Method and system for intelligent tender generation and countermeasures optimization against risks of tender rejection

Abstract

This invention relates to a method and system for intelligent bid generation and bid rejection risk mitigation, within the field of natural language processing technology for the power industry. To address the problems of low success rates, high bid rejection risks, and text style mismatches in existing intelligent bid documents for the power industry, this invention transforms unstructured power bidding data into a computer-reasonable knowledge graph. Based on the knowledge graph, it constructs a bidder preference feature vector model and calculates preference weight distribution using a holographic image of the bidding party and latent preference mining. Adaptive generation of the initial bid draft is performed based on style features derived from sentiment analysis and semantic features of latent preferences. Text similarity is calculated to align the bid tone with project attributes. Bid rejection risk mitigation detection and correction are performed based on counterfactual reasoning. Mitigation detection is conducted based on bid rejection risks, and automatic corrections are implemented to output a compliant final draft. This invention is primarily used for generating intelligent bid documents for the power industry and conducting bid rejection risk mitigation tests.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology in the power industry, and in particular to a method and system for intelligent bid generation and optimization of bid rejection risk mitigation. Background Technology

[0002] In recent years, generative artificial intelligence technology has made continuous progress. General-purpose pre-trained language models have attracted widespread attention due to their powerful semantic understanding, logical reasoning, and text generation capabilities, and have been practically applied in many general fields such as intelligent customer service, assisted writing, and code generation, demonstrating their immense value. As a key technology for improving business efficiency, large language models also have broad application prospects in the vertical field of the power industry.

[0003] While general-purpose models are trained using extensive general knowledge, they lack in-depth training in the power industry, making them difficult to directly apply to complex power bidding scenarios. Some existing research attempts to apply general-purpose models to vertical industries through Retrieval-Enhanced Generation (RAG), but these mainly focus on fields like healthcare and law; research on dedicated models for the complex bidding processes in the power industry is still limited. Currently, power grid companies and related supply chain units face enormous demands for engineering construction and material procurement. High-quality tender documents are not only crucial for companies to win bids but also the first line of defense for ensuring the safe and stable operation of the power system.

[0004] Therefore, exploring the in-depth application of large-scale model technology in the power bidding field is of significant value. By intelligently injecting historical bidding data, safety regulations, and technical specifications from the power industry into the generation process, accurate responses can be achieved to complex projects such as substation construction and transmission and distribution equipment procurement, contributing to the improvement of efficiency and quality in the digital transformation of power companies. The power industry's bidding process is characterized by strong exclusivity and a strict "one-vote veto" mechanism. Bidding parties often have implicit, unspoken preferences, and the requirements for qualifications and safety regulations allow for no errors, posing a significant challenge to the direct application of general-purpose large-scale models.

[0005] Specifically, the power industry's bidding process is highly policy-driven, technically rigorous, and exclusive (e.g., implicit preferences for specific technical routes or equipment stability), and it has a strict "one-vote veto" mechanism for disqualifying bids (e.g., missing key qualifications, safety accident records, etc.). These factors pose significant challenges to the application of large-scale models. There is an urgent need to research highly reliable intelligent bid generation technology for the power industry.

[0006] Furthermore, from a technical perspective, most existing tender document generation technologies are based on a shallow model of "current tender document (RFP) + general model," which has significant limitations. First, the general model lacks insight into the "implicit preferences" of power bidding parties (such as local power grid companies). Power bidding often implicitly includes preferences for specific technical architectures (such as the use of domestically produced chips or specific communication protocols) or past cooperation history. This information is not explicitly stated in the tender document, and the general model's literal response leads to a low success rate. Second, it lacks the ability to mitigate the "risk of bid rejection" unique to the power industry. The power industry has strict compliance requirements regarding safety qualifications, project schedule commitments, and past performance. Existing models are prone to creating "illusions," generating seemingly reasonable commitments that violate power safety regulations or qualification requirements (e.g., ignoring a mandatory type test report), leading to direct bid rejection. Finally, the tone and style are mismatched. The power industry emphasizes "safety, stability, and reliability," while text generated by general models often has a marketing flavor or is overly exaggerated, making it difficult to establish trust with review experts.

[0007] Therefore, there is a need for a method and system for optimizing intelligent bid generation and mitigating bid rejection risks in the power industry, which can enhance the competitiveness of intelligent bids in complex bidding environments. Summary of the Invention

[0008] The purpose of this invention is to address the problems of low success rate, high risk of rejection, and mismatched text style in existing intelligent bidding documents for the power industry. It provides a method and system for optimizing intelligent bidding document generation and rejection risk mitigation, which can improve the success rate of intelligent bidding documents in the power industry, accurately capture the bidding characteristics of the power industry, effectively mine the implicit preferences of the bidding party, and have a strong ability to resist the risk of rejection.

[0009] The method for intelligent bid generation and bid rejection risk mitigation optimization described in this invention includes:

[0010] S1. Transform unstructured power bidding data into a computer-reasonable knowledge graph;

[0011] Obtain multi-source heterogeneous data released by the bidding party, construct a bidding information database, preprocess the text data in the bidding information database and extract key information to obtain a knowledge graph;

[0012] S2. Knowledge Graph-Based Holographic Image and Latent Preference Mining of Bidding Parties;

[0013] Based on the knowledge graph, a preference feature vector model of the bidding party is constructed to quantify the implicit preferences of the bidding party in historical projects and obtain the preference weight distribution.

