personnel experience. Therefore, automated tender document generation technology has become a key demand direction for the industry to improve its tendering management level.

[0003] Currently, existing automated tender document generation solutions mainly revolve around two technical approaches. One is template filling and rule engine technology, which uses a pre-set fixed tender document template, manually configures field mapping relationships and simple logical rules, and substitutes basic project information into the template to complete document generation. The other is shallow natural language processing technology, which uses surface text features such as word frequency statistics and word vectors to achieve keyword matching and clause replacement. However, these existing technologies are difficult to meet the accuracy requirements of complex projects in practical applications and have relatively low accuracy. Summary of the Invention

[0004] To improve the accuracy of tender document generation, this application provides a tender document generation method, system, and storage medium.

[0005] Firstly, a method for generating tender documents is provided, the method comprising:

[0006] Acquire raw data from multiple modalities, including project requirement texts, design model files, legal and regulatory texts, and historical bidding case data;

[0007] The original data is subjected to multi-module semantic deep analysis and structured processing to generate project requirement sub-graph, spatial logic sub-graph, structured legal knowledge base and historical case strategy feature base respectively;

[0008] The project requirement sub-graph and spatial logic sub-graph are dynamically fused across modalities to construct a unified knowledge graph. Dynamic knowledge tracing and strategy injection are performed during the fusion process based on the structured legal knowledge base and historical case strategy feature base.

[0009] A draft of the tender document is generated based on the unified knowledge graph. The draft of the tender document is then verified for compliance, and the draft of the tender document that passes the verification is determined as the final tender document.

[0010] In some embodiments, the multi-module semantic deep parsing and structuring processing of the raw data includes:

[0011] Deep semantic role annotation is performed on the project requirement text to extract the operation subject-behavior-object-constraint tuple and construct the project requirement sub-graph.

[0012] The design model file is subjected to functional-spatial-logical topological relationship extraction to construct a spatial logical sub-graph.

[0013] The legal texts are decomposed into a structured form to generate a rule chain of applicable conditions, responsible parties, prescribed actions, and exemption scenarios, which is then stored in a pre-defined legal knowledge base to obtain a structured legal knowledge base.

[0014] Success and failure attribution analysis is performed on the historical bidding case data to generate feature vectors of success and risk patterns, which are then stored in a preset strategy feature library to obtain the historical case strategy feature library.

[0015] In some embodiments, the cross-modal dynamic fusion of the project requirement subgraph and the spatial logic subgraph to construct a unified knowledge graph includes:

[0016] Convert the project requirement sub-graph and the spatial logic sub-graph into heterogeneous graphs;

[0017] The semantic similarity between nodes from different subgraphs in the heterogeneous graph is calculated using a multi-view graph attention network.

[0018] Based on the semantic similarity, cross-modal semantic links are established between semantically similar nodes to merge the project requirement subgraph and the spatial logic subgraph to form a unified knowledge graph.

[0019] In some embodiments, the dynamic knowledge tracing and strategy injection during the fusion process based on the structured regulatory knowledge base and the historical case strategy feature base includes:

[0020] The key entity nodes in the unified knowledge graph are traversed, and the attributes of the key entity nodes are used as query conditions to match them in the structured legal knowledge base. The matched legal clauses are then attached to the corresponding key entity nodes as compliance evidence to complete dynamic knowledge tracing.

[0021] The current state of the unified knowledge graph is vectorized, and a similarity search is performed in the historical case strategy feature library. The strategy features of the historical case with the highest similarity are then injected as optimization suggestions into the corresponding decision node in the unified knowledge graph, thus completing the strategic knowledge injection.

[0022] In some embodiments, generating a draft tender document based on the unified knowledge graph includes:

[0023] Based on the hierarchical structure of the unified knowledge graph, a table of contents outline for the tender documents is generated.

[0024] A retrieval-enhanced generation architecture is adopted, which takes each node in the unified knowledge graph and its associated attributes, compliance basis and optimization suggestions as generation conditions and context, and inputs them into the large language model to generate corresponding natural language paragraphs. The natural language paragraphs are then combined to obtain the initial draft of the tender document.

[0025] In some embodiments, the compliance verification of the initial draft of the tender document further includes:

[0026] If the verification fails, a compliance warning will be generated;

[0027] Obtain feedback from senior personnel regarding the aforementioned compliance warnings;

[0028] Based on the feedback information, a preference dataset is constructed, and a reinforcement learning algorithm is used to fine-tune the large language model used to generate the initial draft of the tender document.

[0029] In some embodiments, calculating the semantic similarity between nodes from different subgraphs in the heterogeneous graph using a multi-view graph attention network includes:

[0030] The feature vector of each node is input into the view graph attention network, and the attention weights between nodes are calculated under multiple semantic views, including functional view, spatial view and constraint view.

[0031] We weighted and fused the attention weights from multiple perspectives to obtain the comprehensive semantic similarity between cross-modal nodes.

[0032] In some embodiments, the method further includes:

[0033] The knowledge graph update process is triggered in response to the detection of new legal and regulatory releases and / or design model changes and / or new data added to the historical case library.

[0034] Perform multimodal semantic parsing and structuring processing on newly added or modified original data to generate updated sub-maps and / or knowledge bases and / or feature bases.

[0035] Secondly, a tender document generation system is provided, comprising: a multimodal semantic deep parsing and structured representation module, a cross-modal dynamic fusion and intent collaboration central module, and a reinforcement learning-driven iterative generation and compliance verification module; wherein,

[0036] The multimodal semantic deep parsing and structured representation module is used to acquire raw data from multiple modalities, including project requirement text, design model files, legal and regulatory texts, and historical bidding case data. The raw data is subjected to multimodal semantic deep parsing and structured processing to generate project requirement sub-graph, spatial logic sub-graph, structured legal knowledge base, and historical case strategy feature base, respectively.

