Document structured generation and verification method for bidding scene

Abstract

The embodiment of the invention provides a document structured generation and verification method oriented to bidding and tendering scenes, which is applied to the technical field of natural language processing and intelligent document generation, and comprises the following steps: analyzing a bidding document to construct a bidding knowledge graph and generate a first bidding document draft; initializing three intelligent agents of review simulation, competitive confrontation and compliance guards, and generating feedback information in each round of deduction; performing multi-round iterative adversarial deduction on the first draft, encoding a document state and agent feedback into a state vector, inputting the state vector into the reinforcement learning model, outputting a modification operation sequence and revising the document until convergence is performed to obtain a deduction convergence result set; backtracking and extracting a key modification operation sequence; and reconstructing a first draft according to the sequence, and outputting a final-edition bidding document and a compilation basis description document. According to the method, global optimization of the bidding document in three dimensions of score, risk and compliance is realized, a traceable compilation basis is generated, and the intelligent level of document compilation and the interpretability of a result are remarkably improved.

Description

Technical Field

[0001] This application relates to the fields of natural language processing and intelligent document generation technology, and in particular to a method for structured document generation and verification for bidding scenarios. Background Technology

[0002] The preparation of tender documents relies heavily on manual labor, requiring personnel to extract technical specifications, commercial terms, and scoring criteria from dozens of pages of tender documents. Due to the complex structure and interwoven hierarchical structure of the tender documents, manual analysis is prone to overlooking key clauses or misunderstanding constraints, leading to unresponsive or off-target tender documents that directly affect the validity of the bid.

[0003] Risk identification relies heavily on individual experience. Drafters often struggle to comprehensively assess the potential risks posed by the response content, such as whether there are implicit conflicts between commitments in different sections, whether technical parameters are easily questionable, and whether supporting documentation adequately addresses the scoring criteria. The lack of systematic quantitative assessment methods leads to the overlooking of many avoidable risks.

[0004] Document optimization lacks a multi-objective balancing approach. Bidding documents need to strike a balance between improving scores, reducing the risk of being questioned, and ensuring compliance. However, traditional modification methods often address symptoms rather than the root cause, meaning that a change that improves scores may increase the risk of being questioned. Furthermore, the preparation process lacks complete documentation, making it difficult to trace the logic behind the final document's creation and lacking convincing evidence. Summary of the Invention

[0005] This application provides a method for structured document generation and verification in bidding scenarios. Through multi-agent adversarial simulation, it achieves global optimization of bidding documents across three dimensions: score, risk, and compliance, and generates traceable compilation basis. This transforms the document compilation process from uninterpretable experience-based reliance to quantifiable intelligent decision-making. To achieve the above objectives, this application adopts the following technical solution: A method for generating and validating structured documents for bidding and tendering scenarios, the method comprising: The bidding documents are parsed to construct a bidding knowledge graph, and a draft of the bid document is generated based on the bidding knowledge graph. Within the rule space defined by the bidding knowledge graph, multiple agents with different objective functions are initialized. These agents include at least a review simulation agent, a competitive adversarial agent, and a compliance guardian agent. An instantiated review simulation agent is loaded with historical bid evaluation behavior data and the terms, rules and scoring details in the bidding knowledge graph, which are used to generate a scoring report including advantages and potential deductions during the simulation. Instantiate a competitive adversarial agent, load a game strategy library including historical questioning data and differentiated strategies, as well as the terms, rules and scoring details in the bidding knowledge graph, to generate a list of potential questioning points including the location and reasons of the questioning terms; Instantiate the compliance guardian intelligent agent, load regulatory case data and the terms, rules and scoring details in the bidding knowledge graph, and use them to generate a compliance risk list including violation clauses and risk levels; The completed loading of the review simulation agent, the competitive adversarial agent, and the compliance guardian agent will be used as the initial set of agents to start multiple rounds of iterative adversarial simulations. These agents will be used to generate feedback information in each round of simulations. Multiple rounds of iterative adversarial simulations are performed on the initial draft of the tender document until a preset convergence condition is met, resulting in a simulation convergence result set; each round of simulation includes: The current document state and the feedback information of each agent are encoded into a state vector, which is then input into the reinforcement learning decision model to obtain a sequence of modification operations. The document is then revised according to this sequence, and the updated document state is generated as the input for the next round. The above process is repeated until the preset convergence condition is met. Based on the backtracking of the convergence result set accumulated during the deduction process, the sequence of key modification operations performed from the first draft to the final draft is obtained; Based on the aforementioned key modification operation sequence, the initial draft of the tender document is reconstructed, and the final version of the tender document and the corresponding explanatory document on the basis for its preparation are output.

[0006] In some possible implementations, the process of parsing the tender documents, constructing a tender knowledge graph, and generating a draft tender document based on the tender knowledge graph includes: Natural language processing is performed on the tender documents to extract technical specifications, commercial terms, scoring rules and mandatory constraints, resulting in structured clause data; The structured terms data are subjected to entity identification and relationship extraction to extract key technical parameters, delivery deadlines, qualification requirements entities, and to identify the constraints, mutual exclusions, and dependencies between entities, resulting in an entity set and a relationship set. The scoring criteria are broken down into quantifiable scoring points to obtain a set of scoring points; Using chapters as parent nodes and clauses as child nodes, the entity set, relation set, and scoring point set are associated with the corresponding nodes to construct a bidding knowledge graph that includes nodes, attributes, and edge relationships; An initial bid document framework is generated based on the chapter node structure of the bidding knowledge graph. The basic material library is then called to fill in the corresponding clause nodes, forming the first draft of the bid document.

[0007] In some possible implementation methods, the extraction of technical specifications, commercial terms, scoring rules, and mandatory constraints yields structured terms data, including: The tender documents are identified by chapters, chapter titles and clause boundaries are marked, and a chapter structure sequence is obtained; Syntactic analysis is performed on the clause text in the chapter structure sequence to extract key technical parameters and delivery deadlines, resulting in parameter entities; Keyword matching is performed on the text of the terms to identify the qualification requirements and obtain the qualification entity. The text of the scoring rules is parsed to identify the scoring items and their weights and scoring conditions, and obtain the scoring points and their quantitative attributes. The parameter entities, qualification entities, scoring points and their quantitative attributes are associated with the corresponding clauses in the chapter structure sequence to obtain structured clause data including chapter levels and clause content.

[0008] In some possible implementations, encoding the current document state and the feedback information from each agent into a state vector includes: Obtain the scoring report of the review simulation agent output in the current round, the list of potential questionable points of the competitive agent output, and the list of compliance risks of the compliance guardian agent output; The predicted scores, the number of potential deductions, the number of potential points of contention in the potential questioning list, and the risk level data in the compliance risk list in the scoring report are concatenated with the text features of the current document to generate a multi-dimensional state vector.

[0009] In some possible implementations, the convergence condition includes: The scoring prediction values ​​output by the review simulation agent in multiple consecutive rounds are obtained. When the change range of the scoring prediction value is less than a preset threshold and higher than a preset pass line, it is determined that the first convergence condition is met. Obtain a list of potential points of contention for the competitive adversarial agent's output over multiple consecutive rounds. When the list of points of contention is empty or the contents of the list are repeated with the historical output, it is determined that the second convergence condition is met. Obtain the compliance risk list output by the compliance guardian intelligent agent in multiple consecutive rounds. When there are no major risk items in the compliance risk list or all risk items are in an eliminated state, it is determined that the third convergence condition is met.

[0010] In some possible implementations, the reinforcement learning decision model includes: Document modification operations are predefined as a set of actions, including clause rewriting, supplementary proof, and adjustment of commitment values; Based on the changes in document status after performing the modification operations in the set of actions, the combined changes of improved rating, reduced points of contention, and decreased risk level are quantified as reward values. Based on the accumulated state data, the modified operations performed, and the reward values ​​obtained during the historical simulation process, the policy network is trained to establish a mapping relationship between states and action sequences, thus obtaining a reinforcement learning decision model.