[0014] S3. Adaptive generation of the initial draft of the tender document based on style features derived from sentiment computing and semantic features of implicit preferences. By calculating text similarity, the tone of the tender document is aligned with the attributes of the project.

[0015] S4. Counterfactual reasoning-based risk assessment and correction for bid rejection;

[0016] Through a causal inference mechanism, a causal graph is constructed that points from the characteristics of the tender document to the review results. A virtual reviewer model is configured to perform "counterfactual interference" inference on the key features in the initial draft of the tender document, calculate its causal effect on the rejection result, identify the "fatal constraints" that lead to rejection, and call the real database to automatically correct omissions or illusion risks, and output a compliant final draft.

[0017] Furthermore: Following S4, it also includes:

[0018] S5, Multidimensional Quality Assessment and Closed-Loop Iterative Optimization;

[0019] By using human feedback reinforcement learning technology, we collect data on modifications made to tender documents by power experts, construct a preference dataset, fine-tune a large language model, update model parameters by optimizing the objective function, and output the final tender document text after risk optimization.

[0020] Furthermore: In S1, the construction process of the knowledge graph includes:

[0021] S11. Acquisition and preprocessing of multi-source heterogeneous data;

[0022] The bidding data of the target bidding party within a preset historical period T is collected in full. The preprocessing includes text extraction, regularization denoising and / or data cleaning.

[0023] S12. Fine-grained entity extraction based on BERT-BiLSTM-CRF;

[0024] Word embeddings are performed using a pre-trained BERT model, and then input into a BiLSTM-CRF network for named entity recognition.

[0025] S13. Formal definition of power business knowledge graph;

[0026] Construct a knowledge graph G = (V, E, R), where V is the set of entity nodes, E is the set of entities, and R is the set of relations; store the knowledge graph in the Neo4j graph database to form a knowledge graph capable of reasoning.

[0027] Further: In S2, the process of calculating the weights of the implicit preferences includes:

[0028] S21. Calculate the price sensitivity index;

[0029] Traverse all historical bidding records of the bidding party in the knowledge graph, construct a price sensitivity model, and calculate the price sensitivity index;

[0030] S22, Calculate the viscosity index of the technical route;

[0031] Analyze the reuse of specific technologies by the bidding party on specific equipment, and calculate the technology stickiness index;

[0032] S23. Calculate the compliance risk aversion coefficient;

[0033] Calculate the compliance risk aversion coefficient based on the historical rejection rate;

[0034] S24. Calculate the vector coefficients of the holographic image;

[0035] The bidding holographic profile vector coefficient is calculated based on the price sensitivity index, technology stickiness index, and compliance risk aversion coefficient.

[0036] Furthermore: In S3, the process of constructing the initial draft of the tender document includes:

[0037] S31. Calculation of style feature vector of bidding documents;

[0038] Construct prototype vectors for a "document-based robust" style and a "digital innovation" style, and then use BERT encoding on the current tender document to obtain vectors. Calculate its cosine similarity Sim with the style prototype;

[0039] S32, Intonation control parameter mapping;

[0040] Define a set of intonation control parameters for a large language model, establish a mapping function f based on the cosine similarity Sim, and calculate the set of intonation control parameters.

[0041] S33, First draft generated;

[0042] The holographic profile vector coefficients and tone control parameter set of the tender are converted into system prompt words. Based on the tender text of the project to be tendered and combined with the aforementioned preference weight distribution, a draft tender document is generated.

[0043] Furthermore: In S4, the specific steps for detecting and correcting the risk of bid rejection include:

[0044] S41. Virtual reviewer model configuration;

[0045] A counterfactual risk monitoring model (CRM) is constructed to act as a virtual reviewer; the Retrieval Enhanced Generation (RAG) technology is used to obtain real supporting information related to the current tender from the knowledge graph, and historical rejected tender cases are introduced as contextual references to achieve a human-like review of the tender content;

[0046] S42. Explicit logical reasoning mechanism based on thought chain;

[0047] The logical steps of the reasoning process are defined as follows: First, extract the entity values ​​corresponding to the constraint items in the constraint set from the target domain to be reviewed. Then, perform mathematical comparison or semantic analysis between the extracted entity values ​​and the standard values ​​to determine whether there is logical mutual exclusion. Finally, determine whether the deviation constitutes a risk of rejection based on the analysis results.

[0048] S43, Fatal Risk Assessment and Countermeasure Correction;

[0049] A structured query statement is constructed, and the RAG module is called to retrieve the corresponding real credential information from the enterprise's trusted asset base or knowledge graph. This information is defined as unmodifiable truth data. When inconsistencies are detected between the text content and the truth data, or when there are statements that could lead to bid rejection, the relevant content is subjected to adversarial detection and automatic correction. Subsequently, the truth data is injected into the original text paragraph, and a corrected compliant final draft is generated while maintaining the semantic coherence and completeness of the expression. This achieves the detection and correction of bid rejection risks.