[0037] The cross-modal dynamic fusion and intent collaboration hub module is used to dynamically fuse the project requirement sub-graph and spatial logic sub-graph across modalities to construct a unified knowledge graph, and to perform dynamic knowledge tracing and strategy injection during the fusion process based on the structured legal knowledge base and historical case strategy feature base.

[0038] The reinforcement learning-driven iterative generation and compliance verification module is used to generate a draft tender document based on the unified knowledge graph, perform compliance verification on the draft tender document, and determine the draft tender document that passes the verification as the final tender document.

[0039] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon that can run on a processor, wherein when executed by the processor, the computer program implements a tender document generation method as described in the first aspect.

[0040] By employing the aforementioned method, this application first acquires raw data from multiple modalities, including project requirement texts, design model files, legal and regulatory texts, and historical bidding case data. Then, the raw data undergoes multi-module semantic deep parsing and structuring processing to generate a project requirement sub-graph, a spatial logic sub-graph, a structured legal knowledge base, and a historical case strategy feature base. Next, the project requirement sub-graph and the spatial logic sub-graph are dynamically fused across modalities to construct a unified knowledge graph. Dynamic knowledge tracing and strategy injection are performed during the fusion process based on the structured legal knowledge base and the historical case strategy feature base. Finally, a draft bidding document is generated based on the unified knowledge graph. The draft bidding document undergoes compliance verification, and the verified draft bidding document is determined as the final bidding document. In this way, through a multi-modal semantic deep parser, a cross-modal dynamic fusion hub, and a closed-loop compliance enhancement mechanism, a fundamental leap from formal compliance to intelligent optimization of bidding documents is achieved, improving the accuracy of bidding document generation. Attached Figure Description

[0041] Figure 1 This is a flowchart of a method for generating tender documents provided in this application.

[0042] Figure 2 This is a flowchart of the method for multi-module semantic deep parsing and structured processing of raw data provided in this application.

[0043] Figure 3 This is a schematic diagram of the overall technical framework provided in this application.

[0044] Figure 4 This is a flowchart of the phased technical implementation path provided in this application.

[0045] Figure 5 This is a schematic diagram of the connection of a tender document generation system provided in this application. Detailed Implementation

[0046] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but is consistent with the broadest scope claimed in this application.

[0047] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0048] Figure 1 This is a flowchart illustrating a method for generating tender documents provided in this application. For example... Figure 1 As shown, a method for generating tender documents includes the following steps:

[0049] Step S100: Obtain raw data from multiple modalities, including project requirement texts, design model files, legal and regulatory texts, and historical bidding case data.

[0050] The tender document generation method in this application is applied to terminals with processing capabilities, such as mobile phones, tablets, and computers. Step S100 above is the starting point for tender document generation, and its function is to comprehensively and accurately collect raw data related to the tender project from multiple heterogeneous data sources. This data is not only diverse in form and wide in origin, but also carries multi-dimensional information such as project requirements, technical specifications, legal basis, and historical experience, forming the foundation for accurately generating tender documents.

[0051] The project requirement texts in the original data include, but are not limited to, documents provided by the tendering party such as project requirement specifications, feasibility study reports, and technical specifications. These project requirement texts can be obtained through electronic document transfer, downloading from cloud sharing platforms, or directly copying from storage media provided by the tendering party. Metadata such as document version numbers and provision dates are also recorded to ensure data traceability.

[0052] Design model files, including BIM model files and CAD design drawings, can be provided by the design firm or professional team. The design model files can be obtained by communicating with the design firm and securely transmitting them via FTP server, dedicated data exchange platform, or encrypted email. Upon receiving the files, verify their integrity and accuracy to ensure no data corruption or loss.

[0053] Legal and regulatory texts include national, industry, and local bidding and tendering laws and regulations, technical standards, and mandatory provisions. The latest legal and regulatory texts can be obtained in bulk from authoritative official channels such as databases using search technology. Additionally, important laws and regulations can be obtained by purchasing genuine books or subscribing to professional legal database services.

[0054] Historical bidding case data includes historical bidding documents, winning bid announcements, bid evaluation reports, explanations of failed bids, and litigation cases. This data can be retrieved from the organization's or partner organizations' archives, or obtained through public channels such as corporate procurement websites. During the collection process, care must be taken to protect data privacy and security, and to avoid disclosing sensitive information. Furthermore, the collected case data should be categorized and organized to facilitate subsequent analysis.

[0055] Step S200 involves performing multi-module semantic deep analysis and structured processing on the raw data to generate project requirement sub-graphs, spatial logic sub-graphs, structured legal knowledge bases, and historical case strategy feature bases.

[0056] By performing multi-module semantic deep analysis and structured processing on the raw data, project requirement sub-graphs, spatial logic sub-graphs, structured legal knowledge bases, and historical case strategy feature bases are generated respectively. The traditional keyword matching and shallow feature extraction methods are abandoned. Instead, deep semantic analysis technology is used to transform the raw data of different modalities into highly structured, machine-understandable, and reasonable semantic knowledge units. Figure 2 This is a flowchart illustrating the method for multi-module semantic deep parsing and structured processing of raw data provided in this application. For example... Figure 2 As shown, the multi-module semantic deep parsing and structured processing of the original data includes the following steps:

[0057] Step S201: Perform deep semantic role annotation on the project requirement text, extract the operation subject-behavior-object-constraint tuple, and construct the project requirement sub-graph.

[0058] Step S202: Extract the functional-spatial-logical topological relationships from the design model file and construct a spatial logical sub-graph.