[0011] In some possible implementations, the sequence of key modification operations performed from the initial draft to the final draft is obtained by backtracking the accumulated convergence result set during the deduction process, including: Analyze the convergence result set of the deduction, extract the document versions before and after each round of deduction and the modification operations performed in each round, and obtain the document evolution sequence and the corresponding modification operation log; The feedback data output by each agent after each round of deduction is extracted from the set of convergence results of the deduction. The changes of the feedback data relative to the previous round are analyzed to identify the rounds in which the score is improved, the rounds in which the doubts are resolved, and the rounds in which the risks are eliminated. Extract the modification operations performed in the rounds where the rating was improved, the rounds where the doubts were resolved, and the rounds where the risks were eliminated from the modification operation log, and mark them as key modification operations; The key modification operations are arranged according to the round order in which they occur, forming a sequence of key modification operations from the first draft to the final draft.

[0012] In some possible implementations, the output of the final tender document and the corresponding explanatory document on the basis for its preparation includes: Based on the key modification operation sequence, perform modification operations on the initial draft of the tender document to generate the final version of the tender document; Extract the scoring attribution of the final bid document by the review simulation agent from the convergence result set to obtain the review preference response description; extract the resolved objections and corresponding modification operations from the simulation record to obtain the objection avoidance description; extract the eliminated risks and corresponding solutions from the compliance risk list to obtain the compliance compliance description. The review preference response instructions, question avoidance instructions, and compliance compliance instructions are organized according to the chapter structure of the final bid document, generating a compilation basis document, which is then linked and output with the final bid document.

[0013] In some possible implementations, the game strategy library is constructed through the following steps: Collect historical tender documents and their corresponding letters of inquiry and complaint records to obtain historical dispute text pairs; The historical disputed texts were aligned at the clause level, and the texts of the winning bid, the unsuccessful bid, and the challenged bids under the same bidding clause were extracted to obtain a clause alignment sample including the three types of texts. Extract the text of the winning bid and the text of the unwinning bid from the alignment sample of the terms, calculate the text differences between the two, extract the differentiated expression features, and obtain the differentiated winning strategy; The challenge text is extracted from the aligned sample of the terms, the attack logic of the challenge text is analyzed, the attack points are associated with the bidding terms, and the challenge pattern template is obtained. The differentiated winning strategies and the questioning pattern templates are associated according to the clause type to construct a game strategy library that includes the correspondence between clause type, differentiated strategies and questioning patterns.

[0014] As can be seen from the above technical solution, this application has the following beneficial effects: 1. This invention transforms unstructured bidding documents into a structured constraint space by constructing a bidding knowledge graph, solving the problem of easily overlooked clauses during manual parsing; it expands document compilation from single-time generation to multi-agent adversarial inference by initializing multiple agents to generate multi-dimensional feedback during the deduction process, addressing the lack of multi-dimensional evaluation methods; it achieves global optimization of score, risk, and compliance dimensions by encoding document state and agent feedback into state vectors and inputting them into a reinforcement learning model for iterative revision, solving the problem of local modifications failing to balance multiple objective conflicts; and it extracts key modification operation sequences by backtracking the deduction result set, solving the problems of untraceable compilation process and difficult-to-explain basis. The final output document includes the compilation basis for each step of modification, improving the document's credibility and interpretability.

[0015] 2. This invention makes explicit the complex constraints implicit in bidding documents through a bidding knowledge graph, providing a complete view for risk identification; it simulates a real bidding environment through three types of intelligent agents: review simulation, competitive confrontation, and compliance guardian, allowing documents to continuously evolve in adversarial simulation; it simultaneously achieves defense and offense through a game strategy library; it ensures that the final document is the result jointly recognized by the three intelligent agents through multi-dimensional convergence conditions; it eliminates invalid trial and error through key operation backtracking, providing the optimal path for reconstruction; and it establishes a link between each step of modification and intelligent agent feedback through the compilation of supporting documentation, making every key statement verifiable. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is an overall method flowchart provided for embodiments of the present invention; Figure 2 This is a diagram illustrating the bidding knowledge graph construction process in an embodiment of the present invention. Figure 3 This is a diagram illustrating the multi-agent adversarial simulation process in an embodiment of the present invention. Figure 4 This is a structural diagram of the reinforcement learning decision model in an embodiment of the present invention; Figure 5 This is a diagram showing the backtracking of key modification operation sequences and document reconstruction in an embodiment of the present invention. Detailed Implementation

[0018] The terms "first," "second," and "third," etc., used in this application specification, claims, and drawings are used to distinguish different objects, not to define a specific order.

[0019] In the embodiments of this application, the words "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0020] Research has found that existing technologies can only fill in gaps based on templates, and cannot quantitatively assess document quality or optimize it for multiple objectives. Furthermore, risk identification relies on human experience, and the compilation process lacks interpretability and traceability.

[0021] To address the aforementioned issues, this application provides a method for generating and validating structured documents for bidding and tendering scenarios: Terminology Explanation To ensure that those skilled in the art can accurately understand the technical solution of this invention, the key terms appearing in the text are explained below: Bidding Knowledge Graph: A knowledge base organized with a graph structure, including chapter nodes, clause nodes, entity nodes, and scoring point nodes extracted from bidding documents, as well as edge information such as constraints, mutual exclusions, and dependencies between nodes. The knowledge graph transforms unstructured bidding text into a structured, computable knowledge representation.

[0022] Intelligent agent: A computational entity that simulates a specific role, possessing an independent objective function and decision-making logic. The intelligent agents of this invention include a review simulation intelligent agent, a competitive adversarial intelligent agent, and a compliance guardian intelligent agent, respectively simulating the behavior of review experts, competitors, and compliance reviewers.

[0023] Feedback information: The agent outputs the evaluation results based on the current document state and its own objective function. The review simulation agent outputs a scoring report, the competing agent outputs a list of potential points of contention, and the compliance guardian agent outputs a list of compliance risks.

[0024] Adversarial simulation: This involves placing the bidding document in an evaluation environment with multiple agents, and continuously improving the document quality through multiple rounds of "evaluation and modification" iterations. In each round of simulation, the agents generate feedback, the decision model selects modification actions based on the feedback, and the document state is updated before proceeding to the next round.

[0025] State vector: A comprehensive quantitative representation of the current document state and the agent's feedback. It is composed of the document's text feature vector and numerical features such as the predicted rating, number of points of contention, and risk level, and serves as the input to the reinforcement learning decision model.

[0026] Reinforcement learning decision-making model: Based on a policy network trained using deep reinforcement learning, it takes a state vector as input and outputs a sequence of modification operations to be performed in the current state. The model learns how to select actions that maximize cumulative rewards through historical inference data.

[0027] Modification operation sequence: A set of one or more modification operations, each including action type (e.g., clause rewriting), target (e.g., clause position), and parameter value (e.g., new commitment value). A round of document revision is completed by executing the sequence output by the decision model in order.

[0028] Convergence result set: The collection of all data generated throughout the entire process from the start to the end of the simulation, including the document version for each round, the modifications performed, the agent's feedback data, reward values, etc.

[0029] Key Modification Operation Sequence: A set of modification operations obtained by backtracking from the convergence result set and associated with positive changes such as score improvement, resolution of doubts, and risk elimination. These operations are arranged in the order of execution to form the optimal optimization path from the first draft to the final draft.

[0030] Game strategy library: A knowledge base loaded by the competitive adversarial agents, including differentiated winning strategies (extracted from the comparison of winning and unwinning solutions) and questioning pattern templates (extracted from historical questioning texts) to simulate the questioning logic of competitors.

[0031] Explanatory document: An explanatory document accompanying the final version of the tender document, organized by chapter, explaining the triggering reasons, modification process and expected effects of each key modification, making the document preparation process explainable and traceable.

[0032] Example 1: Overview of the overall method.

[0033] This embodiment provides a method for generating and validating structured documents for bidding and tendering scenarios, such as... Figure 1 As shown, the method includes the following steps: Step S101: Parse the bidding documents, construct a bidding knowledge graph, and generate a draft of the bid document based on the bidding knowledge graph.