[0050] The system disclosed in this invention for implementing the intelligent bid generation and bid rejection risk mitigation optimization method includes a knowledge graph generation module, an implicit preference weight calculation module, a draft generation module, and a detection and correction module.

[0051] The knowledge graph generation module is used to generate a knowledge graph based on the bidding information database.

[0052] The implicit preference weight calculation module is used to construct a preference feature vector model of the bidding party and calculate the preference weight distribution;

[0053] The initial draft generation module is used to generate an initial draft of the tender document based on the preference feature vector and document style.

[0054] The detection and correction module is used to perform countermeasure detection based on the risk of rejection and to automatically correct it, thereby outputting a compliant final version.

[0055] The beneficial effects of this invention are:

[0056] Based on the actual needs of bidding and tendering in the power industry, and addressing the limitations of existing general models, this invention proposes a multi-stage generation and optimization method that integrates game theory and psychology.

[0057] The first stage involves constructing a holistic profile of the power bidding party: using knowledge graph analysis to examine the bidding party's winning / rejected bid data over the past 5 years, and uncovering its implicit preferences for "price sensitivity" and "technical routes (such as primary and secondary equipment selection)".

[0058] The second stage is the risk of bid rejection detection: a counterfactual reasoning mechanism is introduced to simulate the consequences of "if a certain power qualification certificate is missing", identify fatal constraints, and set up a virtual reviewer model to conduct adversarial scanning to ensure zero risk of bid rejection;

[0059] The third stage is emotion-adaptive generation: analyzing the semantic features of the tender documents, distinguishing between "infrastructure-oriented" and "digital innovation-oriented" styles, and aligning the tone of the tender documents with the power business scenario. This method utilizes tacit knowledge from historical data and avoids compliance risks through adversarial reasoning, ensuring the generation of high-quality tender documents that are both professional and compliant with power industry standards.

[0060] The significant effects that this invention can bring include the following aspects:

[0061] 1. It enables the quantitative mining of implicit preferences of bidding parties, thereby improving the accuracy of bid response;

[0062] Existing technologies typically only respond to the explicit textual requirements of tender documents, failing to address business practices. This invention, by constructing a knowledge graph, quantifies the tenderer's "price sensitivity" and "technology roadmap stickiness," transforming implicit tendencies into constraints within a large language model. This allows tender documents to better align with the tenderer's historical business logic in terms of technology selection and business strategy.

[0063] 2. A compliance monitoring mechanism based on causal inference has been established, effectively preventing the risk of "one-vote veto";

[0064] In response to the stringent compliance requirements of power bidding, this invention changes the traditional keyword filtering model by introducing counterfactual reasoning technology. By simulating the causal path of "missing key elements leading to bid rejection," the system can accurately identify illusions generated by the large model (such as fictitious qualifications or omissions) at the logical level and correct them for authenticity, ensuring that the bids meet the compliance standard of "zero bid rejection."

[0065] 3. Adaptive intonation based on scene features ensures the professional suitability of the text;

[0066] This invention addresses the issue of style mismatch in specific industry contexts using general-purpose models. By calculating the similarity between project semantic features and standard style prototypes, the lexical weights and sentence structure parameters of the large language model are dynamically adjusted. This method ensures that infrastructure projects emphasize safety regulations and responsibilities, while digital projects highlight technological foresight, resulting in generated text styles that strictly conform to the professional standards of the power industry. Attached Figure Description

[0067] Fig. 1 This is a flowchart illustrating the method of the present invention;

[0068] Fig. 2This is a schematic diagram of the knowledge graph structure;

[0069] Fig. 3 This is a schematic diagram of the generated tender document. Detailed Implementation

[0070] The following are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The embodiments described below are only for explaining the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention should be determined by the scope of the claims. The embodiments of the present invention are described in detail below. In order to facilitate the description of the present invention and simplify the description, the technical terms used in the specification of the present invention should be interpreted broadly, including but not limited to conventional alternatives not mentioned in this application, and including both direct and indirect implementation methods.

[0071] Example 1

[0072] Combination Figs. 1-3 This embodiment describes a method for intelligent bid generation and bid rejection risk mitigation optimization. Utilizing generative artificial intelligence technology, it constructs an intelligent bid generation system for the power industry capable of deeply understanding bidding needs and mitigating compliance risks. This embodiment proposes an intelligent bid generation and optimization method for the power industry based on multi-dimensional knowledge graphs and counterfactual risk deduction. Through a closed-loop process of holographic profile mining, adaptive tone generation, and bid rejection risk mitigation detection, it injects power business logic and compliance risk control capabilities into the model, improving the success rate and security of bids.

[0073] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. This embodiment adopts the following solution:

[0074] S1. Construction of knowledge graph and business knowledge graph for power bidding;

[0075] This step serves as the data foundation layer. Its aim is to transform unstructured power bidding data into a computer-reasonable knowledge graph. The unstructured power bidding data is obtained by collecting all historical bidding documents, technical specifications, and winning / rejecting bid announcements. Deep learning models (such as BERT-BiLSTM-CRF) are used to accurately extract key entities such as the bidding party, bidders, project parameters, qualification certificates, and rejection clauses from the unstructured text. By defining semantic relationships such as "publishing," "participating in bidding," and "parameter constraints," a reasonable power business knowledge graph is constructed, transforming the discrete bidding data into a structured knowledge graph.