[0059] Step S203: The legal text is decomposed into a structured form to generate a rule chain of applicable conditions, responsible parties, prescribed actions, and exemption scenarios, and stored in a preset legal knowledge base to obtain a structured legal knowledge base.

[0060] Step S204: Perform success and failure attribution analysis on historical bidding case data, generate feature vectors for success and risk patterns, and store them in a preset strategy feature library to obtain the historical case strategy feature library.

[0061] To construct the project requirement subgraph, a deep semantic role labeling model based on fine-tuning of a pre-trained language model is employed. This model does not simply identify entities, but performs deep semantic and semantic analysis on sentences, labeling the core predicate verbs with their relevant semantic roles. For example, for the sentence "The contractor must provide centrifugal water pumps conforming to ISO-9001 standards," the model will parse the subject as "contractor," the action as "provide," the object as "centrifugal water pumps," and the constraint as "conforms to ISO-9001 standards." These elements are then organized into structured tuples: contractor - provide - centrifugal water pumps - standard ISO-9001. Finally, by traversing the entire project requirement text, all extracted tuples are used as nodes and relationships to dynamically construct a project requirement subgraph. In this graph, nodes represent entities such as equipment, materials, and contractors, and edges represent their requirements, specifications, and dependencies.

[0062] Pre-trained language models refer to large-scale language models based on the Transformer framework, pre-trained through self-supervised learning on massive amounts of unlabeled text data. Their core feature lies in their ability to deeply understand the syntax, semantics, and contextual logic of natural language through the pre-training process. Specifically, the models that can be used include, but are not limited to, encoder architecture models, encoder-decoder architecture models, and encoder-decoder architecture models. This achieves a leap from word-level processing to semantic-level understanding, accurately capturing the implicit actions, responsibilities, and constraints in the requirement text. This provides a precise and unambiguous semantic foundation for subsequent knowledge fusion and clause generation, fundamentally avoiding deviations in generated content caused by keyword ambiguity.

[0063] The construction of spatial logical subgraphs focuses on design documents such as BIM and CAD, disregarding their graphical rendering information and directly parsing their underlying data. Specifically, a specialized parsing library, such as ifcopenshell, is used to provide entity objects and their attributes from the model. Spatial and logical relationships between entities are extracted, forming entity-relationship-entity triplets. These triplets together constitute the spatial logical subgraph, which describes the physical layout and functional logic of the project. By transforming the unstructured design model into a structured knowledge graph rich in semantic relationships, the physical and functional composition of the project can be understood, enabling cross-modal alignment with textual requirements and ensuring consistency between the technical specifications in the tender documents and the design drawings. For example, the main water pump in the text can be associated with the Pump-001 entity in the BIM model.

[0064] For the structured legal knowledge base, legal provisions described in natural language are converted into machine-executable logical rules using legal text parsing technology. Specifically, named entity recognition is performed using a fine-tuned legal domain LLM to identify key elements in the provisions, such as applicable conditions, responsible parties, prescribed actions, numerical standards, and exemption scenarios. These elements are then populated into a predefined IF-THEN-EXCEPT rule chain template. Each rule chain is assigned a unique ID and associated with metadata such as the original source and effective date, stored in a pre-defined legal knowledge base such as a graph database or relational database, forming a structured legal knowledge base. For example, the provision "If the procurement amount exceeds 5 million yuan, the tenderer must use open tendering" becomes IF{procurement amount > 5 million yuan}THEN{tendering method = open tendering} after structuring. This transforms vague and lengthy legal provisions into precise and structured logical rules, providing a direct and computable prerequisite for subsequent automated compliance tracing and verification, which is key to achieving intrinsic compliance.

[0065] In addition, LAW-BERT and other domain models that have been pre-trained on legal texts, or LLAMA models that have been fine-tuned on legal corpora, can be used to enhance their ability to understand the structure, logic and terminology of legal provisions.

[0066] To obtain a historical case strategy feature library, strategic experience knowledge is mined from historical data. First, bidding documents, winning bids, and evaluation reports from historical projects are collected, and each case is tagged with success, risk, or failure. Then, advanced text embedding models, such as bge-m3, are used to convert key sections like scoring methods and payment terms into high-dimensional vectors. Simultaneously, LLM (Limited Language Management) is used for attribution analysis to generate textual descriptions, such as the case where the excessively high weighting of technical scores led to a shortage of high-quality suppliers. Finally, the vectorized representations of the cases and the attribution conclusions are associated and stored in a pre-defined strategy feature library—a vector database, such as Milvus—forming the historical case strategy feature library. These vectors do not represent specific content, but rather the strategic patterns that led to success or failure. Specifically, GPT series and LLAMA models with strong text understanding and summarization capabilities can be used to analyze and summarize the reasons for failure. This makes the implicit, experience-dependent, silent knowledge explicit and digitized. It not only ensures the compliance of bidding documents, but also proactively incorporates historically validated optimization strategies, such as reasonable scoring weights and risk avoidance clauses, thereby generating more competitive and market-adaptable intelligent optimization documents.

[0067] Step S300 involves cross-modal dynamic fusion of the project requirement sub-graph and the spatial logic sub-graph to construct a unified knowledge graph, and dynamic knowledge tracing and strategy injection during the fusion process based on the structured legal knowledge base and the historical case strategy feature base.

[0068] The core of achieving deep understanding and collaboration of multimodal data lies in dynamically fusing project requirement subgraphs and spatial logic subgraphs across modalities to construct a unified knowledge graph. Its purpose is to break down information silos between project requirements (e.g., text) and design models (e.g., structured data), forming a comprehensive, interconnected, and reasonable unified knowledge system. The process of dynamically fusing project requirement information subgraphs and spatial logic subgraphs across modalities to construct a unified knowledge graph includes the following steps:

[0069] Step S301: Convert the project requirement sub-graph and spatial logic sub-graph into a heterogeneous graph.

