Service platform for completing bidding document production based on automation technology assistance

Abstract

The invention discloses a service platform for completing bidding document production based on automation technology assistance, and belongs to the technical field of bidding services, and the service platform specifically comprises the steps that a user behavior collection module collects operation behavior data of a user in the bidding document production process; the decision preference analysis module processes the operation behavior data through an analysis model to generate user decision preference parameters; the strategy generation module generates a candidate bidding strategy set based on historical cases; the scene simulation module constructs a bidding scene simulation environment and records market competition data in a strategy execution process; the strategy evaluation module calculates a strategy competitiveness index according to the market competition data; the strategy execution module screens an optimal bidding strategy and converts the optimal bidding strategy into a specific bidding file making scheme; through intelligent analysis of user behaviors and dynamic simulation of a bidding environment, the accuracy and competitiveness of bidding document production are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of bidding service technology, and more specifically to a service platform based on automation technology to assist in the preparation of bidding documents. Background Technology

[0002] With increasingly fierce market competition, bidding has become a crucial way for companies to secure commercial projects. As a key document showcasing a company's overall strength, the quality of bid documents directly impacts the bidding outcome. Currently, various organizations invest significant human and material resources in bid preparation, a process involving complex stages such as market analysis, strategy development, and solution design. In the context of digital transformation, how to improve the efficiency and quality of bid document preparation through technological means has become a widespread concern in the industry.

[0003] In the existing technology field, various bidding assistance systems have emerged. These systems are typically based on document template libraries and historical case libraries, providing users with a basic bid document framework through rule matching. Some systems can automatically fill in relevant content paragraphs based on the basic project information entered by the user; others employ keyword extraction technology to recommend similar content modules from historical successful cases. These technical solutions, to some extent, reduce the burden of paperwork for users and provide a standardized document generation path.

[0004] However, existing technical solutions generally rely on static rules and fixed templates to generate bidding strategies, lacking the ability to deeply explore users' potential needs. They fail to effectively capture users' decision-making preferences and risk tolerance during document editing, leading to discrepancies between the generated bidding strategies and the users' true intentions. Furthermore, the lack of a dynamic simulation mechanism for the market competition environment makes it difficult to assess the adaptability and competitiveness of the strategies in actual bidding environments. This disconnect between strategy generation and practical application directly impacts the quality of bid documents and the success rate of bids. Summary of the Invention

[0005] The purpose of this invention is to provide a service platform based on automation technology to assist in the preparation of tender documents, and to solve the following technical problems: Existing tender document generation systems struggle to accurately capture the potential needs revealed by users during the editing process, resulting in discrepancies between the generated bidding strategies and the users' true intentions; at the same time, they lack the ability to dynamically simulate the market competition environment.

[0006] The objective of this invention can be achieved through the following technical solutions: A service platform for assisting in the preparation of tender documents based on automation technology includes: The user behavior collection module is used to collect user operation behavior data during the bid document preparation process. The operation behavior data includes content adjustment records, format modification frequency, and chapter access duration. The decision preference analysis module is used to process operational behavior data through the decision preference analysis model and generate user decision preference parameters, which include scheme selection tendency and risk control preference. The strategy generation module is used to input user decision preference parameters into the bidding strategy generation model, which generates a set of candidate bidding strategies based on historical bidding cases. The scenario simulation module is used to build a bidding scenario simulation environment, execute various strategies in the candidate bidding strategy set in the bidding scenario simulation environment, and record market competition data during the strategy execution process; The strategy evaluation module is used to calculate the strategy competitiveness index based on market competition data during the strategy execution process. The strategy competitiveness index includes the solution advantage and resource utilization efficiency. The strategy execution module is used to select the optimal bidding strategy based on the strategy competitiveness index, transform the optimal bidding strategy into a specific bid document preparation plan, and execute it.

[0007] As a further aspect of the present invention: the specific process of processing operational behavior data through the decision preference analysis model in the decision preference analysis module is as follows: Collect data on the deletion locations of content by users during the editing process of tender documents, and statistically analyze the distribution density of deletion operations in each chapter and paragraph; combine the time series information in the text modification frequency data to establish a time distribution map of modification operations; analyze the duration characteristics in the paragraph dwell time data to identify areas where user attention is concentrated. The data on the distribution density of deletion operations, the time distribution of modification operations, and the concentrated areas of user attention are input into the bidding decision analysis network. The bidding decision analysis network includes a feature extraction layer, a time series processing layer, and a decision output layer. The feature extraction layer calculates spatial distribution features, the time series processing layer analyzes time series patterns, and the decision output layer generates user decision preference parameters, including scheme selection tendency parameters and risk control preference parameters.

[0008] As a further aspect of the present invention: In step S3, the specific process of inputting user decision preference parameters into the bidding strategy generation model is as follows: The user decision preference parameter sequence is input into the bidding strategy encoder network. The bidding strategy encoder network uses a business decision long short-term memory architecture to process the sequence data, calculates the forward business feature sequence and the backward business feature sequence respectively, and merges the forward business feature sequence and the backward business feature sequence to generate a complete business feature vector. The complete business feature vector is then passed to the bidding strategy decoder network. The bidding strategy decoder network uses a business attention mechanism to calculate business context vectors of different dimensions and integrates the business context vectors of different dimensions into a comprehensive business context vector. Based on the comprehensive business context vector, a bidding strategy description sequence is generated step by step, with each strategy description unit corresponding to a specific business strategy element. A business strategy search process is performed on the bidding strategy description sequence, multiple candidate strategy description sequences are maintained and the sequence generation probability is calculated in real time. The number of candidate strategies is dynamically adjusted according to the sequence generation probability, duplicate business strategy descriptions are removed, and an optimized set of candidate bidding strategies is output.