[0076] S11. Acquisition and preprocessing of multi-source heterogeneous data;

[0077] We collect all bidding data from the target bidders within a preset historical period T (e.g., the past 5 years). Data sources include RFP tender documents, technical specifications, winning / rejection notices, and the DL / T industry standard library. We use PyMuPDF (a Python-based PDF processing tool) to extract document text and employ regular expressions to remove header and footer noise. Specific cleaning rules are defined for power industry text: meaningless short sentences shorter than 5 characters are removed, and paragraphs containing industry keywords such as "kV," "MVA," and "type test" are retained.

[0078] S12. Fine-grained entity extraction based on BERT-BiLSTM-CRF;

[0079] Word embeddings are performed using a pre-trained BERT model (Bidirectional Encoder Representations from Transformers), and then input into a BiLSTM-CRF network (a deep learning model including BiLSTM and CRF) for named entity recognition. The defined entity set E includes:

[0080] E={ };

[0081] in, Indicates the bidding party, Indicates the bidder, Indicates the project type. Indicates equipment parameters, Indicates qualification certificate, This indicates a clause indicating a rejection of the bid.

[0082] S13. Formal definition of power business knowledge graph;

[0083] Construct a knowledge graph G = (V, E, R), where V is the set of entity nodes and R is the set of relations. Define key relations. include:

[0084] ;

[0085] in, Indicates a publishing relationship, connecting the bidding party and the project, indicating < , , >, meaning "the bidding party has released a project notice"; Indicates the bidding relationship, connecting the bidders and the project, indicating < , , > This means that the bidder participated in the bidding for a certain project; Indicates the bidding relationship, connecting the bidder and the project, indicating < , , > This means that the bidder wins the bid for a certain project. This indicates parameter constraint relationships, connecting project and equipment parameters, and represents < , , > This means that the project requires specific equipment parameters; This indicates the relationship between qualification requirements, connecting the project and the qualification certificate, and represents < , , This means that the project requires bidders to possess specific qualification certificates. This indicates the triggering relationship for bid rejection, connecting the bidder / project with the rejection clause, indicating < , , > This means that the bidder violated a certain clause that would disqualify the bid; This indicates the standard definition relationship, connecting equipment parameters with rejected clauses / industry standards, and represents < , , The deviation range of a certain technical parameter defines the boundary of the rejected standard. The knowledge graph is stored in the Neo4j graph database to form a reasonable knowledge graph.

[0086] S2. Knowledge Graph-Based Holographic Image and Latent Preference Mining of Bidding Parties;

[0087] This step uses mathematical statistics to uncover implicit rules not explicitly stated by the bidding party. As a decision support layer, this step aims to "understand" the bidding party. Based on the knowledge graph built using S1, statistical modeling is used to quantify the three core characteristics of the bidding party:

[0088] (1) Price sensitivity: Determine whether the bidding party prefers "lowest price priority" or "technology priority";

[0089] (2) Technical route stickiness: Explore the implicit preferences of the bidding party for specific technical routes beyond the standard documents;

[0090] (3) Compliance risk aversion coefficient: Based on the historical rejection rate, assess the tenderer's tolerance for minor deviations.

[0091] Finally, the above indicators are integrated to generate a holographic profile vector of the bidding party, providing personalized guidance for the subsequent generation process.

[0092] S21, Price Sensitivity Index Quantification;

[0093] Traverse all historical winning bid records of this bidding party in the knowledge graph to construct a price sensitivity model. Let... Let be the winning bid price for the i-th bidding project. As the budget ceiling, The time decay factor (with higher weight for more recent projects) determines the price sensitivity of the bidding party. Calculate according to formula (1):

[0094] ;

[0095] Where N is the total number of tenders. The closer the value is to 1, the more likely the bidding party is to significantly lower the price, belonging to the "price-sensitive" type; the smaller the value, the more tolerant they are of high prices, belonging to the "technology-oriented" type.

[0096] S22, Probabilistic modeling of the stickiness index of technical routes;

[0097] This analysis examines the reuse of specific technologies by bidding parties in specific equipment (such as smart meters). Let the number of times a certain technology, k, appears in historical winning bids be denoted as k. If the total number of tenders is N, then the technology stickiness index is... Calculate according to formula (2):

[0098] ;

[0099] in, This represents the frequency of the technology's adoption in industry-standard practices (inverse document frequency concept). If... If a preset threshold is reached, the technology is determined to be a “strong implicit preference” of the bidding party and must be forcibly included in the tender documents.

[0100] S23, Compliance Risk Aversion Coefficient;

[0101] This index is calculated based on historical bid rejection rates. Let the number of bidders rejected due to non-compliance in the i-th project be . The total number of valid bidders is Similarly, a time decay factor is introduced. The compliance risk aversion coefficient The formula is as follows:

[0102] ;

[0103] in, The larger the value, the higher the bid rejection rate in the past projects, indicating that the bidding party has a high rate of bid rejection, extremely strict review, and low tolerance for minor deviations (i.e., extreme aversion to compliance risks), suggesting that the highest level of compliance check should be carried out when generating the bid documents; The smaller the value, the more lenient the bidding party is.