[0070] Step S302: Use a multi-view graph attention network to calculate the semantic similarity between nodes from different subgraphs in the heterogeneous graph.

[0071] Step S303: Based on semantic similarity, establish cross-modal semantic links between semantically similar nodes to integrate the project requirement subgraph and the spatial logic subgraph to form a unified knowledge graph.

[0072] First, the project requirement subgraph and spatial logic subgraph generated in step S200 are converted into a heterogeneous graph, in which two types of nodes and edges coexist. The two types of nodes include project requirement nodes and spatial logic nodes. Project requirement nodes originate from text parsing, such as main water pump, fire protection system, contractor, etc. Spatial logic nodes originate from design models, such as Pump-001, Fire-Fighting-System-01, Room-B1-102, etc. The edges within each subgraph remain unchanged; initially, there are no connections between nodes in two subgraphs.

[0073] Then, to achieve deep fusion of the two sub-graphs, a multi-view graph attention network is used to calculate the semantic similarity between nodes in different sub-graphs and establish cross-modal links. The first step is node feature vectorization, which encodes the attributes, type, and structural relationships of each node in the original graph into high-dimensional feature vectors using a graph neural network. The second step is a multi-view attention mechanism, where the graph attention network does not focus on a single perspective but rather on three key semantic dimensions: functional, spatial, and constraint perspectives, assigning attention weights between nodes. For the functional perspective, the focus is on the node's purpose and behavior, such as calculating the semantic association between the text node "water pump" and the model node in terms of functions like water supply and pressurization. For the spatial perspective, the focus is on the node's spatial location and layout relationships, such as determining whether the equipment in Room-B1-102 of the BIM model described in the text as located in the underground pump room has consistent locations. For the constraint perspective, the focus is on the technical parameters and standard constraints imposed on the node. For example, comparing whether the requirement in the text that the water pump flow rate is not less than a certain value matches the corresponding model node. The third step involves weighted fusion and link establishment, which involves weighted fusion of the attention weights calculated from the three perspectives to obtain a comprehensive semantic similarity score. When this score exceeds a preset threshold, a semantically equivalent or instantiated cross-modal semantic link is automatically established between the two nodes. The fourth step forms a unified knowledge graph, which deeply merges two previously isolated sub-graphs into a single unified knowledge graph. In this graph, all information about a concept is integrated. This achieves deep and interpretable semantic fusion. Through a multi-perspective mechanism, the semantics of nodes can be comprehensively understood from multiple core dimensions such as function, space, and constraints, ensuring the accuracy and rationality of cross-modal links, reducing semantic ambiguity, and making the fusion process transparent and interpretable. Furthermore, it breaks down information silos between modalities, forming a network rich in semantic relationships, enabling complex reasoning and thus enhancing reasoning capabilities. In addition, the unified knowledge graph, as the sole trusted source, ensures that subsequently dynamically added legal basis and strategic suggestions can be accurately mapped to the correct entities and clauses, thereby indirectly improving the compliance, strategy, and overall quality of the final generated tender documents.

[0074] The process of calculating the semantic similarity between nodes from different subgraphs in the heterogeneous graph using a multi-view graph attention network includes the following steps:

[0075] Step S302-1: Input the feature vector of each node into the view graph attention network, and calculate the attention weights between nodes under multiple semantic views, wherein the multiple semantic views include functional view, spatial view and constraint view.

[0076] Step S302-2: Weighted fusion of attention weights from multiple perspectives is performed to obtain the comprehensive semantic similarity between cross-modal nodes.

[0077] The high-dimensional feature vector of each node is projected into three perspective-specific subspaces: functional perspective, spatial perspective, and constraint perspective, generating perspective-specific feature representations. Then, parallel attention computation is performed, that is, within each perspective subspace, an independent graph attention network works in parallel, calculating the attention weights of node pairs from different subgraphs under that perspective. This weight represents the importance of one node to another from the current perspective.

[0078] Next, a weighted fusion of attention weights is performed. The first step is weight fusion, where, after obtaining the attention weights from the functional, spatial, and constraint perspectives, instead of simply averaging them, a weighted fusion is performed. This involves assigning a learnable or pre-defined importance coefficient to each perspective, reflecting the relative importance of different perspectives in a specific domain or task. The second step is comprehensive semantic similarity, where the attention weights from the three perspectives are combined into a comprehensive semantic similarity score using a weighted formula. This score is the final, comprehensive quantitative assessment of the cross-modal association strength between two nodes. The third step is link decision, where a decision is made based on this comprehensive score regarding whether to establish a cross-modal semantic link. A link is only established between nodes when the comprehensive score exceeds a pre-defined confidence threshold, thus ensuring the accuracy and reliability of the fusion. This approach allows for more accurate judgments by combining multiple perspectives, and the multi-perspective mechanism, through multi-dimensional cross-validation, effectively avoids semantic ambiguity and significantly reduces the probability of false links and missed links. Furthermore, the functional, spatial, and constraint perspectives together constitute the core dimensions for understanding entities in the engineering field. Through collaborative information processing, it moves beyond superficial word or graphic matching to a deeper semantic understanding of the entity's essence, thus achieving true deep semantic fusion. On the other hand, the weighted fusion mechanism of perspectives is flexible; the importance of each perspective can be adjusted for different types of projects, exhibiting good domain adaptability and enabling its wide application in bidding projects of various natures.