[0034] In this step, the tender document refers to the complete collection of documents issued by the tendering party, including technical specifications, commercial terms, scoring rules, and instructions to bidders, usually in PDF or Word format. The purpose of parsing is to transform unstructured text content into a structured knowledge representation. The tender knowledge graph is a knowledge base organized in a graph structure, where nodes represent entities such as chapters, clauses, technical parameters, qualification requirements, and scoring points in the tender document, and edges represent constraints, dependencies, and attribution relationships between entities. Generating a preliminary draft of the tender document based on this graph means calling upon pre-built basic material libraries to populate the corresponding clause nodes according to the chapter node order in the graph, forming a preliminary tender document.

[0035] Step S102: Within the rule space defined by the bidding knowledge graph, initialize multiple agents with different objective functions. These agents are used to generate feedback information in each round of deduction.

[0036] In this step, the rule space refers to the set of constraints comprised of all node attributes and edge relationships in the bidding knowledge graph. Any modifications to the bidding document must be made within the scope defined by these rules. For example, technical parameters cannot exceed the range specified in the bidding documents, and the promised delivery deadline cannot be later than the latest deadline required by the bidding documents. An agent is a computational entity that simulates a specific role. Each agent has an independent objective function, which defines the criteria for evaluating the quality of the bidding document. Initializing agents refers to instantiating these agents and loading them with the corresponding data resources, enabling them to generate feedback information during the deduction process. The feedback information refers to the evaluation results output by the agent based on the current document state and its own objective function, such as scores, lists of points of contention, and risk lists.

[0037] Step S103: Perform multiple rounds of iterative adversarial simulations on the initial draft of the tender document until the preset convergence condition is met, and obtain the simulation convergence result set.

[0038] This step is the core of the method. Adversarial simulation involves placing the initial draft of the bidding document under the "siege" of multiple agents. Each round of simulation includes: First, encoding the current document state (the current version of the bidding document) and the feedback information generated by each agent into a state vector. This state vector is a comprehensive quantitative representation of the document's current health status. Then, this state vector is input into a reinforcement learning decision model, which outputs a sequence of modification operations to be performed in the current state based on historical learning experience. Next, the document is revised according to this sequence of modification operations, generating an updated document state. Finally, the updated document state is used as the input for the next round of simulation, and the above process is repeated. This process is repeated until a preset convergence condition is met, i.e., the document state no longer changes significantly or has reached the expected quality standard. The simulation convergence result set refers to the collection of all intermediate document versions, modification operation records, agent feedback data, and other information generated during the simulation process.

[0039] Step S104: Based on the backtracking of the convergence result set accumulated during the deduction process, obtain the sequence of key modification operations performed from the first draft to the final draft.

[0040] In this step, backtracking refers to analyzing the convergence results of the deduction process in reverse to identify the modifications that truly played a crucial role in improving document quality. Specifically, by analyzing changes in agent feedback data before and after each round of deduction—such as whether scores improved, whether points of contention decreased, and whether risks were eliminated—modifications that brought about positive changes are identified and arranged in the order of their execution in each round, forming a sequence of key modification operations. This sequence represents the optimal optimization path from the initial draft to the final draft.

[0041] Step S105: Reconstruct the initial draft of the tender document according to the key modification operation sequence, and output the final version of the tender document and the corresponding preparation basis explanation document.

[0042] In this step, reconstruction refers to strictly following the operational sequence recorded in the key modification operation sequence, re-executing these modification operations from the initial draft to generate the final version of the tender document. Simultaneously, during the reconstruction process, each modification operation is linked to the feedback information from the agent that triggered it, forming a documentation document explaining the basis for the compilation. This documentation document is organized according to the chapter structure of the final tender document, and each key modification is accompanied by explanatory content such as the reason for the modification, the triggering conditions, and the expected effects.

[0043] This system enables fully intelligent processing from parsing the tender documents to generating the final bid document. The specific implementation methods for each step are explained in detail below.

[0044] Example 2: Construction and draft generation of a bidding knowledge graph.

[0045] like Figure 2As shown, the construction of the bidding knowledge graph includes the following sub-steps: Step S201: Extract structured clause data.

[0046] First, chapter identification is performed. Chapter boundaries are determined using a title pattern library and regular expression matching. Untitled text is clustered to obtain a chapter structure sequence. Next, each chapter is divided into clauses, segmented into clause units based on numbering patterns. Unnumbered paragraphs are grouped into natural paragraphs, forming a chapter-clause hierarchical structure. Finally, specific extractions are performed: technical specifications are extracted using syntactic analysis to obtain parameter entities (names, values, units, comparison relationships) and delivery deadlines; business clauses are extracted using keyword matching to obtain qualification requirement entities; and scoring rules are parsed to parse scoring item names, score ranges, and scoring conditions to obtain preliminary scoring points. The extraction results are then linked by clause to form structured data with clauses as units.

[0047] Step S202: Entity relation extraction.

[0048] Based on the extracted entities, the constraints (such as parameter thresholds), mutual exclusion (such as choosing between two technical solutions), and dependencies (such as parameter-dependent qualifications) between entities are identified. Relationship extraction can be performed using rules or pre-trained models, and the results are stored in the form of triples (entity A, relation type, entity B).

[0049] Step S203: Quantification of scoring points.

[0050] Each scoring condition in the scoring rules is transformed into a quantifiable scoring point, including the scoring item name, score range, scoring condition, supporting material requirements, and quantification rules (unit of measurement, score increment, upper limit), forming a set of scoring points.

[0051] Step S204: Construct a bidding knowledge graph.

[0052] Using chapters as parent nodes and clauses as child nodes, create chapter nodes (number, title, level), clause nodes (number, text, type), entity nodes (name, type, value), and scoring point nodes (name, score, condition, quantification rule). Connect the nodes with edges such as "belongs to," "originates from," "based on," and "requirements," and add the relationships extracted in step S202 as "constraints," "mutually exclusive," and "dependencies" edges to the graph, forming a complete knowledge graph including nodes, attributes, and edges.

[0053] Step S205: Generate the initial draft of the tender document.

[0054] Generate a document framework (table of contents, titles, and blank clauses) according to the chapter order of the tender document. For each clause node, retrieve a matching response template from the basic resource library (the resource library includes standard response templates, technical solutions, qualification descriptions, etc., labeled with applicable clause types and keywords) based on its type and associated entities. Match the optimal resource through semantic similarity, replace the placeholders with the actual parameter values ​​required by the tender document, and fill them into the corresponding clause positions to form a draft tender document including all chapter clauses, which will be further optimized.

[0055] Example 3: Initialization and adversarial simulation of multi-agent systems.

[0056] like Figure 3 As shown, the multi-agent adversarial simulation involves three types of core agents: review simulation agent, competitive adversarial agent, and compliance guardian agent.

[0057] (a) Initialization of the agent.

[0058] Step S301: Instantiate the review simulation agent, load historical bidding behavior data and the terms, rules and scoring details in the bidding knowledge graph.

[0059] The goal of the review simulation agent is to simulate the evaluation behavior of real review experts, scoring bid documents and pointing out their strengths and weaknesses. Its initialization includes: Load historical bid evaluation data. This data comes from the bid evaluation records of historical projects, including expert scoring sheets for each bid, review comments, and reasons for deductions. By analyzing this data, we can learn the experts' scoring preferences; for example, experts particularly value certain technical indicators, certain wording styles are likely to earn points, and certain common errors are likely to result in deductions.

[0060] Load the terms, rules, and scoring details from the bidding knowledge graph. The terms and rules define the basic constraints that bid responses must meet, such as parameter ranges and delivery deadlines; the scoring details define the weight and scoring conditions for each scoring item. The review simulation agent needs to accurately understand these rules in order to generate a scoring report that meets the requirements of this bidding process.

[0061] After loading, the review simulation agent possesses the following capabilities: It can take a tender document as input, evaluate it line by line against the scoring criteria, and output a scoring report including strengths (which clauses are well-answered and likely to receive high scores) and potential deductions (which clauses are deficient and may result in point deductions). The scoring report also includes a comprehensive score prediction value, calculated based on historical evaluation data and the current document's response quality. The score prediction value is calculated as follows: first, a satisfaction score is calculated for each scoring point based on the response content; then, it is calibrated by combining the score distribution of similar clauses in historical evaluation data; finally, the scores of all scoring points are summed to obtain the comprehensive score prediction value.