[0104] S24. Holographic image vector construction;

[0105] By integrating the above indicators, a holographic profile vector coefficient for the bidding process is generated:

[0106] .

[0107] S3. Adaptive generation of initial bid drafts based on sentiment computing and semantic features;

[0108] This step serves as the content generation layer. It calculates the stylistic features of the current tender document (e.g., "Document-based" or "Digitally Innovative"), and combines this with the holographic profile generated by S2 to construct intonation control parameters that include weighted security vocabulary and strict sentence structure constraints. The profile features and intonation parameters are then converted into system prompts, driving the large language model to generate a draft tender document aligned with the tenderer's business attributes. This step achieves alignment between the tender document's tone and project attributes by calculating text similarity.

[0109] S31. Calculation of style feature vector of bidding documents;

[0110] Constructing a prototype vector for a "documentary robust" style (High-weighted keywords: safety, compaction, stability) and "digital innovation" style prototype vectors (High-weight keywords: empowerment, perception, iteration). Regarding the current tender document... BERT encoding is used to obtain the document semantic vector. Calculate its cosine similarity Sim with the style prototype, as shown in the formula:

[0111] ;

[0112] in, express or The item attributes are determined based on cosine similarity (Sim), that is: if > Then the tone is set as a conservative style of writing, if < Therefore, it is defined as digital innovation-oriented;

[0113] S32, Intonation control parameter mapping;

[0114] Define the set of intonation control parameters for a large language model A mapping function f is established based on the cosine similarity Sim obtained from S31:

[0115] ;

[0116] For example, when identified as document robust, the function outputs a weighted list of highly secure terms. and highly rigorous sentence structure constraints When identified as digital innovation, the function outputs a low-security vocabulary weighted average. and low strict sentence structure constraints ;

[0117] For example, when identified as document robust At that time, the function outputs a weighted list of highly secure terms. (e.g., 0.6-1.0) and highly rigorous sentence structure constraints (e.g., 0.6-1.0); when identified as digital innovation type At that time, the function outputs a low-security vocabulary weighted average. (e.g., 0.1-0.5) and low-strictness sentence structure constraints (e.g., 0.1-0.5). Among them... The density of safety-related technical terms in the tender document is controlled, with a value range of [0,1]. The higher the coefficient, the more required the model to incorporate high-weight terms in the field of power safety, such as "dual protection," "anti-misoperation," "redundant configuration," and "fault blocking," into key paragraphs such as technical descriptions and solution commitments. The coefficient controls the logical rigor of the generated sentences, ranging from 0 to 1. A higher coefficient requires the model to employ more complex sentence structures with explicit conditional clauses, preconditions, and result guarantees, such as "Under the operating conditions of substations at voltage levels of 110kV and above, through a dual-redundancy protection mechanism, ensure that the accuracy of relay protection actions continuously meets the requirements of DL / T 559-2018 'Technical Specification for Automatic Safety Devices for Relay Protection'", to reflect the rigor and traceability of the technical solution.

[0118] S33, First draft generated;

[0119] The vector coefficient of the holographic image in the tender and intonation control parameter set Converted into system prompts, combined with the requirements of the tender documents provided by the tendering party, and input into a large language model, a draft tender document is generated. .

[0120] ;

[0121] Here, LLM stands for Large Language Model. Indicates a prompt word, This indicates the context required by the tender documents.

[0122] S4. Counterfactual reasoning-based risk assessment and correction for bid rejection.

[0123] This step serves as a quality risk control layer, addressing the risk of "one-vote veto." A causal inference mechanism is introduced to construct a causal graph from tender features to review results. A virtual reviewer model is configured to perform "counterfactual interference" exercises (simulating scenarios where the feature is missing) on ​​key features in the initial draft, calculating their causal effect on the rejection result. "Fatal constraints" leading to rejection are identified, and for omissions or illusory risks, a real database is invoked for automatic correction, outputting a compliant final draft.

[0124] S41, Virtual Reviewer Model Configuration.

[0125] A counterfactual risk monitoring model (CRM) is constructed as a virtual reviewer. This model uses a large language model as its core for logical reasoning and simulates the cognitive process of a human reviewer through cue word engineering. During reasoning, the model utilizes Retrieval Enhanced Generation (RAG) technology to retrieve relevant authentic evidence from a knowledge graph and incorporates historical rejected bid cases as contextual references, achieving a human-like review of the bid content. To transform unstructured industry rules into instructions understandable to the model, a structured three-dimensional cue input sequence X is designed, which is composed of the following three mapping domains:

[0126] ;

[0127] in, The instruction for the model is as follows: "You are an extremely rigorous power bidding review expert with the authority to screen bids using a 'one-vote veto' system. Your task is to identify any logical flaws, vague promises, or parameter deviations in the bid documents based on the context of discarded cases." This represents a dynamic constraint injection domain. Utilizing RAG and contextualization techniques, based on the currently generated tender document fragment content, it retrieves associated hard rejection clause nodes from the knowledge graph constructed in S13. Technical parameter standard node This structured data is transformed into a set of constraints described in natural language. (For example: {“Clause 3.1: Registered capital > 50 million yuan”, “Standard GB / T: Transformer loss must be 120kW”}, serving as an insurmountable “review red line” for the model; among which This indicates the target domain to be reviewed.