[0079] During the construction of a unified knowledge graph, i.e., the integration process, dynamic knowledge tracing and strategic knowledge injection are required based on a structured legal knowledge base and a historical case strategy feature base. This dynamic knowledge tracing and strategy injection based on the structured legal knowledge base and the historical case strategy feature base during the integration process includes the following steps:

[0080] Step S205: Traverse the key entity nodes in the unified knowledge graph, use the attributes of the key entity nodes as query conditions to match them in the structured legal knowledge base, and attach the matched legal clauses as compliance basis to the corresponding key entity nodes to complete dynamic knowledge tracing.

[0081] Step S206: Vectorize the current state of the unified knowledge graph, perform similarity retrieval in the historical case strategy feature library, and inject the strategic knowledge of the historical case corresponding to the largest similarity found.

[0082] After forming a unified knowledge graph through a multi-view graph attention network, it is not regarded as a static result, but as an augmentable intelligent hub. It injects legal basis and historical wisdom into the graph through the following two parallel dynamic processes.

[0083] The dynamic knowledge tracing component automatically and in real time finds and associates the legal basis for each key requirement node in the knowledge graph, ensuring that the content of the bidding documents is compliant from the source.

[0084] Specifically, the first step is to traverse the key entity nodes, that is, to traverse the unified knowledge graph and identify all key entity nodes that represent requirements, specifications, standards, or responsibilities, such as pressure vessels, fire ratings, contractor qualifications, payment terms, etc.

[0085] Then, a query matching mechanism is constructed, which involves providing the core attributes of the aforementioned nodes, such as the design pressure of the pressure vessel and the toxicity of the medium, and converting them into structured query conditions. Subsequently, real-time matching and retrieval are performed in the structured regulatory knowledge base.

[0086] Finally, compliance evidence is attached, namely specific clauses matched in the knowledge base. If a clause in the Pressure Vessel Safety Technical Supervision Regulations stipulates that vessels with a design pressure greater than or equal to 1.6 MPa require Class Y testing, this will be used as a compliance evidence attribute and directly attached to the corresponding key entity node, completing dynamic knowledge traceability. This transforms the original ex-post, manual compliance check into an in-process, endogenous, and automated compliance embedding. In the generated tender documents, every technical or commercial requirement can be traced back to its regulatory source, greatly reducing legal risks and significantly improving the professionalism and authority of the document preparation. Moreover, legal clauses are no longer external reference documents but are directly structured into the inherent attributes of the knowledge graph, ensuring that every requirement in the initial draft of the tender documents is legally sound and has rules to follow. Compliance checks are moved from manual verification after generation to an automated and standardized process during generation, reducing the possibility of citing outdated or incorrect regulations due to negligence.

[0087] The strategic knowledge injection section involves applying successful experiences and lessons learned from past projects to the current project in the form of strategic suggestions, so that the generated tender documents are not only compliant, but also market-adaptable and strategically superior.

[0088] Specifically, the first step is to vectorize the graph state, which involves transforming the overall topology, node attributes, and relationships of the current unified knowledge graph into a global state vector through graph embedding technology. This vector comprehensively represents the core features of the current project.

[0089] Then, a historical case similarity retrieval is performed, which involves using the aforementioned global state vector as the query condition to perform a similarity retrieval in the historical case strategy feature library to find the most similar historical successful and risky cases to the current project.

[0090] Finally, the injection optimization is established by using the strategy features adopted in the most similar successful cases as optimization suggestion attributes and injecting them into the corresponding strategy nodes in the unified knowledge graph, thus completing the strategy injection. This makes tacit knowledge explicit and applicable, allowing the experience of senior bidding personnel accumulated in historical cases to be applied on a large scale to new projects. This makes the generated bidding documents more strategic and competitive in terms of clause settings and scoring criteria, thus moving from formal contracts to intelligent optimized contracts.

[0091] In addition, the strategic knowledge injection mechanism can stand on the shoulders of giants, continuously learn successful bidding strategies and avoid known risks from historical data, and dynamically recommend this strategic knowledge to new projects. This makes the output bidding documents more market-foresighted and strategically adaptable, and better achieve procurement objectives, rather than just completing a compliance paperwork.

[0092] Furthermore, the dynamically enhanced knowledge graph contains rich original information, cross-modal connections, legal basis, and optimization suggestions at each node. The tender documents generated based on this foundation are far more accurate, in-depth, and practical than simple template filling or keyword replacement, thus improving the accuracy and depth of the generated content.

[0093] Step S400: Generate a draft of the tender document based on the unified knowledge graph, perform compliance verification on the draft of the tender document, and determine the draft of the tender document that passes the verification as the final tender document.

[0094] Step S400 above is the core step in transforming the previously constructed and enhanced intelligent knowledge kernel into the final deliverable tender document. Its purpose is to ensure that the generated document is not only accurate in content and rigorous in structure, but also possesses good readability and standardization. First, a draft tender document needs to be generated based on a unified knowledge graph. Generating a draft tender document based on a unified knowledge graph includes the following steps:

[0095] Step S401: Generate the outline of the tender document based on the hierarchical structure of the unified knowledge graph.

[0096] Step S402: Using a retrieval-enhanced generation architecture, the attributes, compliance basis, and optimization suggestions of each node in the unified knowledge graph and their associations are used as generation conditions and context, and input into the large language model to generate corresponding natural language paragraphs. The natural language paragraphs are then combined to obtain the initial draft of the tender document.

[0097] The process of generating the initial draft of the tender document adopts a structured and controlled generation strategy, which can avoid the instability of content generation by traditional large language models.