[0062] Step S302: Instantiate the competitive adversarial agent, and load the game strategy library including historical questioning data and differentiated strategies, as well as the terms, rules and scoring details in the bidding knowledge graph.

[0063] The goal of the competitive adversarial agent is to simulate the behavior of its competitors and identify potential weaknesses in the bidding documents that could be questioned. Its initialization includes: Load the game strategy library. The construction method of this library will be described in detail in Example 4. Its core content includes two parts: first, the questioning pattern template, which is derived from the attack logic extracted from historical questioning texts; second, the differentiated winning strategy, which is derived from the comparative analysis of winning and unwinning solutions.

[0064] Load the terms, rules, and scoring details from the bidding knowledge graph. Similar to the review simulation agent, the competitive adversarial agent also needs to understand the bidding requirements, but with a different purpose—to identify loopholes in the bid responses that could be exploited by competitors.

[0065] After loading, the competitive adversarial agent possesses the following capabilities: Given a tender document, it can simulate the perspective of a competitor, scan the document for potentially questionable sections, and output a list of potential points of contention, including the location and reasons for any questioning clauses. For example, a slight deviation in the description of a technical parameter from the tender requirements might be questioned as "not meeting the tender requirements"; inconsistencies in the content of a commitment across different chapters might be questioned as "self-contradictory"; insufficient supporting documentation for a scoring point might be questioned as "unable to prove that the scoring conditions are met".

[0066] Step S303: Instantiate the Compliance Guardian intelligent agent, and load the regulatory case data and the terms, rules and scoring details in the bidding knowledge graph.

[0067] The goal of the Compliance Guardian intelligent agent is to ensure that bidding documents comply with the mandatory requirements of relevant laws, regulations, and bidding documents. Its initialization includes: Load legal and regulatory case data. This data includes relevant laws and regulations concerning bidding and tendering, industry standards, historical cases of rejected bids, and administrative penalty cases. By analyzing this data, we can identify which types of violations are likely to lead to bid rejection or penalties.

[0068] Load the terms, rules, and scoring details from the bidding knowledge graph. The compliance guardian intelligent agent focuses on the mandatory constraints and veto conditions, which are the bottom-line requirements that must be met; any violation will directly result in the bid being invalid.

[0069] After loading, the Compliance Guardian AI agent possesses the following capabilities: It can input a bidding document, scan for content that may violate laws, regulations, or mandatory requirements of the bidding documents, and output a compliance risk list including the violation clauses and their risk levels. The risk levels are divided into three levels: "Significant Risk" (potentially leading to bid rejection), "General Risk" (potentially resulting in point deductions or clarification requirements), and "Informative Risk" (suggests optimization but will not lead to serious consequences).

[0070] Step S304: The completed loading of the review simulation agent, the competitive adversarial agent, and the compliance guardian agent is used as the initial set of agents for initiating multiple rounds of iterative adversarial simulation.

[0071] At this point, all three agents are ready, and in each subsequent round of deduction, they will independently generate their own feedback information based on the current document state.

[0072] (ii) Convergence conditions for adversarial inference.

[0073] During adversarial simulations, explicit convergence criteria need to be set to determine when the simulation can terminate. In this embodiment, the convergence criteria are directly linked to the feedback from the three agents; the simulation will only stop when all three dimensions simultaneously meet the convergence criteria.

[0074] Step S305: Obtain the score prediction values ​​output by the review simulation agent in multiple consecutive rounds. When the change range of the score prediction values ​​is less than a preset threshold and higher than a preset pass line, it is determined that the first convergence condition is met.

[0075] This condition means that the score has stabilized, there is no longer significant room for improvement, and it has reached an acceptable quality level. The specific settings are as follows: A preset threshold is used to judge the magnitude of change. For example, it can be set to "the change in the predicted score for three consecutive rounds is less than one percent". This means that in the most recent rounds of simulation, the score has basically stopped rising, indicating that the document is close to its optimal level in terms of scoring.

[0076] A preset passing score is used to determine whether the basic requirements have been met; for example, it could be set to "predicted score higher than 80". If the score is stable but consistently below the passing score, it indicates a fundamental problem with the document itself, which may not be solvable by simple deduction. In this case, manual intervention should be triggered instead of automatic convergence.

[0077] If both of these sub-conditions are satisfied, then the first convergence condition is satisfied.

[0078] Step S306: Obtain a list of potential points of contention from the competitive adversarial agent's output over multiple rounds. When the list of points of contention is empty or the contents of the list are the same as the historical output, it is determined that the second convergence condition is met.

[0079] This condition means that no new points of contention have been found, and all previously identified points of contention have been resolved. The specific settings are as follows: An empty list of points of contention means that, in the current document state, the competing adversarial agents have conducted a comprehensive scan and believe that there is nothing to question, which is the ideal convergence state.

[0080] If the list is not empty, but its contents are completely identical to the historical outputs from multiple consecutive rounds—meaning the list of points of contention has remained unchanged for several rounds—it indicates that while points of contention exist, subsequent modifications have failed to effectively resolve them. In this case, it's necessary to analyze the reasons: are the points of contention truly unsolvable, such as contradictions in the bidding requirements themselves, or is the modification strategy ineffective? In convergence determination, this situation also satisfies the second convergence condition because further deduction yields no improvement.

[0081] Step S307: Obtain the compliance risk list output by the compliance guardian intelligent agent in multiple consecutive rounds. When there are no major risk items in the compliance risk list or all risk items are in an eliminated state, it is determined that the third convergence condition is met.

[0082] This condition means that there are no significant risks that could lead to the rejection of the bid, and all identified risks have been addressed. The specific settings are as follows: "No major risk items" means that the compliance risk list does not include items marked as "major risks". "General risks" and "noticeable risks" are permissible, for example, some suggested optimizations may not be fully resolved due to cost or objective limitations.

[0083] All risk items are now in an eliminated state. This means that if a risk item once appeared in the list, but the corresponding modification operations have been executed in subsequent rounds, and the document version after execution has been confirmed as eliminated by the Compliance Guardian's intelligent re-examination.

[0084] Step S308: When the first convergence condition, the second convergence condition, and the third convergence condition are simultaneously satisfied, it is determined that the deduction satisfies the preset convergence condition, and the deduction process is terminated.

[0085] By combining the above three conditions, the final converged document version is guaranteed to achieve a satisfactory state in terms of score, resistance to questioning, and compliance.

[0086] Example 4: Construction of a game strategy library.

[0087] The game strategy base is the core knowledge resource of competitive adversarial agents, and its construction is based on in-depth mining of historical bidding dispute data. Specifically, it includes the following sub-steps: Step S401: Collect historical tender documents and their corresponding letters of inquiry and complaint records to obtain historical dispute text pairs.

[0088] From the historical bidding database, projects that have been subject to objections or complaints are selected. For each project, three types of documents are collected: first, all bid documents submitted by bidders; second, the response documents to objections issued by the tendering party (including the original text of the objection letter); and third, the complaint handling decision issued by the relevant regulatory department (including the reasons for the complaint and the handling result). The bid documents for the same project are linked with the objection letter / complaint record to form historical dispute text pairs.

[0089] Step S402: Perform clause-level alignment on the historical dispute text pairs, extract the texts of the winning bid, the unsuccessful bid, and the challenged bid under the same bidding clause, and obtain a clause alignment sample including the three types of texts.

[0090] The core of this step is to achieve clause-level alignment across multiple texts. The specific method is as follows: First, based on the tender documents, they are broken down into clause units (corresponding to the clause nodes in the aforementioned tender knowledge graph). For each tender clause, the corresponding paragraph needs to be found in the tender documents and the objection documents.

[0091] The alignment process employs a two-stage strategy: Phase 1: Preliminary Alignment Based on Numbering. Extract the clause numbering system from the tender documents (e.g., "3.1", "Article 4"), and search for texts with the same number in the tender documents and challenge documents as candidate alignment pairs.