[0128] S42. Explicit logical reasoning mechanism based on thought chain.

[0129] To prevent the Counterfactual Risk Monitoring (CRM) model from generating "illusions" or directly drawing unfounded conclusions, the model is required to perform explicit chain-of-thought reasoning before outputting its final judgment. The reasoning process is defined as follows: It includes the following three strict logical steps:

[0130] S421. The model first starts from the target domain to be reviewed. Extracting and constraining sets The entity value corresponding to the constraint. For example, extract "Bidder's registered capital: 48 million yuan" from the text and anchor it to "50 million yuan" in the rule.

[0131] S422. Logical Implication Analysis. Perform mathematical comparison or semantic analysis between the extracted values ​​and the standard values. The model needs to identify numerical relationships (e.g., 4800 < 5000) or semantic conflicts (e.g., a rule requiring "must" responds to "satisfied in principle"), and determine whether logical mutual exclusion exists.

[0132] S423. Risk Classification and Assessment. Based on the analysis results of S2, determine whether the deviation constitutes a risk of bid rejection.

[0133] The model calculation process is represented as follows:

[0134] Output = CRM(R_reasoning, X);

[0135] Output is the final compliance conclusion, which is the result of counterfactual reasoning-based risk assessment of bid rejection.

[0136] S43, Fatal Risk Assessment and Countermeasure Correction;

[0137] This step aims to achieve a closed loop from risk diagnosis to content self-healing based on the output of counterfactual analysis of bid cancellation risk detection results. For diagnosed bid cancellation risks, the system triggers a correction mechanism:

[0138] A structured query statement is constructed, and the Retrieval Enhancement Generation (RAG) module is invoked to retrieve authentic credential information corresponding to the current tender content from the enterprise's trusted asset base or knowledge graph. This information includes certificate numbers, key data from audit reports, and relevant qualification information. The retrieval results are then validated for consistency, and the validated data is designated as unmodifiable truth values. Subsequently, semantic similarity calculations are used to determine the text fragments corresponding to these truth values. Key entities in the truth values ​​(such as "Level 1 Installation Qualification" and "Transformer No-Load Loss 8.5kW") are converted into query vectors. Simultaneously, the tender text is segmented by sentences or paragraphs, and each text fragment is also converted into a vector. The cosine similarity between the query vector and each text fragment vector is calculated, and the text fragment with the highest similarity is selected as the position to be corrected. For example, when the truth value is "Level 1 Installation Qualification Certificate Number DL-2023-0886," the system will find the sentence in the tender containing words such as "installation," "qualification," and "Level 1" that is semantically closest, such as "Our company holds a Level 1 license for the installation (repair, testing) of power facilities."

[0139] Based on the aforementioned location results, the system performs controlled local rewriting, replacing or supplementing the truth information to the corresponding text positions. By using the prompt "ensure the modified text flows naturally with the context and is semantically coherent," the system delegates the task of maintaining semantic coherence to the large language model itself. This allows the system to generate the corrected final tender text while preserving contextual semantic coherence. .

[0140] ;

[0141] in, For the original text paragraph where risks exist, Ouput indicates the final compliance conclusion of the risk assessment.

[0142] When inconsistencies are detected between the text content and the truth data, or expressions that could lead to bid rejection, the relevant content is subjected to adversarial detection and automatic correction. Subsequently, the truth data is injected into the original text paragraph, and a corrected and compliant final draft is generated while maintaining the semantic coherence and completeness of the expression, thereby realizing the detection and correction of bid rejection risks.

[0143] The description of the risk of bid rejection was retrieved from the Enterprise Trusted Asset Database or Knowledge Graph Platform, and its content format is roughly as follows:

[0144] (1) Qualification-related risks: fabrication, exaggeration, or omission of mandatory qualification certificates (such as "License for Installation (Repair, Test) of Power Facilities" of incorrect level, "Safety Production License" expired, "Type Test Report" missing, etc.);

[0145] (2) Risks related to indicators: The promised technical parameters deviate from the bottom line requirements of the tender documents (such as the promised value of "transformer no-load loss" being higher than the upper limit specified in the tender documents, or "short-circuit impedance" exceeding the allowable deviation range, etc.).

[0146] (3) Commitment-related risks: making commitments that exceed the company's actual capabilities or asset scope (such as promising to "complete emergency repairs within 48 hours" but the company's actual on-site personnel are insufficient, or promising to "provide a local spare parts warehouse" but the company has no warehousing facilities in the region, etc.).

[0147] (4) Consistency risk: Inconsistent statements in the tender document (e.g., promising "A brand circuit breakers" in the technical solution, but listing B brand in the equipment list; or inconsistencies between the commercial quotation and the summary of itemized quotations).

[0148] (5) Compliance risks: statements that violate mandatory regulations in the power industry (such as promising "live work is allowed" but not mentioning safety distance requirements, or selecting equipment in explosion-proof areas that does not meet the GB3836 standard).