[0098] Specifically, the first step is to generate a table of contents outline. This is achieved by analyzing the hierarchical topology of a unified knowledge graph. The root node of the graph is typically a project's tender document, and its next-level child nodes naturally correspond to the core chapters of the tender document. For example, Chapter 1 is "Instructions for Bidders," corresponding to the set of nodes in the graph related to the bidding process and rules; Chapter 2 is "Project Technical Requirements," corresponding to the subtree of nodes representing all technical requirements and equipment specifications; Chapter 3 is "Evaluation Methods and Standards," corresponding to the scoring nodes and strategy suggestions in the graph; and Chapter 4 is "Contract Terms and Formats," corresponding to the legally relevant nodes in the graph such as responsible parties and prescribed actions. By traversing the graph in this way, and based on the type and aggregation relationship of the nodes, a logically clear and complete tender document table of contents outline is automatically generated, ensuring the structure and standardization of the generated document.

[0099] Then, a retrieval-enhanced generation framework is used to generate natural language paragraphs. This is achieved by employing a retrieval-enhanced generation architecture, generating content chapter by chapter and clause by clause. This includes context retrieval, conditional generation, and paragraph generation and combination. Context retrieval involves retrieving all relevant nodes, node attributes, associated compliance evidence, and optimization suggestions from a unified knowledge graph for each chapter in the table of contents. Conditional generation assembles all retrieved structured information, such as nodes, attributes, evidence, and suggestions, into a highly dense and precise package of generated conditions and context. This context package is then used as prompts and input into the large language model. Paragraph generation and combination, under specific and strict contextual constraints, involves the large language model organizing and polishing the text to generate natural language paragraphs that conform to official document standards and are grammatically correct.

[0100] Finally, following the directory structure, all generated paragraphs are combined to output a complete draft of the tender document. This approach achieves several advantages: First, every core fact in the generated content originates directly from a verified knowledge graph, fundamentally ensuring the authenticity and accuracy of the tender document content. Second, the generated tender document is no longer a black-box production; its structure stems from the hierarchy of the knowledge graph, and its specific clauses originate from specific nodes and attributes within the graph. This allows each sentence and requirement to be traced back to its semantic source, resulting in strong interpretability and auditability. Third, because the generation context includes compliance evidence and optimization suggestions from dynamic tracing and strategy injection, the generated clauses are not merely dry technical parameters but rather comprehensive, high-quality text integrating legal basis and historical wisdom. Fourth, the unified knowledge graph avoids common problems such as structural chaos, content duplication, or loss when generating long tender document drafts, reducing the complexity of the generation task, improving generation speed and quality stability, and ensuring the accuracy of the tender document draft.

[0101] After generating a draft tender document using the enhanced architecture, it was not directly used as the final deliverable. Instead, an automated and iterative compliance verification process was introduced to act as the final checklist, ensuring the document's legality and high quality. This process employs a reverse mapping and logical consistency check method to perform final verification of the generated content.

[0102] Specifically, the first step is to perform text parsing and key information extraction. This involves analyzing the generated draft tender document and using natural language processing technology to identify and extract key requirements, commitments, timelines, numerical standards, liability clauses, and other entities and relationships.

[0103] Then, the data is back-mapped to the knowledge source, meaning these extracted key assertions are back-mapped to the unified knowledge graph and the structured regulatory knowledge base. Specifically, it checks whether the clauses in the generated file are consistent with the original requirements and attributes of the nodes in the knowledge graph, thus completing the mapping to the knowledge graph. It also checks whether the mandatory requirements in the generated file match the clauses in the latest regulatory base, thus completing the mapping to the regulatory knowledge base.

[0104] Next, a logical consistency check is performed, which involves cross-clause logical checks. Finally, a compliance warning is generated. If the above mapping fails or a logical conflict is found, a compliance warning will be automatically generated, clearly indicating the problematic paragraph, the type of problem, and possible modification suggestions. If the above mapping is successful and no logical conflicts are found, the verification is considered passed, and this verified draft of the tender document will be designated as the final tender document. This establishes the final quality defense line, reducing legal and enforcement risks.

[0105] In addition, after verifying the compliance of the initial draft of the tender document, the process also includes: generating a compliance warning if the verification fails; obtaining feedback from senior personnel regarding the compliance warning; constructing a preference dataset based on the feedback information; and using reinforcement learning algorithms to fine-tune the large language model used to generate the initial draft of the tender document.

[0106] Specifically, if a compliance alert requiring manual intervention is generated, the alert, along with the initial draft, will be submitted to senior professionals in the field. These professionals will review the draft and provide feedback. This feedback will be recorded and compiled into a preference dataset. Reinforcement learning algorithms will be used periodically to fine-tune the large language model used in the content generation process based on this dataset. This allows the model to learn the decision-making preferences of senior professionals, reducing the occurrence of similar compliance issues in future generation processes, forming a closed-loop optimization system that becomes increasingly intelligent with use. This self-checking and learning capability enhances the reliability and credibility of the entire technical solution. Furthermore, although a verification process is introduced, its high degree of automation makes it far more efficient than manual review from start to finish. Senior professionals are freed from tedious comprehensive proofreading and only need to identify a few challenging alerts, thus achieving a significant leap in overall compilation efficiency while ensuring high-quality documents.

[0107] Preferably, in response to the detection of new legal and regulatory releases and / or design model changes and / or new data added to the historical case library, the knowledge graph update process is triggered; and the newly added or changed original data is subjected to multimodal semantic parsing and structuring processing to generate updated sub-graphs and / or knowledge bases and / or feature bases.