[0009] As a further aspect of the present invention: the specific process by which the bidding strategy encoder network processes the user decision preference parameter sequence is as follows: Initialize the forward and backward business states of the business decision long short-term memory network; sequentially input each business parameter vector in the user decision preference parameter sequence, and update the forward and backward business state sequences respectively; calculate the final state vector of the forward business state sequence, and calculate the initial state vector of the backward business state sequence. The final state vector of the forward business state sequence and the initial state vector of the backward business state sequence are merged into a complete business feature vector; attention distribution data is calculated through a business attention mechanism to generate an enhanced business context vector; the final business feature vector and business context vector are output to the bidding strategy decoder network.

[0010] As a further aspect of the present invention: the specific process of constructing the bidding scenario simulation environment in the scenario simulation module is as follows: Initialize the basic parameters of the virtual bidding environment and load complete case records from the historical bidding project database; configure multiple competitor business models, and set the bidding strategy parameter set and market reaction parameter set for each competitor business model. The bidding strategy parameter set includes initial price parameters, price adjustment parameters, and price limit parameters, while the market reaction parameter set includes reaction time parameters and competition intensity parameters; define bidding market environment variables, including demand change parameters, price sensitivity parameters, and market capacity parameters. The candidate bidding strategy set is sequentially input into the bidding simulation environment to execute multiple rounds of bidding competition simulation. Each round of simulation includes a bidding stage, a bid evaluation stage, and a result confirmation stage. During each round of simulation, the bidding market status data is updated in real time, and detailed price data, winning bid result data, and project revenue data for each bidding strategy are recorded. The complete bidding simulation process data is stored in the bidding analysis database.

[0011] As a further aspect of the present invention: the specific process for configuring the competitor's business model is as follows: Define basic competitor attribute parameters, including competitor identifier, historical bidding behavior data, and business decision preference parameters; set up a competitor decision rule system, including pricing calculation rules, market reaction rules, and risk assessment rules; initialize competitor instance objects and allocate an initial set of business state parameters; invoke the decision rule system in real time during the bidding simulation to generate pricing decision data based on the current bidding market status; update the competitor's internal state parameters and record complete competitor behavior log data, including decision time points, decision input information, and decision output results.

[0012] As a further aspect of the present invention: the specific process of calculating the strategy competitiveness index based on market competition data during strategy execution in the strategy evaluation module is as follows: Complete pricing sequence data for each bidding strategy is extracted from the bidding analysis database. The commercial deviation of the pricing sequence from the market average price is calculated, including absolute difference data and relative fluctuation data. Market share time series data for each bidding strategy is analyzed to calculate real-time growth data and long-term trend data. Combining the commercial deviation of the pricing and the market share growth characteristics, a scheme advantage index is generated through commercial weighting. Simultaneously, time-series data of resource consumption for each bidding strategy are extracted, and real-time and long-term efficiency data of resource input and output are calculated to generate resource utilization efficiency indicators. The scheme advantage index and resource utilization efficiency index are commercially standardized, and a comprehensive competitiveness score of the bidding strategy is generated through business combination calculation.

[0013] As a further aspect of the present invention: the specific process of selecting the optimal bidding strategy based on the strategy competitiveness index in the strategy execution module is as follows: The system receives comprehensive competitiveness scores for bidding strategies and selects the highest-scoring strategy as the target strategy. It then parses the complete description of the target strategy, identifying the necessary adjustments to the bid document paragraphs and content. A detailed set of bid document modification instructions is generated, including the modification location information, the modified content text, and the modification format requirements. The system executes document modification operations according to the instruction set, updating the bid document content and adjusting its format. Finally, the bidding document processing engine generates the final document, outputting a standardized bid document according to the required specifications.

[0014] The beneficial effects of this invention are: This invention effectively solves the technical challenge of capturing potential user needs in existing bidding technologies by constructing a user behavior collection and decision preference analysis model. By analyzing user actions during document editing, such as deletion locations, modification frequency, and dwell time, a precise user preference profile is established, generating bidding strategies that align with the user's true intentions. Furthermore, by establishing a virtual bidding scenario simulation environment, the system can conduct multiple rounds of market competition simulations on candidate strategies, dynamically evaluating their adaptability and competitiveness based on real-time market data. A multi-objective optimization screening mechanism based on strategy competitiveness indicators ensures that the final output bidding strategy possesses both advantages and resource utilization efficiency. The entire solution forms a complete technical closed loop from behavior collection, strategy generation, simulation evaluation to optimized output. Through a continuous learning mechanism, the strategy generation effect is constantly optimized, significantly improving the quality and market competitiveness of bid documents and effectively increasing the bid success rate. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, this invention is a service platform for assisting in the preparation of tender documents based on automation technology, comprising: The user behavior collection module is used to collect user operation behavior data during the bid document preparation process. The operation behavior data includes content adjustment records, format modification frequency, and chapter access duration. The decision preference analysis module is used to process operational behavior data through the decision preference analysis model and generate user decision preference parameters, which include scheme selection tendency and risk control preference. The strategy generation module is used to input user decision preference parameters into the bidding strategy generation model, which generates a set of candidate bidding strategies based on historical bidding cases. The scenario simulation module is used to build a bidding scenario simulation environment, execute various strategies in the candidate bidding strategy set in the bidding scenario simulation environment, and record market competition data during the strategy execution process; The strategy evaluation module is used to calculate the strategy competitiveness index based on market competition data during the strategy execution process. The strategy competitiveness index includes the solution advantage and resource utilization efficiency. The strategy execution module is used to select the optimal bidding strategy based on the strategy competitiveness index, transform the optimal bidding strategy into a specific bid document preparation plan, and execute it.