[0092] The second stage: Semantic-based secondary alignment. For clauses without numbers or with inconsistent numbers, a sentence BERT model is used to encode the tender clause text into vectors. Similarly, the tender document paragraphs and the challenge text paragraphs are also encoded into vectors. The cosine similarity between the tender clause vector and the tender document paragraph vector is calculated. Those with a similarity greater than a preset threshold (e.g., 0.75) are considered alignment candidates. The alignment between the tender clause and the challenge text is also calculated.

[0093] For each alignment candidate, perform two-way verification: treat the bid paragraph as a query, search for the most similar clause in the tender document, and if the retrieved clause is consistent with the original tender clause, then confirm the alignment.

[0094] After alignment, for the same bidding clause, three types of texts can be obtained: the winning bid text (the content of the successful bidder's response), the unsuccessful bid text (the content of the responses of other bidders), and the objection text (the reasons for objecting to the clause). These three types of texts together constitute a clause alignment sample.

[0095] Step S403: Extract the text of the winning bid and the text of the unwinning bid from the clause alignment sample, calculate the text differences between the two, extract the differentiated expression features, and obtain the differentiated winning strategy.

[0096] The goal of this step is to learn from winning bids how to write proposals that are more likely to succeed. Specific methods include: A comparative analysis was conducted on the texts of the winning and unsuccessful bids under the same bidding terms. First, sentence-level alignment was performed to obtain corresponding sentence pairs. Then, for each sentence pair, the following difference indicators were calculated: Edit distance: Normalized edit distance, which measures the degree of literal difference.

[0097] Semantic similarity: Cosine similarity of sentence vectors is calculated using BERT.

[0098] Keyword differences: Extract the keyword sets of the two sentences and calculate the Jaccard distance.

[0099] Based on the difference indicators, the differences can be divided into the following types: word differences (such as synonym substitution, terminology usage), sentence structure differences (such as active voice vs. passive voice), information completeness differences (whether explanations and citations are included, whether evidence is cited), and structural differences (whether lists, tables, and charts are used).

[0100] The analysis revealed frequent patterns of difference across multiple comparisons. For instance, a large sample size analysis showed that winning bids tended to use a combination of flowcharts and textual descriptions, while unsuccessful bids relied more on purely textual descriptions. Winning bids also tended to state parameters as "no less than X" rather than "approximately X," and added "see supporting documentation X" after their commitments. These frequent patterns of difference were extracted as differentiated winning strategies.

[0101] Each differentiation winning strategy includes the following structured information: strategy type, applicable clause type, specific wording template, confidence score (when the confidence score is greater than a set threshold (e.g., 1.5), the pattern is considered a valid differentiation winning strategy. For example, if a pattern appears in 60% of the winning bids and 30% of the unsuccessful bids, the confidence score is 2.0, exceeding 1.5, and is adopted. If the confidence score is less than 0.67, it indicates that the pattern is more common in the unsuccessful bids and should be recorded as a 'wording to be avoided' in the strategy library for reference during document optimization), and application example.

[0102] Step S404: Extract the challenge text from the clause alignment sample, analyze the attack logic of the challenge text, associate the attack points with the bidding clauses, and obtain the challenge pattern template.

[0103] The goal of this step is to learn from historical critiques how others have attacked. Specific methods include: Syntactic analysis and semantic understanding are performed on the challenged text to extract the core logic of the challenge. Common challenge logics include: parameter non-compliance ("Parameter X in the tender document is A, while the tender document requires it to be no less than B, which is not actually met"), inconsistent commitments ("The tender document commits to X in Chapter 3 and to Y in Chapter 5, which is contradictory"), missing evidence ("The tender document claims to meet condition X, but fails to provide supporting documentation Y as required"), vague descriptions ("The description of X in the tender document is too vague to determine whether the requirement is met"), and questionable qualifications ("The qualification certificate provided in the tender document has expired / is not of the required type"), etc.

[0104] For the identified objection logic, it is associated with the targeted tender terms to form objection pattern templates. Each objection pattern template includes the following structured information: applicable clause type, objection triggering conditions (what situations would lead to an objection), objection expression paradigm (how an objection is usually expressed), severity assessment (what consequences the objection might cause), and mitigation suggestions (how to modify the template to avoid such objections).

[0105] For example, for technical parameter clauses, the following challenge template is formed: "When the deviation of the parameter value in the bid response from the tender requirements exceeds the preset threshold and no reasonable explanation is given, it may be challenged as 'not meeting the tender requirements'. The challenge statement is usually 'The tender document requires the parameter value to be no lower than the required value, and the parameter value provided in the bid document is lower than the requirement, which constitutes a substantial deviation'." Step S405: Associate the differentiated winning strategy with the questioning mode template according to the clause type to construct a game strategy library that includes the correspondence between clause type, differentiated strategy and questioning mode.

[0106] Finally, the differentiation strategies and challenge templates are organized and linked according to clause type. For example, for "technical parameter" clauses, the strategy library includes both "how to write better" (differentiation strategy) and "how it might be attacked" (challenge template). This linkage enables the competitive adversarial agent to conduct a comprehensive analysis from both "positive learning" and "reverse defense" perspectives when evaluating documents.

[0107] Once the game strategy library is built, it will be loaded into the competitive adversarial agent as the core basis for generating a list of potential points of contention.

[0108] Example 5: Encoding of State Vectors and Reinforcement Learning Decision Model.

[0109] (a) Encoding of state vectors.

[0110] At the beginning of each round of inference, the current document state and the feedback information from each agent need to be encoded into a state vector, which serves as the input to the reinforcement learning decision model. The encoding process for the state vector is as follows: Step S501: Obtain the scoring report of the review simulation agent output, the list of potential questionable points of the competitive agent output, and the list of compliance risks of the compliance guardian agent output in the current round.

[0111] The scoring report includes the following key data: overall score prediction (a numerical value), a list of strengths (including the location markers of several clauses), and a list of potential deductions (including the location markers of several clauses and the expected deduction amount).

[0112] The list of potential points of contention includes: the total number of points of contention (a numerical value), the clause location of each point of contention, the type of contention (such as parameter non-compliance, inconsistent commitments, etc.), and the intensity of contention (three levels: high, medium, and low).

[0113] The compliance risk list includes: the total number of risk items (a numerical value), the number of major risk items (a numerical value), the clause location of each risk item, the risk level, and the risk type (such as expired qualifications, excessive parameters, etc.).

[0114] Step S502: Combine the predicted score, the number of potential deductions, the number of potential points of contention in the potential questioning list, and the risk level data in the compliance risk list in the scoring report with the text features of the current document to generate a multi-dimensional state vector.

[0115] The specific encoding method is as follows: First, extract numerical features and perform normalization: Predicted score: Divide by 100 to map to the range of zero to one.

[0116] Number of potential deductions: Divide by the maximum possible number of deductions. This maximum value is twice the number of all scoring points in the bidding knowledge graph (because each scoring point may have a maximum of two deductions: one is a technical deduction due to incomplete content, and the other is a proof deduction due to insufficient evidence, so it is doubled).

[0117] Total number of points of contention: Divide by the maximum possible number of points of contention. This maximum value is twice the total number of clauses (because each clause may generate a maximum of two points of contention: one is whether the content of the response itself meets the requirements, and the other is whether the supporting materials are sufficient, so it is doubled).

[0118] Total number of risk items: Divide by the maximum possible number of risk items, which is the total number of clauses (each clause can generate at most one risk item, because multiple violations of the same clause are usually combined into one risk item record).

[0119] Number of major risk items: Divide by the maximum possible number of major risk items. The maximum value is the number of mandatory clauses (only mandatory clauses can generate major risks, so the upper limit of the number of major risk items is equal to the total number of mandatory clauses).

[0120] Secondly, extract the textual features of the document. The entire document is encoded, and a pre-trained language model (such as BERT-base-Chinese) is used to extract document-level semantic vectors. Specifically, the document is segmented into sentences, and each sentence is input into the model to obtain a sentence vector (768 dimensions). All sentence vectors are then averaged and pooled to obtain document-level vectors. If the document length exceeds the model's maximum input length (512 tokens), it is segmented before pooling.