[0149] The adversarial detection uses RAG retrieval, and the correction includes comparing the RAG retrieval with the original text content, which is then corrected by CRM.

[0150] This step initiates an automatic correction process based on the rejection risks identified by S42. First, a query is constructed targeting the rejection risks, and the RAG module is invoked to retrieve corresponding real information (such as certificate numbers, product parameters, performance contracts, etc.) from the enterprise's trusted asset base or knowledge graph. Then, semantic similarity calculations are used to locate the corresponding positions requiring modification in the initial draft of the tender document. The retrieved real information is compared with the original text to determine if there are any issues that could lead to rejection. For content confirmed to require modification, a large language model is invoked, inputting both the original text and the real information. The CRM automatically rewrites the paragraph, seamlessly integrating the real information into the text while maintaining contextual coherence. Finally, the modified content replaces the original text, and the corrected, compliant final draft is output.

[0151] S5, Multidimensional Quality Assessment and Closed-Loop Iterative Optimization;

[0152] This step serves as the model training layer. Data encompassing compliance, preference hit rate, and intonation style is designed. A human feedback reinforcement learning (RLHF) mechanism is introduced, utilizing data from power industry experts' revisions of tender documents to construct a preference dataset. The proximal policy optimization (PPO) algorithm is then used to fine-tune the large language model. By maximizing the reward function, the model is encouraged to continuously learn from the implicit experience of experts, achieving continuous evolution in the quality of generated content.

[0153] Human feedback reinforcement learning (RLHF) technology was used to collect data on modifications made to tender documents by power industry experts, constructing a preference dataset. The proximal policy optimization (PPO) algorithm was then employed to fine-tune the large language model by maximizing the objective function. To update the LLM parameters of the large language model The objective function formula is as follows:

[0154] ;

[0155] Here, (s, a) represents a state-action pair, where state s refers to the input "holographic image vector + tender document context Prompt", and action a refers to the "tender document text content" generated by the model. This indicates the current strategy, i.e., in the parameters. Under control, given an input context s, the model generates the concept distribution of the tender text a. The reference policy typically refers to a supervised fine-tuning (SFT) model before it is fine-tuned by reinforcement learning. This represents the KL divergence penalty coefficient. It is an adaptive hyperparameter used to control the balance between "reward maximization" and "language stability". This represents the mathematical expectation, which is the average calculation of a batch of tender documents generated by the current model.

[0156] The independent variable in the formula is the parameter φ (i.e., the weight value in the neural network) of the large language model. (Reward model) and reference model The parameters remain constant throughout the fine-tuning process; only the parameters φ of the large language model are continuously updated and optimized.

[0157] Fine-tuning employs the gradient ascent method: each time, the gradient of the objective function J(φ) with respect to the parameter φ is calculated (i.e., the adjustment direction), and then the parameter φ is moved a small step along the gradient direction, increasing the value of J(φ). This process is repeated thousands of times, gradually optimizing the parameter φ from its initial value to its optimal value, ultimately resulting in increasingly higher quality text generated by the model.

[0158] The state s (i.e., "holographic image vector + tender document context Prompt") and the text a generated by the model are concatenated together and sent as a complete input to the reward model. The reward model processes this input internally and outputs a score.

[0159] The reward model is a pre-trained model that has learned the criteria for judging "what is good text and what is bad text". When given a set of (s, a), the reward model will automatically give a score based on the learned criteria - if the text conforms to expert preferences (accurate terminology, reasonable promises, professional style), the score is high; if the text has risks (exaggerated qualifications, falsely labeled parameters, vague expression), the score is low.

[0160] This step uses the Proximal Policy Optimization (PPO) algorithm to fine-tune the Large Language Model (LLM). The fine-tuning targets the parameters φ of the Large Language Model itself (i.e., the weights of each layer in the neural network), with the goal of enabling the text generated by the model to obtain a higher reward model score without deviating from the original distribution of the reference model.

[0161] By learning from expert revision behavior, the large language model can automatically avoid expressions that do not conform to bidding specifications, improve the hit rate of bidding scoring points, and enhance the consistency of professional expression in the bid document text.

[0162] By fine-tuning the existing RLHF technology, the large language model can automatically generate tender documents that conform to industry standards, have a stable style, and cover more scoring points during the inference stage, based on the input tender document requirements and the company's holographic profile information, thereby improving the quality of the final tender and its competitiveness in winning bids.