[0019] In a preferred embodiment of the present invention, the specific process of processing operational behavior data through the decision preference analysis model in the decision preference analysis module is as follows: First, an event listener is deployed in the tender document editing interface. This listener captures three types of user actions during the document editing process in real time. For content deletion, the system records the specific location of each deletion, including the paragraph number, sentence index, and character offset, as well as the type of content deleted: technical description, commercial terms, or pricing information. For example, when a user frequently deletes specific terms in the technical solution section, the system records that these deletions are concentrated in the third paragraph of the document and mainly involve descriptions related to "cloud computing architecture."

[0020] The system collects operation data in time windows, each set to 30 minutes. Within a single time window, the system calculates the deletion operation density for each paragraph by dividing the number of deletions by the paragraph length and then multiplying by 100 to obtain a standardized density value. For example, if a technical solution paragraph is 500 characters long and 5 deletion operations occur within 30 minutes, the deletion density is calculated to be 1.0.

[0021] For text modification frequency data, the system records the timestamp and modification type of each modification. Modification types are divided into three categories: content replacement, format adjustment, and structural adjustment. The system builds a time distribution map of modification operations at 15-minute intervals. This map contains 24 time points, and each time point records the number of modification operations within that time period. For example, if 8 content replacements and 3 format adjustments are detected between 9:00 AM and 9:15 AM, the modification intensity at that time point is marked as 11.

[0022] When analyzing paragraph dwell time, the system identifies user reading patterns by monitoring cursor position changes. When the cursor remains in a paragraph area for more than 10 seconds, the system begins recording the dwell time. Simultaneously, the system identifies abnormal dwell times; for example, if a budget description paragraph has an average dwell time of 45 seconds, while other paragraphs have an average dwell time of only 15 seconds, that paragraph is marked as an area of ​​concentrated attention.

[0023] This pre-processed behavioral data is then fed into the bidding decision analysis network. The network employs a three-layer architecture. The first layer is a feature extraction layer, using 128 convolutional kernels to process spatial distribution features. The kernel size is 3x3, and the stride is 1. This layer specifically processes the distribution density data of deletion operations, identifying patterns in key areas of user interest through convolutional operations.

[0024] The second layer is the time-series processing layer, which employs a long short-term memory network structure. This layer contains 64 memory units and processes the temporal distribution data of modification operations. This network layer captures the temporal patterns of user operations through a memory gating mechanism. For example, it identifies that users tend to modify technical content in the morning and focus more on adjusting business terms in the afternoon.

[0025] The third layer is the decision output layer. This layer fuses the features extracted from the first two layers and generates user decision preference parameters through a fully connected neural network. The fully connected layer contains three hidden layers, with 256, 128, and 64 nodes per layer, respectively, using the ReLU activation function. The final output layer generates two-dimensional decision parameters: a preference parameter for choosing a particular option and a preference parameter for risk control.

[0026] In practical implementation, the solution selection preference parameter is calculated by analyzing how users handle technical and commercial solutions differently. For example, when a user frequently modifies the depth and breadth of the technical solution but makes fewer adjustments to the commercial terms, the system will determine that the user is more focused on the technical solution. The risk control preference parameter is inferred by analyzing the frequency with which users edit risk clauses, disclaimers, and other content. For example, if a user repeatedly modifies the content of risk response measures, the system will determine that the user has a high risk control need.

[0027] Throughout the process, the system employs a sliding window mechanism to update data, recalculating user decision preference parameters every 5 minutes to ensure timely capture of changes in user needs. All intermediate processing results are stored in an in-memory database for subsequent modules to access and analyze.

[0028] In another preferred embodiment of the present invention, the specific process of inputting user decision preference parameters into the bidding strategy generation model in step S3 is as follows: The bidding strategy encoder network adopts a bidirectional long short-term memory architecture, which consists of two independent LSTM networks, responsible for forward and backward sequence processing respectively. During the network initialization phase, the hidden state and cell state of the forward business state are set to zero vectors with a vector dimension of 256; similarly, the hidden state and cell state of the backward business state are also initialized to zero vectors of the same dimension.

[0029] The user decision preference parameter sequence is input into the encoder network in chronological order. Each business parameter vector contains 16 dimensions of feature data, including 8 dimensions of option selection preference parameters and 8 dimensions of risk control preference parameters. The sequence length is fixed at 32 time steps, corresponding to the 32 most recently recorded user decision preference snapshots in the system. For each input vector, the forward LSTM network processes it chronologically from time step 1 to time step 32, calculating the hidden state at each time step. Simultaneously, the backward LSTM network processes it in reverse order from time step 32 to time step 1, independently calculating its hidden state sequence.