[0121] Then, the normalized numerical features (five dimensions) are concatenated with the document vector (768 dimensions) to form the final state vector with 773 dimensions.

[0122] This state vector includes both the semantic information of the document itself and the quantitative evaluation results given by the three agents from different dimensions, providing the decision-making model with a comprehensive environmental perception.

[0123] (ii) Construction of reinforcement learning decision-making model.

[0124] like Figure 4 As shown, the construction of a reinforcement learning decision model includes the following sub-steps: Step S503: Predefine the document modification operation as a set of actions.

[0125] The action set is the collection of all modification operations that a reinforcement learning model can perform. Based on the characteristics of bidding documents, the following types of modification operations are predefined: Clause rewriting action: Rewrites the text content of a specified clause. This action includes parameters: clause location identifier, rewriting style (e.g., "more professional," "more concise," "more detailed"). The rewriting operation can invoke a pre-trained language generation model to generate new text based on the current clause content and the specified style.

[0126] Supplementary Supporting Documents: Provides supporting documentation for a specified clause. This action includes parameters: clause location identifier, and document type (e.g., "Certificate of Qualification," "Performance Record," "Technical Specification"). The Supplementary Supporting Documents action inserts a link or explanatory text to the supporting documentation at the appropriate location in the document.

[0127] Adjusting the commitment value action: Modifying the specific numerical value of the commitment in the clause. This action includes parameters: clause location identifier, parameter name, and new commitment value. The adjustment must be performed within the rule space defined by the bidding knowledge graph, meaning the new commitment value cannot violate the mandatory constraints of the bidding documents.

[0128] In addition to the three core actions mentioned above, other auxiliary actions can also be defined, such as deleting redundant expressions, adjusting wording and tone, and merging duplicate clauses.

[0129] Each action defines input parameters and output effects. The size of the action set depends on the types of operations that can be performed and the number of operable clauses. For example, if a document contains two hundred clauses, and each clause has five operable actions, then the action space size is one thousand.

[0130] Step S504: Based on the changes in document status after performing the modification operations in the action set, quantify the combined changes of score improvement, reduction of questionable points, and decrease in risk level into reward value.

[0131] The reward value is a benchmark for measuring the quality of an action in reinforcement learning. In this embodiment, the reward value is directly linked to the feedback changes of the three agents.

[0132] Specifically, suppose that in a certain round of simulation, a sequence of modification operations is executed, and the document changes from state to state. Become a state Comparison status and Based on the agent's feedback, calculate the following changes: Rating improvement: New rating prediction minus old rating prediction.

[0133] Reduction in points of contention: Total number of old points of contention minus total number of new points of contention.

[0134] Risk reduction amount: Total number of old risk items minus total number of new risk items.

[0135] The amount of major risks eliminated: the number of old major risk items minus the number of new major risk items.

[0136] Because these changes have different dimensions, they need to be normalized: Normalization of score improvement: Divide by 100 to get a value in the range of zero to one.

[0137] Normalization of the reduction in questionable points: Divide by the maximum possible reduction in questionable points (the maximum possible reduction in questionable points is preset to 20, which is based on statistical analysis of historical bidding data: in the past 1,000 bidding projects, the maximum number of reductions in questionable points in a single round of simulation did not exceed 20, and taking this value as the normalization denominator can cover more than 99% of the actual situation), to obtain a value in the range of zero to one.

[0138] Risk reduction normalization: Divide by the reduction in the maximum possible risk item (similarly, the default is ten) to obtain a value in the range of zero to one.

[0139] Normalization of major risk elimination: Divide by the reduction of the most likely major risk item (similarly, the default is five) to obtain a value in the range of zero to one.

[0140] Reward Value It can be designed as a weighted sum of the above normalized changes:

[0141] in: The normalized value of the rating improvement is calculated as (new rating prediction value - old rating prediction value) / 100, with a value range of [-1, 1]. The normalized value of the reduction in the number of points of contention is calculated as (total number of old points of contention - total number of new points of contention) / 20, with a value range of [-1, 1]. The normalized value of the reduction in risk items is calculated as (total number of old risk items - total number of new risk items) / 10, with a value range of [-1, 1]. The normalized value of the reduction in major risk items is calculated as (number of old major risk items - number of new major risk items) / 5, with a value range of [-1, 1]. Weighting coefficients, with default values ​​of 1.0, 2.0, 2.0, and 5.0, are used to balance the importance of different optimization objectives; among which The highest value reflects the priority given to eliminating major risks; The values ​​of denominators 20, 10, and 5 are based on the statistical analysis of 1,000 historical bidding projects. In a single round of simulation, the maximum reduction in the number of questionable points does not exceed 20, the reduction in the number of risk items does not exceed 10, and the reduction in the number of major risk items does not exceed 5. Using the above denominators can normalize each change to the interval [-1, 1].

[0142] It is important to note that if a modification results in a negative effect (such as a drop in rating or an increase in points of criticism), the corresponding change will be negative, and the reward value may also be negative, thus penalizing the action.

[0143] Step S505: During the system cold start phase, the reinforcement learning decision model uses a stochastic strategy or a rule-based heuristic strategy for initial inference. As the number of inference rounds increases, the actual generated data (states, actions, rewards) is stored in an experience pool. When the amount of data in the experience pool reaches a preset threshold (e.g., 10,000 data points), model training begins. Training is conducted offline, with model parameters periodically sampled from the experience pool to update. Inference can still be performed before the model is fully trained, but the optimization effect may not be as good as that of a fully trained model.

[0144] This step uses a deep reinforcement learning algorithm (such as the proximal policy optimization algorithm) to train the policy network. The structure of the policy network is as follows: Input layer: Receives state vectors, with a dimension of 773.

[0145] Hidden layers: Two fully connected layers are used, each with 512 neurons, and the activation function is ReLU. To capture the temporal dependencies in sequential decisions, a long short-term memory network layer can also be added to the hidden layers.

[0146] Action Output Layer: Outputs the action probability distribution, with a dimension equal to the size of the action space. This layer uses the Softmax activation function and outputs the probability of selecting each action in the current state.

[0147] Value output layer: outputs state value estimates, with one dimension, and uses a linear activation function.

[0148] During training, state-action-reward tuples are sampled from historical projection data. The model is updated after a certain number of projection rounds (e.g., ten rounds). During the update, the advantage function is first calculated using generalized advantage estimation, then the policy network loss and value network loss are calculated, and finally, the network parameters are updated through backpropagation. The specific calculation formulas are as follows: The advantage function (A(s,a)) is calculated using generalized advantage estimation:

[0149] in: This is called the timing difference error; : The reward value obtained in step t; :state The value estimate is output by the value network; Discount factor, with a value of 0.99, is used to balance the importance of current and future rewards; The GAE parameter, with a value of 0.95, is used to control the trade-off between bias and variance. (T): The number of steps of modification operations in the current round of simulation.

[0150] The policy network loss function uses a pruned proxy objective:

[0151] in: , where is the ratio of the probabilities of the new and old strategies; The current policy network is in state. Select action The probability of; The probability given by the old policy network before the update; The value of the dominance function at time step t; The pruning hyperparameter, with a value of 0.2, is used to limit the magnitude of policy updates and prevent destructive updates. :Will Limited to Functions within an interval.

[0152] The loss function for the value network is the mean squared error:

[0153] in: Value network for state The estimated value; Temporal difference objective, used to supervise the training of value networks; Discount factor, taken as 0.99 as before.

[0154] The total loss function combines policy loss, value loss, and entropy regularization:

[0155] in: Value loss coefficient, with a value of 0.5, is used to balance strategy loss and value loss; Entropy regularization coefficient, with a value of 0.01, is used to encourage policy exploration and avoid getting trapped in local optima too early;

[0156] in It is the set of all possible actions.

[0157] (State): Represents the current document's state vector. It is a multi-dimensional vector (e.g., 773-dimensional) obtained by concatenating the textual features of the current document with feedback information from three agents (review simulation, competitive adversarial, and compliance guardian) in each iteration. It describes the document's health status at the current moment.