[0108] This application possesses continuous evolution capabilities. The specific implementation methods are as follows: First, it updates the triggering conditions and detection mechanisms. For updates to laws and regulations, it continuously detects the release, revision, or repeal of laws and regulations through official enterprise APIs, authoritative legal database push services, or configured timed web crawlers. Once a regulatory change related to the procurement field is detected, the update process is immediately triggered. For changes to the design model, it monitors file version update times to detect the version iteration of the project design model in real time and triggers an update. For additions to the historical case library, when a new bidding project is completed, its related documents and data are automatically entered into the library through a preset data interface, or manually imported by the administrator, and then identified as new data, triggering an update.

[0109] Then, incremental multimodal semantic parsing and structured processing are performed. Instead of reconstructing all knowledge at every trigger, incremental processing is used. For new regulations, they are structurally decomposed to generate new or updated rule chains of applicable conditions, responsible parties, prescribed actions, and exemption scenarios, which are incrementally stored in the structured regulatory knowledge base. Simultaneously, old rules that have been replaced or repealed are marked. For changed design models, the new version of the document is parsed, extracting the functional-spatial-logical topological relationships of the changed parts, generating an incremental spatial logic sub-graph update package. For newly added historical cases, cost attribution analysis is performed to generate feature vectors of new success or risk patterns, which are incrementally stored in the historical case strategy feature library.

[0110] Next, the knowledge graph undergoes system updates and consistency maintenance. For regulatory updates, all nodes in the graph that reference the updated regulations are traversed, their compliance attributes are updated to the latest clauses, and version change logs are recorded. For model changes, the incremental spatial logic sub-graph is re-aligned and merged with the existing graph, updating or replacing changed entities and relationships. For new cases, the strategy feature library is expanded to prepare for injecting richer optimization suggestions into similar projects in the future. This transforms it from a fixed, delivered tool into an organic entity capable of continuous learning from a constant influx of new cases. Each completed project, regardless of success or failure, becomes nourishment for optimization, making its strategy recommendation capabilities increasingly accurate and intelligent, enabling self-evolution and knowledge accumulation.

[0111] Figure 3 This is a schematic diagram of the overall technical framework provided in this application. (For example...) Figure 3As shown, a microservice architecture is adopted, with core components including a multimodal parsing engine, a fusion hub, a knowledge base, a generation engine, and a verification feedback loop. All services communicate and are scheduled through a unified API gateway, while data is shared and persistently stored via a central knowledge graph database. This architecture clearly demonstrates the entire technical path from multimodal data input, semantic parsing and fusion, knowledge enhancement and tracing, to final document generation and compliance verification, highlighting the collaborative relationships and data flow between modules.

[0112] Figure 4 This is a flowchart illustrating the phased technical implementation path provided in this application. For example... Figure 4 As shown, the implementation path is divided into four main stages: Stage 1 is data standardization and knowledge base construction, covering the establishment of a regulatory knowledge base and a historical case strategy feature base; Stage 2 is the development of a multimodal semantic deep parser, including the structured processing of raw data such as text and design models; Stage 3 is the implementation of a cross-modal dynamic fusion hub, focusing on the semantic alignment and dynamic knowledge injection mechanism based on graph neural networks; Stage 4 is condition generation and compliance verification, demonstrating the knowledge graph-based tender document generation and closed-loop verification and optimization process. This flowchart systematically outlines the complete technical chain from data input to document generation and system self-evolution, clarifying the core tasks and outputs of each stage.

[0113] Figure 5 This is a schematic diagram of the connection of a tender document generation system provided in this application. For example... Figure 5 As shown, a tender document generation system includes: a multimodal semantic deep parsing and structured representation module, a cross-modal dynamic fusion and intent collaboration central module, and a reinforcement learning-driven iterative generation and compliance verification module.

[0114] The system comprises several modules: a multimodal semantic deep parsing and structured representation module, which acquires raw data from multiple modalities, including project requirement texts, design model files, legal and regulatory texts, and historical bidding case data; a cross-modal dynamic fusion and intent collaboration module, which dynamically fuses the project requirement sub-graph and spatial logic sub-graph to construct a unified knowledge graph; and a module that dynamically traces knowledge sources and injects strategies based on the structured legal knowledge base and historical case strategy feature base during the fusion process. Finally, a reinforcement learning-driven iterative generation and compliance verification module, which generates a draft bidding document based on the unified knowledge graph, performs compliance verification on the draft, and designates the verified draft as the final bidding document.

[0115] The other functions performed by the aforementioned multimodal semantic deep parsing and structured representation module, cross-modal dynamic fusion and intent collaboration hub module, and reinforcement learning-driven iterative generation and compliance verification module, as well as the technical details of each function, are the same or similar to the features in the tender document generation method described above, and therefore will not be repeated here.

[0116] This application also provides a computer storage medium storing a computer program that, when run on a computer, enables the computer to execute the steps in the tender document generation method described above.

[0117] It should be understood that although the steps in the flowcharts in the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order in which these steps are performed, and they may be performed in other orders.

[0118] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for generating tender documents, characterized in that, The method includes: Acquire raw data from multiple modalities, including project requirement texts, design model files, legal and regulatory texts, and historical bidding case data; The original data is subjected to multi-module semantic deep analysis and structured processing to generate project requirement sub-graph, spatial logic sub-graph, structured legal knowledge base and historical case strategy feature base respectively; The project requirement sub-graph and spatial logic sub-graph are dynamically fused across modalities to construct a unified knowledge graph. Dynamic knowledge tracing and strategy injection are performed during the fusion process based on the structured legal knowledge base and historical case strategy feature base. A draft of the tender document is generated based on the unified knowledge graph. The draft of the tender document is then verified for compliance. The draft of the tender document that passes the verification is determined as the final tender document. Specifically, the design model file is subjected to functional-spatial-logical topological relationship extraction to construct a spatial logical sub-graph. The process of dynamically fusing the project requirement subgraph and the spatial logic subgraph across modalities to construct a unified knowledge graph includes: Convert the project requirement sub-graph and the spatial logic sub-graph into heterogeneous graphs; The semantic similarity between nodes from different subgraphs in the heterogeneous graph is calculated using a multi-view graph attention network. Based on the semantic similarity, cross-modal semantic links are established between semantically similar nodes to merge the project requirement subgraph and the spatial logic subgraph to form a unified knowledge graph; The process of generating a draft tender document based on the unified knowledge graph includes: Based on the hierarchical structure of the unified knowledge graph, a table of contents outline for the tender documents is generated. A retrieval-enhanced generation architecture is adopted, which takes each node in the unified knowledge graph and its associated attributes, compliance basis and optimization suggestions as generation conditions and context, and inputs them into the large language model to generate corresponding natural language paragraphs. The natural language paragraphs are then combined to obtain the initial draft of the tender document.