[0030] In the first time step, the feedforward LSTM receives the initial hidden state and the first business parameter vector. Through input, forget, and output gates, it updates the cell state and outputs the first feedforward hidden state. The forget gate uses a sigmoid activation function to calculate the proportion of historical information retained, the input gate determines the degree of new information inclusion, and the output gate controls the output content of the current hidden state. The computation of these gating mechanisms involves a weight matrix and bias terms. The weight matrix has a dimension of 256 x 272, where 272 is the concatenation dimension of the 256-dimensional hidden state and the 16-dimensional input vector.

[0031] As sequence processing progresses, the forward LSTM gradually constructs a sequence of hidden states containing historical information about the sequence. For example, by the 16th time step, the forward hidden state has incorporated user preference features from the previous 16 time steps. Meanwhile, the backward LSTM starts processing from the end of the sequence, and by the 16th time step, its hidden state already contains feature information from time steps 17 to 32.

[0032] After sequence processing is complete, the forward LSTM outputs the hidden state vector at the final time step, which has a dimension of 256 and contains the forward context information of the entire sequence. The backward LSTM outputs its hidden state vector at the initial time step, also with 256 dimensions, containing the backward context information of the entire sequence. These two vectors are concatenated to form a complete 512-dimensional business feature vector.

[0033] The complete business feature vector is then fed into the bidding policy decoder network. The decoder network employs an attention-based sequence-to-sequence architecture, with its initial hidden states set to the complete business feature vector output by the encoder. The decoder processes all the encoder's hidden states through a multi-head attention mechanism, including 32 hidden states from the forward LSTM and 32 hidden states from the backward LSTM, for a total of 64 hidden states.

[0034] The attention mechanism computation is divided into 8 attention heads, each responsible for a different semantic space. For each attention head, the query vector, key vector, and value vector are first computed. The query vector comes from the current hidden state of the decoder, while the key and value vectors come from the set of hidden states of the encoder. Each attention head independently computes the attention weight distribution, which is then normalized using a softmax function to obtain the attention distribution for the 64 hidden states.

[0035] For example, the first attention head might focus more on hidden states related to technical solutions, with its attention weight exhibiting a higher value on the technical feature dimension; the second attention head might focus more on hidden states related to risk control, allocating more weight on the risk feature dimension. Each attention head performs a weighted summation of the value vectors based on its attention weights to generate the corresponding context vector.

[0036] The context vectors generated by the eight attention heads are concatenated into a 2048-dimensional comprehensive business context vector. This vector is then reduced to 512 dimensions through a linear layer and fused with the current hidden state of the decoder. The fused vector is processed by the tanh activation function and then input into the output layer.

[0037] The output layer is a fully connected neural network that maps a 512-dimensional input vector to a probability distribution of 1024 policy tokens. The policy token vocabulary contains common business strategy elements, such as differentiated pricing, emphasis on technological advantages, and risk mitigation strategies. Each policy token corresponds to a specific strategy description unit, such as using a tiered pricing strategy to highlight technological innovation capabilities and strengthen after-sales service commitments.

[0038] The decoder generates a sequence of policy descriptions in an autoregressive manner, starting with the initial token and progressively generating each policy token. Upon generating each new token, the decoder updates its hidden state and recalculates the attention weights. This generation process continues until an ending token is generated or the maximum sequence length of 50 tokens is reached.

[0039] Throughout the generation process, the system executes a beam search algorithm with a beam width of 5. This means the system maintains 5 candidate sequences simultaneously, each with a corresponding cumulative log probability. When a new token is generated, each candidate sequence expands into 5 new sequences with the highest probabilities, and then the 5 sequences with the highest cumulative probabilities are selected from the 25 candidate sequences to continue generation.

[0040] For example, when generating the 10th token, a candidate sequence might already contain 9 strategy tokens, such as those emphasizing technological innovation, cost control advantages, and highlighting localized services, with a cumulative logarithmic probability of -12.5. The system will calculate the probability distribution of the next token based on the current hidden state and attention context, and select the candidate tokens with the highest probabilities for expansion.

[0041] The system monitors the generation quality of candidate sequences in real time and optimizes the output results through the following mechanisms: First, the system calculates the repetition penalty for each candidate sequence, and reduces the probability weight of policy tokens that appear repeatedly; second, the system implements length normalization, dividing the cumulative probability by the power of alpha of the sequence length, where alpha is set to 0.8 to balance the impact of sequence length; finally, the system uses a coverage mechanism to prevent attention from repeatedly focusing on the same encoder state.

[0042] When all candidate sequences generate an end token or reach the maximum length, the system selects the sequence with the highest cumulative probability as the primary output. Simultaneously, the system retains the top 5 best sequences as alternative strategies. These strategy description sequences undergo a post-processing module to remove redundant semantic content, ensuring that each strategy description unit has a unique value proposition.

[0043] For example, a generated strategy description sequence might contain the following strategy token sequence: adopting a differentiated pricing strategy with a base price set at 95 percent of the market level, but using a premium strategy for core technology services with a premium of 15 percent, while emphasizing the advantages of a localized service team, providing a 24-hour response commitment, and proposing a three-tiered guarantee mechanism for risk control.