[0158] (Action): Represents the modification operation that the agent (i.e., the reinforcement learning decision model) can perform. Examples include: rewriting clauses, supplementing proofs, adjusting commitment values, etc. Each action corresponds to a specific document modification instruction.

[0159] The policy function represents the strategy in a given state. Next, the model selects the action to execute. The probability of.

[0160] : The natural logarithm of the probability.

[0161] Policy entropy measures the randomness of a policy. The higher the entropy value, the more the model tends to explore different actions; the lower the entropy value, the more the model tends to choose the deterministic optimal action.

[0162] As the number of iterations increases and more data accumulates, the decision-making ability of the policy network continuously improves. After training, the reinforcement learning decision model has the following capabilities: inputting the state vector of the current document, it outputs a sequence of modification operations (i.e., a series of actions and their parameters), which is predicted to maximize the cumulative reward.

[0163] In actual simulations, each time the model is invoked, an optimized modification plan for the current document can be obtained.

[0164] Example 6: Backtracking of key modification operation sequences and document reconstruction.

[0165] (a) Backtracking of key modification operation sequence.

[0166] like Figure 5 As shown, the backtracking of the key modification operation sequence includes the following sub-steps: Step S601: Analyze the convergence result set of the deduction and extract the document evolution sequence and modification operation log.

[0167] Extract the document versions (V0, V1, …, Vn) before and after each round from the convergence result set, as well as the modification operations performed in each round. Record the action type, effect clause, parameter value and content comparison of each operation to obtain a complete document evolution sequence and modification operation log.

[0168] Step S602: Analyze feedback changes and mark positive optimization rounds.

[0169] Compare the agent's feedback data before and after each round of simulation: if the score improves by more than a preset threshold (e.g., 0.5 points), mark it as "Round of Score Improvement"; if the number of points of contention disappearing is greater than zero, mark it as "Round of Contention Resolution"; if the number of risk items disappearing is greater than zero, mark it as "Round of Risk Elimination". Record the specific changes for each round.

[0170] Step S603: Extract key modification operations.

[0171] Extract all modification operations executed in the marked round from the modification operation log, and use them as candidate key modification operations.

[0172] Step S604: Sort to form a sequence of key modification operations.

[0173] The extracted candidate operations are arranged according to the original round order and the execution order within the round to obtain the key modification operation sequence from the first draft to the final draft.

[0174] (ii) Dependency verification and document refactoring.

[0175] Step S604A: Dependency verification.

[0176] Dependency detection uses the following methods: Dependency detection based on operation objects: For each operation Oi in the sequence, record its operation clause position Pi. Traverse subsequent operations Oj (j>i). If the operation clause position Pj of Oj is the same as Pi, then Oi and Oj have a clause-level dependency, and Oi must be executed before Oj.

[0177] Content-based dependency detection: For each modified clause text of operation Oi, entity recognition is performed to extract the identifier of the cited supporting material, such as 'See Attachment X for details'. Subsequent operations Oj are then traversed. If Oj involves deleting or modifying the supporting material, then there is an evidence-level dependency between Oi and Oj.

[0178] Semantic-based dependency detection: For dependencies that cannot be determined by the above rules, a pre-trained model is used to calculate the semantic relevance of document fragments before and after the operation. If the relevance exceeds the threshold, a possible dependency relationship is indicated, and the user confirms or the system executes in the default order.

[0179] Step S605: Perform the refactoring sequentially to generate the final document.

[0180] Starting with the initial draft V0, modifications were executed strictly according to the verified operation sequence. After each operation, checks were performed on commitment consistency, relevance of supporting materials, and contextual references to ensure the logical consistency of the revised document. Once all operations were completed, the final tender document V_final was obtained.

[0181] Step S606: Extract rating attributions and generate reviewer preference response descriptions.

[0182] Extract the scoring reasons of the review simulation agent for the final document from the convergence result set, and organize them into review preference response explanations to reflect how the document responds to the expert scoring preferences.

[0183] Step S607: Extract the records of resolving doubts and generate an explanation of how to avoid doubts.

[0184] The resolved points of contention and their corresponding modifications were identified from the simulation records and compiled into a statement of contention avoidance, demonstrating that the document has specifically avoided potential attacks.

[0185] Step S608: Extract risk elimination records and generate compliance statements.

[0186] Identify the eliminated risk items and their solutions from the compliance risk list, compile them into a compliance compliance statement, and prove that the documentation meets all mandatory requirements.

[0187] Step S609: Organize and compile the supporting documentation and associate it with the output.

[0188] The above three sections are organized according to the chapter structure of the final document to form a compilation basis document. This document is then linked at the chapter level with the final tender document and output accordingly. This document provides an explainable basis for each key modification, for example: "Chapter 3 Technical Solution: In response to the 'insufficient operability of the solution' pointed out by the review simulation agent, a flowchart was added in the third round of simulation; in response to the 'parameter ambiguity' questioned by the competitive adversarial agent, the parameter range was clarified in the fifth round of simulation; and the compliance guardian agent confirmed that there is no compliance risk." Example 7: Application scenario.

[0189] To better understand the practical application of this invention, a specific application scenario example is given below.

[0190] A large infrastructure project has issued a tender notice. The tender documents, totaling over 300 pages, include technical specifications, commercial terms, and scoring rules. A construction company wishing to participate in the bidding uses the system described in this invention to assist in preparing its tender documents.

[0191] Step 1: The user uploads the tender document (PDF format) to the system. The system activates the tender document parsing module, automatically identifying the chapter structure, extracting technical parameters such as concrete strength grade and steel reinforcement specifications, delivery deadlines such as "completion within 360 days after contract signing," qualification requirements such as "Special Grade Qualification for General Contracting of Building Construction," and scoring details such as "Construction organization design plan accounts for 30 points, of which the rationality of the schedule plan accounts for 10 points." The system constructs a tender knowledge graph including over 200 clause nodes, over 500 entity nodes, and over 800 relationships, and based on this graph, calls upon the basic material library to generate a tender document including the initial draft content.

[0192] Step 2: The system initializes the review simulation agent, the competitive adversarial agent, and the compliance guardian agent. The review simulation agent loads evaluation data of similar projects in the region over the past three years and learns the scoring preferences of local review experts; the competitive adversarial agent loads a game strategy library, which includes all challenging cases in the construction engineering field over the past five years; the compliance guardian agent loads the Construction Law, the Bidding and Tendering Law, and related administrative regulations.

[0193] Step 3: The system initiates adversarial simulation. In the first round of simulation, the review simulation agent points out that the initial draft of the construction organization design plan is too general and is expected to lose points; the competitive adversarial agent predicts that "ambiguity in the description of concrete strength parameters" may become a point of contention; the compliance guardian agent indicates that "the safety production license has expired" is a major risk. The system encodes the current document state and these feedbacks into a state vector, inputs it into the reinforcement learning decision model, and the model outputs a sequence of modification operations: for the expired safety production license issue, immediately upload a new license; for the concrete strength parameters, modify the vague description to a clear range description; for the construction organization design plan, invoke the clause rewriting action to generate a more detailed description.

[0194] After multiple rounds of simulation, the score improved from the initial 72 points to 91 points, the list of points of contention decreased from the initial 23 to zero, and all major risk items in the risk list were eliminated. The system detected that the score change was less than the preset threshold for multiple consecutive rounds, the list of points of contention remained empty, and the risk list had no major risks, thus determining that the convergence condition was met and stopping the simulation.

[0195] Step 4: The system backtracks and deduces the process, identifying key rounds related to score improvement, dispute resolution, and risk elimination. The modification operations within these rounds are extracted, forming a sequence of twenty-seven key modification operations. Dependency verification is then performed on this sequence to ensure smooth execution during refactoring.

[0196] Step 5: The system reconstructs the initial draft based on the key modification operation sequence, generating the final version of the tender document. Simultaneously, it extracts the scoring attributions from the review simulation agent for the final document, such as "the construction organization design received a high score because it included a Gantt chart and risk response plan," records of resolved objections (e.g., "the ambiguity in concrete strength parameters was resolved through clear expression"), and records of eliminated risk items (e.g., "the expired safety production license issue was resolved by re-uploading"). The system then organizes and generates a compilation basis explanation document by chapter.