Claims

1. A method for intelligent bid generation and optimization to mitigate the risk of bid rejection, characterized in that: include: S1. Transform unstructured power bidding data into a computer-reasonable knowledge graph; Obtain multi-source heterogeneous data released by the bidding party, construct a bidding information database, preprocess the text data in the bidding information database and extract key information to obtain a knowledge graph; S2. Knowledge Graph-Based Holographic Image and Latent Preference Mining of Bidding Parties; Based on the knowledge graph, a preference feature vector model of the bidding party is constructed to quantify the implicit preferences of the bidding party in historical projects and obtain the preference weight distribution. In S2, the process of calculating the weights of the implicit preferences includes: S21. Calculate the price sensitivity index; Traverse all historical bidding records of the bidding party in the knowledge graph, construct a price sensitivity model, and calculate the price sensitivity index; S22, Calculate the viscosity index of the technical route; Analyze the reuse of specific technologies by the bidding party on specific equipment, and calculate the technology stickiness index; S23. Calculate the compliance risk aversion coefficient; Calculate the compliance risk aversion coefficient based on the historical rejection rate; S24. Calculate the vector coefficients of the holographic image; The bidding holographic profile vector coefficient is calculated based on the price sensitivity index, technology stickiness index, and compliance risk aversion coefficient. S3. Adaptive generation of the initial draft of the tender document based on style features derived from sentiment computing and semantic features of implicit preferences. By calculating text similarity, the tone of the tender document is aligned with the attributes of the project. In S3, the process of constructing the initial draft of the tender document includes: S31. Calculation of style feature vector of bidding documents; Construct prototype vectors for a "document-robust" style and a "digital-innovative" style, and then use BERT encoding on the current tender document to obtain a document semantic vector. Calculate its cosine similarity Sim with the style prototype vector; S32, Intonation control parameter mapping; Define a set of intonation control parameters for a large language model, establish a mapping function f based on the cosine similarity Sim, and calculate the set of intonation control parameters. S33, First draft generated; The vector coefficients of the holographic profile of the tender and the tone control parameter set are converted into system prompt words. Based on the tender text of the project to be tendered and combined with the aforementioned preference weight distribution, a draft tender document is generated. S4. Counterfactual reasoning-based risk assessment and correction for bid rejection; Through a causal inference mechanism, a causal graph is constructed that points from the characteristics of the tender document to the review results. A virtual reviewer model is configured to perform "counterfactual interference" inference on the key features in the initial draft of the tender document, calculate its causal effect on the rejection result, identify the "fatal constraints" that lead to rejection, and call the real database to automatically correct omissions or illusion risks, and output a compliant final draft. In S4, the specific steps for detecting and correcting the risk of bid rejection include: S41. Virtual reviewer model configuration; A counterfactual risk monitoring model (CRM) is constructed to act as a virtual reviewer; retrieval enhancement generation technology is used to obtain real supporting information related to the current tender from the knowledge graph, and historical rejected tender cases are introduced as contextual references to achieve a human-like review of the tender content; S42. Explicit logical reasoning mechanism based on thought chain; The logical steps of the reasoning process are defined as follows: First, extract the entity values ​​corresponding to the constraint items in the constraint set from the target domain to be reviewed. Then, perform mathematical comparison or semantic analysis between the extracted entity values ​​and the standard values ​​to determine whether there is logical mutual exclusion. Finally, determine whether the deviation constitutes a risk of rejection based on the analysis results. S43, Fatal Risk Assessment and Countermeasure Correction; A structured query statement is constructed, and the RAG module is called to retrieve the corresponding real credential information from the enterprise's trusted asset base or knowledge graph. This information is defined as unmodifiable truth data. When inconsistencies are detected between the text content and the truth data, or when there are statements that could lead to bid rejection, the relevant content is subjected to adversarial detection and automatic correction. Subsequently, the truth data is injected into the original text paragraph, and a corrected compliant final draft is generated while maintaining the semantic coherence and completeness of the expression. This achieves the detection and correction of bid rejection risks.

2. The method for intelligent bid generation and bid rejection risk mitigation optimization according to claim 1, characterized in that, Following S4, it also includes: S5, Multidimensional Quality Assessment and Closed-Loop Iterative Optimization; By using human feedback reinforcement learning technology, we collect data on modifications made to tender documents by power experts, construct a preference dataset, fine-tune a large language model, update model parameters by optimizing the objective function, and output the final tender document text after risk optimization.

3. The method for intelligent bid generation and bid rejection risk mitigation optimization according to claim 1, characterized in that, In S1, the construction process of the knowledge graph includes: S11. Acquisition and preprocessing of multi-source heterogeneous data; The bidding data of the target bidding party within a preset historical period T is collected in full. The preprocessing includes text extraction, regularization denoising and / or data cleaning. S12. Fine-grained entity extraction based on BERT-BiLSTM-CRF; Word embeddings are performed using a pre-trained BERT model, and then input into a BiLSTM-CRF network for named entity recognition. S13. Formal definition of power business knowledge graph; Construct a knowledge graph G = (V, E, R), where V is the set of entity nodes, E is the set of entities, and R is the set of relations; store the knowledge graph in the Neo4j graph database to form a knowledge graph capable of reasoning.

4. A system for implementing the intelligent bid generation and bid rejection risk mitigation optimization method as described in any one of claims 1-3, characterized in that, It includes a knowledge graph generation module, an implicit preference weight calculation module, a draft generation module, and a detection and correction module; The knowledge graph generation module is used to generate a knowledge graph based on the bidding information database. The implicit preference weight calculation module is used to construct a preference feature vector model of the bidding party and calculate the preference weight distribution; The initial draft generation module is used to generate an initial draft of the tender document based on the preference feature vector and document style. The detection and correction module is used to perform countermeasure detection based on the risk of rejection and to automatically correct it, thereby outputting a compliant final version.

Images