[0044] The final output set of candidate bidding strategies contains 3 to 5 complete strategy descriptions, each including core elements such as pricing strategy, key technical solutions, service commitments, and risk control measures. These strategy descriptions are sorted according to their comprehensive scores, which take into account the probability of sequence generation, strategy innovation, and the degree of matching with user preferences.

[0045] In another preferred embodiment of the present invention, the specific process of constructing the bidding scenario simulation environment in the scenario simulation module is as follows: First, complete case records from the past three years are loaded from the historical bidding project database. These records cover bidding projects in different industry sectors, including information technology services, engineering construction, and equipment procurement. The system loads 1,000 case records, each containing complete bid documents, evaluation criteria, competitor quotations, winning bid results, and subsequent implementation data. Environmental baseline parameters are set, including a 90-day simulation time span, daily simulation granularity, and market area division into five major economic zones.

[0046] The competitor's business model is constructed using an agent-based modeling approach. The system creates an independent business model instance for each competitor, with each instance containing complete attribute parameters and decision rules. Basic competitor attribute parameters include a competitor identifier, represented by an 8-bit code: the first two bits represent the industry category, the middle three bits represent the company size, and the last three bits represent the geographical distribution. Historical bidding behavior data is extracted from the database, including information such as the competitor's past bidding frequency, success rate, typical pricing strategies, and areas of technological advantage. Business decision preference parameters are derived from analyzing historical bidding data, including dimensions such as risk tolerance, innovation inclination, and price sensitivity, each quantified from 0 to 100.

[0047] The competitor decision-making rule system is implemented using a combination of rule engines and machine learning. The pricing calculation rules include a basic pricing calculation module and a dynamic adjustment module. The basic pricing calculation module generates an initial price using a random forest algorithm based on project size, complexity, and historical pricing data. For example, for an IT project with a budget of 5 million yuan, competitor A's initial price might be calculated as 85% of the budget, or 4.25 million yuan. The pricing adjustment module dynamically adjusts the price based on market feedback during the simulation process. The adjustment range is influenced by factors such as the competitor's current capacity utilization rate, market share target, and cash flow status.

[0048] The market reaction rules are implemented using a deep Q-learning algorithm trained on historical competitor behavior data. Each competitor's business model contains an independent neural network. The input layer includes 15 dimensions of market state features, such as current market share, recent winning bids, and bids from other competitors. The output layer includes eight possible action options: maintain bid, reduce bid by 3%, reduce bid by 5%, and increase bid by 2%. The network has three hidden layers, each with 128 nodes, using the ReLU activation function.

[0049] The risk assessment rules employ a multi-factor weighted evaluation model, considering three dimensions: project risk, competition risk, and execution risk. Project risk includes factors such as technical difficulty, schedule requirements, and payment terms; competition risk includes factors such as the number of competitors and their relative strengths; and execution risk includes factors such as the availability of internal resources and experience with similar projects. Each factor is assigned a different weight based on historical data, ultimately calculating a comprehensive risk coefficient.

[0050] When initializing competitor instance objects, the system assigns an initial set of business status parameters to each competitor, including current market share, financial status, technical capability score, and brand influence. Market share is set based on the competitor's market research data from the most recent quarter; financial status is derived by analyzing its publicly available financial statements and industry research data; technical capability score is based on its number of patents, the proportion of technical personnel, and the quality assessment of past projects; and brand influence is calculated comprehensively through market research and media exposure.

[0051] The bidding simulation environment is executed in multiple rounds, each representing a complete bidding cycle. The system first initializes market environment variables, including demand change parameters, price sensitivity parameters, and market capacity parameters. Demand change parameters are derived through time series analysis, taking into account factors such as seasonal fluctuations, industry cycles, and policy impacts. Price sensitivity parameters are calculated using market research data, with different price elasticity coefficients set for different industry sectors. Market capacity parameters are determined based on macroeconomic data and industry reports.

[0052] After inputting the set of candidate bidding strategies into the simulation environment, the system begins executing multi-round bidding competition simulations. Each simulation round comprises three distinct phases: the bidding phase, the evaluation phase, and the result confirmation phase. During the bidding phase, all competitors' business models generate bidding proposals in parallel based on the current market conditions, including technical proposals, commercial proposals, and pricing proposals. Technical proposals highlight their respective strengths, commercial proposals emphasize differentiated service commitments, and pricing proposals are calculated based on their respective decision-making rules.

[0053] The bid evaluation phase simulates the real-world project evaluation process. The system quantitatively scores each bid based on pre-set evaluation criteria. These criteria include three parts: technical solution score, commercial solution score, and price score, with weightings determined by the project type. For example, for a technologically innovative project, the technical solution might account for 50%, the commercial solution for 30%, and the price for 20%. The technical solution score uses natural language processing to analyze the innovativeness, feasibility, and completeness of the solution; the commercial solution score assesses after-sales service, delivery capabilities, and risk control measures; and the price score is calculated using a benchmark price method, with the average of all valid bids as the benchmark.

[0054] The results confirmation phase determines the winning bidder based on the evaluation scores and updates the business status of all competitors. The winning bidder's market share increases accordingly, its financial situation improves, and its brand influence strengthens. Unsuccessful bidders adjust their business strategies based on the bidding results, such as increasing investment in technology research and development, adjusting pricing strategies, or exploring new market areas.