[0197] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this application. Various changes and modifications can be made to this application without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of this application as claimed. The scope of protection of this application is defined by the appended claims and their equivalents.

Claims

1. A method for generating and validating structured documents for bidding and tendering scenarios, characterized in that, The method includes: The bidding documents are parsed to construct a bidding knowledge graph, and a draft of the bid document is generated based on the bidding knowledge graph. Within the rule space defined by the bidding knowledge graph, multiple agents with different objective functions are initialized. These agents include at least a review simulation agent, a competitive adversarial agent, and a compliance guardian agent. An instantiated review simulation agent is loaded with historical bid evaluation behavior data and the terms, rules and scoring details in the bidding knowledge graph, which are used to generate a scoring report including advantages and potential deductions during the simulation. Instantiate a competitive adversarial agent, load a game strategy library including historical questioning data and differentiated strategies, as well as the terms, rules and scoring details in the bidding knowledge graph, to generate a list of potential questioning points including the location and reasons of the questioning terms; Instantiate the compliance guardian intelligent agent, load regulatory case data and the terms, rules and scoring details in the bidding knowledge graph, and use them to generate a compliance risk list including violation clauses and risk levels; The completed loading of the review simulation agent, the competitive adversarial agent, and the compliance guardian agent will be used as the initial set of agents to start multiple rounds of iterative adversarial simulations. These agents will be used to generate feedback information in each round of simulations. Multiple rounds of iterative adversarial simulations are performed on the initial draft of the tender document until a preset convergence condition is met, resulting in a simulation convergence result set; each round of simulation includes: The current document state and the feedback information of each agent are encoded into a state vector, which is then input into the reinforcement learning decision model to obtain a sequence of modification operations. The document is then revised according to this sequence, and the updated document state is generated as the input for the next round. The above process is repeated until the preset convergence condition is met. Based on the backtracking of the convergence result set accumulated during the deduction process, the sequence of key modification operations performed from the first draft to the final draft is obtained; Based on the aforementioned key modification operation sequence, the initial draft of the tender document is reconstructed, and the final version of the tender document and the corresponding explanatory document on the basis for its preparation are output.

2. The method according to claim 1, characterized in that, The process of parsing the bidding documents, constructing a bidding knowledge graph, and generating a draft bid document based on the bidding knowledge graph includes: Natural language processing is performed on the tender documents to extract technical specifications, commercial terms, scoring rules and mandatory constraints, resulting in structured clause data; The structured terms data are subjected to entity identification and relationship extraction to extract key technical parameters, delivery deadlines, qualification requirements entities, and to identify the constraints, mutual exclusions, and dependencies between entities, resulting in an entity set and a relationship set. The scoring criteria are broken down into quantifiable scoring points to obtain a set of scoring points; Using chapters as parent nodes and clauses as child nodes, the entity set, relation set, and scoring point set are associated with the corresponding nodes to construct a bidding knowledge graph that includes nodes, attributes, and edge relationships; An initial bid document framework is generated based on the chapter node structure of the bidding knowledge graph. The basic material library is then called to fill in the corresponding clause nodes, forming the first draft of the bid document.

3. The method according to claim 2, characterized in that, The extraction of technical specifications, business terms, scoring rules, and mandatory constraints yields structured clause data, including: The tender documents are identified by chapters, chapter titles and clause boundaries are marked, and a chapter structure sequence is obtained; Syntactic analysis is performed on the clause text in the chapter structure sequence to extract key technical parameters and delivery deadlines, resulting in parameter entities; Keyword matching is performed on the text of the terms to identify the qualification requirements and obtain the qualification entity. The text of the scoring rules is parsed to identify the scoring items and their weights and scoring conditions, and obtain the scoring points and their quantitative attributes. The parameter entities, qualification entities, scoring points and their quantitative attributes are associated with the corresponding clauses in the chapter structure sequence to obtain structured clause data including chapter levels and clause content.

4. The method according to claim 1, characterized in that, Encoding the current document state and the feedback information from each agent into a state vector includes: Obtain the scoring report of the review simulation agent output in the current round, the list of potential questionable points of the competitive agent output, and the list of compliance risks of the compliance guardian agent output; The predicted scores, the number of potential deductions, the number of potential points of contention in the potential questioning list, and the risk level data in the compliance risk list in the scoring report are concatenated with the text features of the current document to generate a multi-dimensional state vector.

5. The method according to claim 1, characterized in that, The convergence conditions include: The scoring prediction values ​​output by the review simulation agent in multiple consecutive rounds are obtained. When the change range of the scoring prediction value is less than a preset threshold and higher than a preset pass line, it is determined that the first convergence condition is met. Obtain a list of potential points of contention for the competitive adversarial agent's output over multiple consecutive rounds. When the list of points of contention is empty or the contents of the list are repeated with the historical output, it is determined that the second convergence condition is met. Obtain the compliance risk list output by the compliance guardian intelligent agent in multiple consecutive rounds. When there are no major risk items in the compliance risk list or all risk items are in an eliminated state, it is determined that the third convergence condition is met.

6. The method according to claim 1, characterized in that, The reinforcement learning decision model includes: Document modification operations are predefined as a set of actions, including clause rewriting, supplementary proof, and adjustment of commitment values; Based on the changes in document status after performing the modification operations in the set of actions, the combined changes of improved rating, reduced points of contention, and decreased risk level are quantified as reward values. Based on the accumulated state data, the modified operations performed, and the reward values ​​obtained during the historical simulation process, the policy network is trained to establish a mapping relationship between states and action sequences, thus obtaining a reinforcement learning decision model.

7. The method according to claim 1, characterized in that, The backtracking based on the accumulated convergence result set during the deduction process yields the sequence of key modification operations performed from the initial draft to the final draft, including: Analyze the convergence result set of the deduction, extract the document versions before and after each round of deduction and the modification operations performed in each round, and obtain the document evolution sequence and the corresponding modification operation log; The feedback data output by each agent after each round of deduction is extracted from the set of convergence results of the deduction. The changes of the feedback data relative to the previous round are analyzed to identify the rounds in which the score is improved, the rounds in which the doubts are resolved, and the rounds in which the risks are eliminated. Extract the modification operations performed in the rounds where the rating was improved, the rounds where the doubts were resolved, and the rounds where the risks were eliminated from the modification operation log, and mark them as key modification operations; The key modification operations are arranged according to the round order in which they occur, forming a sequence of key modification operations from the first draft to the final draft.

8. The method according to claim 1, characterized in that, The output of the final tender document and the corresponding explanatory document explaining the basis for its preparation includes: Based on the key modification operation sequence, perform modification operations on the initial draft of the tender document to generate the final version of the tender document; Extract the scoring attribution of the final bid document by the review simulation agent from the convergence result set to obtain the review preference response description; extract the resolved objections and corresponding modification operations from the simulation record to obtain the objection avoidance description; extract the eliminated risks and corresponding solutions from the compliance risk list to obtain the compliance compliance description. The review preference response instructions, question avoidance instructions, and compliance compliance instructions are organized according to the chapter structure of the final bid document, generating a compilation basis document, which is then linked and output with the final bid document.

9. The method according to claim 1, characterized in that, The game strategy library is constructed through the following steps: Collect historical tender documents and their corresponding letters of inquiry and complaint records to obtain historical dispute text pairs; The historical disputed texts were aligned at the clause level, and the texts of the winning bid, the unsuccessful bid, and the challenged bids under the same bidding clause were extracted to obtain a clause alignment sample including the three types of texts. Extract the text of the winning bid and the text of the unwinning bid from the alignment sample of the terms, calculate the text differences between the two, extract the differentiated expression features, and obtain the differentiated winning strategy; The challenge text is extracted from the aligned sample of the terms, the attack logic of the challenge text is analyzed, the attack points are associated with the bidding terms, and the challenge pattern template is obtained. The differentiated winning strategies and the questioning pattern templates are associated according to the clause type to construct a game strategy library that includes the correspondence between clause type, differentiated strategies and questioning patterns.

Images