2. The method according to claim 1, characterized in that, The multi-module semantic deep parsing and structured processing of the original data includes: Deep semantic role annotation is performed on the project requirement text to extract the operation subject-behavior-object-constraint tuple and construct the project requirement sub-graph. The legal texts are decomposed into a structured form to generate a rule chain of applicable conditions, responsible parties, prescribed actions, and exemption scenarios, which is then stored in a pre-defined legal knowledge base to obtain a structured legal knowledge base. Success and failure attribution analysis is performed on the historical bidding case data to generate feature vectors of success and risk patterns, which are then stored in a preset strategy feature library to obtain the historical case strategy feature library.

3. The method according to claim 1, characterized in that, The dynamic knowledge tracing and strategy injection during the fusion process based on the structured legal knowledge base and the historical case strategy feature base includes: The key entity nodes in the unified knowledge graph are traversed, and the attributes of the key entity nodes are used as query conditions to match them in the structured legal knowledge base. The matched legal clauses are then attached to the corresponding key entity nodes as compliance evidence to complete dynamic knowledge tracing. The current state of the unified knowledge graph is vectorized, and a similarity search is performed in the historical case strategy feature library. The strategy features of the historical case with the highest similarity are then injected as optimization suggestions into the corresponding decision node in the unified knowledge graph, thus completing the strategic knowledge injection.

4. The method according to claim 1, characterized in that, The compliance verification of the initial draft of the tender documents also includes: If the verification fails, a compliance warning will be generated; Obtain feedback from senior personnel regarding the aforementioned compliance warnings; Based on the feedback information, a preference dataset is constructed, and a reinforcement learning algorithm is used to fine-tune the large language model used to generate the initial draft of the tender document.

5. The method according to claim 1, characterized in that, The calculation of semantic similarity between nodes from different subgraphs in the heterogeneous graph using a multi-view graph attention network includes: The feature vector of each node is input into the view graph attention network, and the attention weights between nodes are calculated under multiple semantic views, including functional view, spatial view and constraint view. We weighted and fused the attention weights from multiple perspectives to obtain the comprehensive semantic similarity between cross-modal nodes.

6. The method according to claim 1, characterized in that, The method further includes: The knowledge graph update process is triggered in response to the detection of new legal and regulatory releases and / or design model changes and / or new data added to the historical case library. Perform multimodal semantic parsing and structuring processing on newly added or modified original data to generate updated sub-maps and / or knowledge bases and / or feature bases.

7. A tender document generation system, characterized in that, The system includes: a multimodal semantic deep parsing and structured representation module, a cross-modal dynamic fusion and intent collaboration central module, and a reinforcement learning-driven iterative generation and compliance verification module; wherein, The multimodal semantic deep parsing and structured representation module is used to acquire raw data from multiple modalities, including project requirement text, design model files, legal and regulatory texts, and historical bidding case data. The raw data is subjected to multimodal semantic deep parsing and structured processing to generate project requirement sub-graph, spatial logic sub-graph, structured legal knowledge base, and historical case strategy feature base, respectively. The cross-modal dynamic fusion and intent collaboration hub module is used to dynamically fuse the project requirement sub-graph and spatial logic sub-graph across modalities to construct a unified knowledge graph, and to perform dynamic knowledge tracing and strategy injection during the fusion process based on the structured legal knowledge base and historical case strategy feature base. The reinforcement learning-driven iterative generation and compliance verification module is used to generate a draft of the tender document based on the unified knowledge graph, perform compliance verification on the draft of the tender document, and determine the draft of the tender document that passes the verification as the final tender document. Specifically, the design model file is subjected to functional-spatial-logical topological relationship extraction to construct a spatial logical sub-graph. The process of dynamically fusing the project requirement subgraph and the spatial logic subgraph across modalities to construct a unified knowledge graph includes: Convert the project requirement sub-graph and the spatial logic sub-graph into heterogeneous graphs; The semantic similarity between nodes from different subgraphs in the heterogeneous graph is calculated using a multi-view graph attention network. Based on the semantic similarity, cross-modal semantic links are established between semantically similar nodes to merge the project requirement subgraph and the spatial logic subgraph to form a unified knowledge graph; The process of generating a draft tender document based on the unified knowledge graph includes: Based on the hierarchical structure of the unified knowledge graph, a table of contents outline for the tender documents is generated. A retrieval-enhanced generation architecture is adopted, which takes each node in the unified knowledge graph and its associated attributes, compliance basis and optimization suggestions as generation conditions and context, and inputs them into the large language model to generate corresponding natural language paragraphs. The natural language paragraphs are then combined to obtain the initial draft of the tender document.

8. A computer-readable storage medium having a computer program stored thereon that can run on a processor, characterized in that, When the computer program is executed by the processor, it implements a tender document generation method as described in any one of claims 1 to 6.

Images