[0055] Throughout the simulation, the system updates bidding market status data in real time, recording detailed pricing data, winning bid results, and project revenue data for each round and bidding strategy. Pricing data includes initial bids, final bids, and the process of bid adjustments; winning bid results data includes score details, rankings, and analysis of reasons for rejection; project revenue data includes direct profits, brand value enhancement, and assessment of potential business opportunities.

[0056] Simulation data is fully stored in the bidding analysis database, which uses a time-series data model. Each data point contains three fields: timestamp, data type, and data value. The system generates a unique session identifier for each simulation run, ensuring data integrity and traceability. For example, a complete simulation run may consist of 30 rounds, generating approximately 15,000 data records, covering the behavioral decisions and market reactions of all competitors in each round.

[0057] Competitor behavior log data is recorded using a structured storage method, recording decision input information and decision output results at each decision point in time. Decision input information includes a snapshot of the current market state, the company's own business status, and competitor dynamics; decision output results include specific actions taken, expected return assessments, and risk assessment conclusions. This log data is used for subsequent analysis and model optimization, helping to understand competitors' behavioral patterns and decision-making logic.

[0058] In another preferred embodiment of the present invention, the specific process of calculating the strategy competitiveness index based on market competition data during strategy execution in the strategy evaluation module is as follows: First, the complete dataset generated during the simulation is extracted from the bidding analysis database. This data is organized according to the simulation rounds, with each round containing detailed execution records of each bidding strategy. The system uses a 30-day analysis period to extract the pricing sequence data for all bidding strategies within that period. The pricing sequence data includes daily pricing records, pricing adjustment times, and adjustment amounts.

[0059] The system employs a dual-indicator evaluation system when calculating the commercial deviation of a price quote sequence from the market average. The absolute difference data is derived by calculating the difference between the strategy's price quote and the market average price at each point in time, with the difference value categorized as either positive or negative. For example, if a strategy's price quote at a certain point in time is 4.5 million yuan, while the market average price is 4.8 million yuan, the absolute difference is -300,000 yuan. The relative volatility data is derived by calculating the coefficient of variation of the price quote sequence. First, the standard deviation of each strategy's price quote sequence is calculated, then divided by the average value of the sequence to obtain the relative volatility. For example, if a strategy's standard deviation over 30 days is 250,000 yuan, and the average price is 4.6 million yuan, the relative volatility is 5.43%.

[0060] When analyzing market share time series data, the system uses a sliding window technique. Real-time growth data is derived by calculating the rate of change in market share between adjacent time points, specifically on a daily basis. This involves calculating the difference between today's market share and yesterday's market share, then dividing by yesterday's market share. For example, if a strategy's market share increased from 15% yesterday to 18% today, the real-time growth rate is 20%. Long-term trend data is derived through linear regression analysis. The system fits a trend line to 30 days of market share data and calculates the slope of that trend line. A positive slope indicates an upward trend in market share, while a negative slope indicates a downward trend.

[0061] The calculation of the scheme's advantage index employs a multi-factor weighted evaluation model. This model includes two primary indicators: the degree of business deviation and market share growth characteristics. The degree of business deviation indicator is further divided into absolute difference and relative fluctuation sub-items; the market share growth characteristics indicator is divided into immediate growth and long-term trend sub-items. Each sub-item is standardized and converted into a score from 0 to 100. The weighting is determined based on expert evaluation and historical data statistical analysis: the total weight of the degree of business deviation is 45%, with absolute difference accounting for 25% and relative fluctuation for 20%; the total weight of the market share growth characteristics is 55%, with immediate growth accounting for 30% and long-term trend for 25%.

[0062] The extraction of resource consumption time series data covers three dimensions: human resource input, financial input, and time input. Human resource input data includes the number of various personnel involved in the project and their working hours; financial input data includes direct cost expenditures and indirect cost allocation; and time input data includes the time for solution preparation, bid response time, and follow-up service time. The system aggregates this data on a weekly basis to form a time series of resource consumption.

[0063] Real-time efficiency data is derived by calculating the input-output ratio within a single week. Output indicators include the amount of bids won, the number of new customers acquired, and brand exposure during the period. For example, if 200 man-hours of human resources are invested in a week with a capital cost of 150,000 yuan, and a bid amount of 1 million yuan is won, then the real-time efficiency is output divided by input, which is calculated to be 3.33. Long-term efficiency data is derived by calculating the cumulative input-output ratio over the entire analysis period. It also considers the time value of money and uses the net present value method to standardize the inputs and outputs at different points in time.

[0064] The resource utilization efficiency index was calculated using the analytic hierarchy process (AHP). A judgment matrix was established to compare the three resource dimensions pairwise to determine the relative importance of each dimension. After consistency testing, the weights for human resource input were determined to be 35%, financial input 40%, and time input 25%. The efficiency values ​​for each dimension were standardized, and the final weighted sum was used to obtain the resource utilization efficiency index.

[0065] The business standardization process employs a minimum-maximum normalization method, mapping the raw values ​​of the scheme advantage index and resource utilization efficiency index to a standardized score range of 0 to 100. Specifically, the minimum and maximum values ​​of each index in the historical dataset are first determined, and then a formula is used to calculate the score for each data point. For example, if the historical minimum value of the scheme advantage index is 20 points and the maximum value is 90 points, and the raw value of a certain strategy is 75 points, then the standardized score would be 79.3 points.

[0066] The overall competitiveness score of the bidding strategy is calculated using a linear weighted combination, with the advantage index accounting for 60% and the resource utilization efficiency index accounting for 40%. The weighting is based on regression analysis of historical winning bid data to ensure a high correlation between the score and the actual winning bid rate. Each strategy ultimately receives a comprehensive score from 0 to 100, with higher scores indicating stronger competitiveness.

[0067] In another preferred embodiment of the present invention, the specific process of selecting the optimal bidding strategy based on the strategy competitiveness index in the strategy execution module is as follows: After receiving this comprehensive competitiveness score data, the strategy execution module first sorts the scores from highest to lowest and selects the bidding strategy with the highest score as the target bidding strategy. For example, in an evaluation, if strategy A scores 92 points, strategy B scores 87 points, and strategy C scores 83 points, then strategy A is selected as the target strategy.

[0068] When parsing the complete description of the target bidding strategy, the system uses natural language processing technology to identify key elements in the strategy text. These elements include the pricing strategy description, key technical solutions, service commitment terms, and risk control measures. The system establishes a mapping relationship between these elements and paragraphs in the bid document; for example, the pricing strategy description corresponds to the commercial pricing section of the bid document, and the key technical solutions section corresponds to the technical solutions section.

[0069] The modification location identification process is based on the standard template structure of the tender documents. The system maintains an index containing all standard chapters and paragraphs. Each paragraph has a unique identifier; for example, the second paragraph of section three in the technical solution chapter is identified as TECH3.2. The system determines the specific location that needs to be modified based on the correspondence between strategy elements and paragraph identifiers.

[0070] The set of bid document modification instructions is generated using a structured representation method. Each modification instruction contains four fields: operation type, target location, modification content, and formatting requirements. Operation types include insertion, deletion, and replacement; the target location uses paragraph identifiers for precise positioning; the modification content provides the specific text content; and the formatting requirements specify layout parameters such as font, font size, and spacing. For example, a specific modification instruction might be placed in the fifth paragraph of the commercial quotation section, replacing the original quotation description with a new tiered pricing scheme, and using SimSun font, size 12 (small four).

[0071] Document modification operations are executed using a transaction processing mechanism, ensuring that all modifications either succeed completely or are rolled back completely. The system first creates a copy of the tender document, and then executes each modification command sequentially on the copy. During execution, syntax checks and format validations are performed to ensure that the modified document meets the specifications. For example, replacement operations check whether the length of the new text is appropriate for the original position, and format adjustments verify whether the font and font size meet the standards.

[0072] The tender document processing engine employs template-based document generation technology. The engine incorporates multiple standard tender document templates covering various industries and project types. Based on the modified document content and formatting requirements, the engine selects the appropriate template for rendering. The rendering process includes steps such as text flow layout, chart insertion, and page number generation, ultimately outputting a tender document that conforms to industry standards.

[0073] Standardized output formats ensure that tender documents meet the technical requirements of the tendering party, including file format, font specifications, page layout, and table of contents structure. The system supports outputting tender documents in multiple formats, including PDF, Word, and HTML. A final quality check is performed before output, including spell checking, format consistency verification, and content integrity verification.

[0074] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A service platform for assisting in the preparation of tender documents based on automation technology, characterized in that, include: The user behavior collection module is used to collect user operation behavior data during the bid document preparation process. The operation behavior data includes content adjustment records, format modification frequency, and chapter access duration. The decision preference analysis module is used to process operational behavior data through the decision preference analysis model and generate user decision preference parameters, which include scheme selection tendency and risk control preference. The strategy generation module is used to input user decision preference parameters into the bidding strategy generation model, which generates a set of candidate bidding strategies based on historical bidding cases. The scenario simulation module is used to build a bidding scenario simulation environment, execute various strategies in the candidate bidding strategy set in the bidding scenario simulation environment, and record market competition data during the strategy execution process; The strategy evaluation module is used to calculate the strategy competitiveness index based on market competition data during the strategy execution process. The strategy competitiveness index includes the solution advantage and resource utilization efficiency. The strategy execution module is used to select the optimal bidding strategy based on the strategy competitiveness index, transform the optimal bidding strategy into a specific bid document preparation plan, and execute it.

2. The service platform for assisting in the preparation of tender documents based on automation technology as described in claim 1, characterized in that, In the decision preference analysis module, the specific process of processing operational behavior data through the decision preference analysis model is as follows: Collect data on the deletion locations of content by users during the editing process of tender documents, and statistically analyze the distribution density of deletion operations in each chapter and paragraph; combine the time series information in the text modification frequency data to establish a time distribution map of modification operations; analyze the duration characteristics in the paragraph dwell time data to identify areas where user attention is concentrated. The data on the distribution density of deletion operations, the time distribution of modification operations, and the concentrated areas of user attention are input into the bidding decision analysis network. The bidding decision analysis network includes a feature extraction layer, a time series processing layer, and a decision output layer. The feature extraction layer calculates spatial distribution features, the time series processing layer analyzes time series patterns, and the decision output layer generates user decision preference parameters, including scheme selection tendency parameters and risk control preference parameters.

3. The service platform for assisting in the preparation of tender documents based on automation technology as described in claim 1, characterized in that, In step S3, the specific process of inputting user decision preference parameters into the bidding strategy generation model is as follows: The user decision preference parameter sequence is input into the bidding strategy encoder network. The bidding strategy encoder network uses a business decision long short-term memory architecture to process the sequence data, calculates the forward business feature sequence and the backward business feature sequence respectively, and merges the forward business feature sequence and the backward business feature sequence to generate a complete business feature vector. The complete business feature vector is passed to the bidding strategy decoder network. The bidding strategy decoder network uses a business attention mechanism to calculate business context vectors of different dimensions and integrates the business context vectors of different dimensions into a comprehensive business context vector. Based on the comprehensive business context vector, a bidding strategy description sequence is generated step by step, with each strategy description unit corresponding to a specific business strategy element. A business strategy search process is performed on the bidding strategy description sequence, multiple candidate strategy description sequences are maintained and the sequence generation probability is calculated in real time. The number of candidate strategies is dynamically adjusted according to the sequence generation probability, duplicate business strategy descriptions are removed, and an optimized set of candidate bidding strategies is output.

4. The service platform for assisting in the preparation of tender documents based on automation technology as described in claim 3, characterized in that, The specific process by which the bidding strategy encoder network processes the user decision preference parameter sequence is as follows: Initialize the forward and backward business states of the business decision long short-term memory network; sequentially input each business parameter vector from the user decision preference parameter sequence, and update the forward and backward business state sequences respectively; Calculate the final state vector of the forward business state sequence, and calculate the initial state vector of the backward business state sequence. The final state vector of the forward business state sequence and the initial state vector of the backward business state sequence are merged into a complete business feature vector. Attention distribution data is calculated using a business attention mechanism to generate an enhanced business context vector; the final business feature vector and business context vector are then output to the bidding strategy decoder network.

5. The service platform for assisting in the preparation of tender documents based on automation technology according to claim 1, characterized in that, The specific process of constructing the bidding scenario simulation environment in the scenario simulation module is as follows: Initialize the basic parameters of the virtual bidding environment and load complete case records from the historical bidding project database; configure multiple competitor business models, and set the bidding strategy parameter set and market reaction parameter set for each competitor business model. The bidding strategy parameter set includes initial price parameters, price adjustment parameters, and price limit parameters, while the market reaction parameter set includes reaction time parameters and competition intensity parameters; define bidding market environment variables, including demand change parameters, price sensitivity parameters, and market capacity parameters. The candidate bidding strategy set is sequentially input into the bidding simulation environment to execute multiple rounds of bidding competition simulation. Each round of simulation includes a bidding stage, a bid evaluation stage, and a result confirmation stage. During each round of simulation, the bidding market status data is updated in real time, and detailed price data, winning bid result data, and project revenue data for each bidding strategy are recorded. The complete bidding simulation process data is stored in the bidding analysis database.

6. The service platform for assisting in the preparation of tender documents based on automation technology according to claim 5, characterized in that, The specific process for configuring the competitor's business model is as follows: Define the basic attribute parameters of competitors, which include competitor identifiers, historical bidding behavior data, and business decision preference parameters. Set up a competitor decision-making rule system, which includes pricing calculation rules, market reaction rules, and risk assessment rules; initialize competitor instance objects and allocate an initial set of business state parameters. During the bidding simulation, the decision-making rule system is invoked in real time to generate pricing decision data based on the current bidding market status. Update the internal status parameters of competitors and record complete competitor behavior log data, including decision time points, decision input information, and decision output results.

7. The service platform for assisting in the preparation of tender documents based on automation technology according to claim 1, characterized in that, In the strategy evaluation module, the specific process of calculating the strategy competitiveness index based on market competition data during strategy execution is as follows: Extract complete price series data for each bidding strategy from the bidding analysis database, calculate the commercial deviation of the price series from the market average price, including absolute difference data and relative volatility data; analyze the market share time series data for each bidding strategy, and calculate the real-time growth data and long-term trend data of market share. By combining the degree of commercial deviation of the quoted price and the characteristics of market share growth, a solution advantage index is generated through commercial weighting. Simultaneously, time-series data of resource consumption for each bidding strategy are extracted, and real-time and long-term efficiency data of resource input and output are calculated to generate resource utilization efficiency indicators. The scheme advantage index and resource utilization efficiency index are commercially standardized, and a comprehensive competitiveness score of the bidding strategy is generated through business combination calculation.

8. The service platform for assisting in the preparation of tender documents based on automation technology according to claim 1, characterized in that, In the strategy execution module, the specific process of selecting the optimal bidding strategy based on the strategy competitiveness index is as follows: Receive comprehensive competitiveness score data for bidding strategies, select the highest-scoring bidding strategy as the target bidding strategy; parse the complete content description of the target bidding strategy, identify the paragraph positions and content modification requirements that need to be adjusted in the bid documents; generate a detailed set of bid document modification instructions, which includes modification location information, modification content text, and modification format requirements; Execute document modification operations according to the set of tender document modification instructions to complete the update of tender document content and adjustment of format specifications; The final document is generated using a tender document processing engine, which outputs a standardized tender document in accordance with the tender document specification requirements.

Images