Bid invitation information accurate matching and pushing system based on artificial intelligence

Abstract

The invention discloses a bid invitation information accurate matching and pushing system based on artificial intelligence, and relates to the field of intelligent bid invitation. Comprising an enterprise portrait construction module used for constructing an enterprise semantic portrait vector; the bid invitation information processing module is used for constructing a bid invitation semantic vector; the semantic matching module is used for calculating the semantic similarity between the enterprise semantic portrait vector and the bid invitation semantic vector and generating a correlation score; the competition perception recommendation module is used for constructing a competition perception model based on group game, regarding a plurality of enterprises as game participants, constructing a dynamic game model, simulating a bid invitation competition process by using a multi-agent reinforcement learning algorithm, and outputting expected bid winning income of each enterprise in each bid invitation; and the sorting and pushing module is used for sorting the bid invitation information according to the correlation score and the expected bid winning income and pushing the bid invitation information to the enterprise. The method solves the problems that a traditional method is inaccurate in matching and cannot adapt to competition dynamics.

Description

Technical Field

[0001] The present invention relates to the field of intelligent bidding, and more specifically, to a precise matching and pushing system for bidding information based on artificial intelligence. Background Art

[0002] Bidding information plays an important role in modern business activities. Especially for enterprises that need to obtain engineering projects, equipment procurement, or service contracts, it is crucial to timely obtain bidding opportunities that match their capabilities. With the development of digital technologies, the number of released bidding information has increased rapidly, covering a wide range of fields and with complex content. However, traditional information acquisition methods mainly rely on manual screening or simple keyword matching. This method often fails to accurately understand the deep relationship between bidding requirements and enterprise capabilities, resulting in enterprises missing potential business opportunities. In addition, the competition in the bidding market is intensifying. Enterprises not only need to focus on the fit between themselves and project requirements but also consider the participation of competitors, and the existing technologies lack effective support in this regard. At the same time, some bidding announcement information is expressed vaguely or has an incomplete structure, further increasing the difficulty of matching. Therefore, how to achieve precise matching in the massive information and adapt to the dynamic competition environment has become a key problem to be solved urgently. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a precise matching and pushing system for bidding information based on artificial intelligence to solve the problems mentioned in the background art.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A precise matching and pushing system for bidding information based on artificial intelligence, comprising:

[0006] An enterprise portrait construction module, configured to collect enterprise information and construct an enterprise semantic portrait vector;

[0007] A bidding information processing module, configured to collect bidding announcement information and construct a bidding semantic vector;

[0008] A semantic matching module, configured to generate a relevance score between a bidding project and an enterprise based on the calculation of the semantic similarity between the enterprise semantic portrait vector and the bidding semantic vector;

[0009] A competition-aware recommendation module, configured to construct a competition-aware model based on group game, regard multiple enterprises as game participants, construct a dynamic game model, use a multi-agent reinforcement learning algorithm to simulate the bidding competition process, and output the expected winning bid benefits of each enterprise in each bidding;

[0010] The sorting and pushing module is used to sort and push tender information to enterprises based on the comprehensive relevance score and the expected winning bid income. The sorting methods include any one or more of the following: relevance score sorting, expected winning bid income sorting, and comprehensive score sorting; where the comprehensive score is the weighted average of the relevance score and the expected winning bid income. The weights for weighting can be selected according to specific circumstances, such as each being one-half, etc., and enterprises can customize the selection.

[0011] In an alternative embodiment, the dynamic game model uses a revenue function to calculate the expected winning bid income r of enterprise i i , and the calculation formula of the revenue function is:

[0012]

[0013] where s i represents the relevance score between the enterprise and the tender project, and a j represents the participation decision of the enterprise. When the value is 1, it means participation, and when the value is 0, it means non-participation.

[0014] In an alternative embodiment, the semantic similarity calculation uses cosine similarity to measure the closeness between the enterprise semantic portrait vector and the tender semantic vector. The calculation formula is:

[0015]

[0016] where e represents the enterprise semantic portrait vector, t represents the tender semantic vector, e·t represents the dot product of the two vectors, ∥e∥ represents the Euclidean norm of the enterprise semantic portrait vector, and ∥t∥ represents the Euclidean norm of the tender semantic vector.

[0017] In an alternative embodiment, the enterprise information includes any one or more of the following: the industry it belongs to, the main business, the service area, the historical bidding records, and the winning bid cases.

[0018] In an alternative embodiment, the system further includes:

[0019] The enterprise behavior and signal acquisition module is used to collect the behavior data of the enterprise in the system, including any one or more of the following: the search history of the enterprise, the browsing records, and the click behavior.

[0020] In an alternative embodiment, the system further includes:

[0021] The enterprise intention modeling module uses a variational autoencoder to process the collected enterprise behavior data and external business signals, generates an implicit latent distribution of the enterprise intention, and combines time series analysis techniques to analyze the change of the implicit latent distribution over time through a long short-term memory network to identify the change trend of the enterprise intention.

[0022] In an alternative embodiment, the system further comprises:

[0023] A feature weight dynamic adjustment module that dynamically adjusts the feature weights in the semantic matching module based on the identified trend of changes in enterprise intentions. By increasing or decreasing the feature weights related to emerging interests of the enterprise, the recommendation accuracy for the potential interest directions of the enterprise is improved.

[0024] In an alternative embodiment, the system further comprises:

[0025] A tender information generation module for inputting a tender announcement input generator with missing structure and generating background context information related to the tender announcement according to industry corpus and historical announcement data.

[0026] In an alternative embodiment, the system further comprises:

[0027] A generated content verification module that uses a convolutional neural network or a BERT model as a discriminator, receives the original tender announcement and the background context information generated by the generator, and outputs a consistency score.

[0028] In an alternative embodiment, the generator introduces industry domain constraints during the training process. By adding domain consistency loss in the training of the generator, it is urged that the generated content contains industry keywords and conforms to the terms and specifications of a specific industry.

[0029] The advantages of the present invention over the prior art are that the present invention significantly improves the accuracy and practicality of tender information matching by constructing a competition perception model based on group game and combining multi-agent reinforcement learning technology. This method not only recommends based on the semantic similarity between the enterprise and the tender project, but also evaluates the expected winning bid income of the enterprise in each project by simulating the competition process, so as to preferentially push opportunities with high matching degree and high winning probability, effectively solving the problem that traditional matching methods ignore market competition and helping enterprises increase the winning bid probability. In addition, the system can also dynamically adjust the feature weights according to enterprise behavior data, capture the changing trend of enterprise intentions, and further optimize the personalization and timeliness of the recommendation results. For the situation where the tender announcement information is incomplete, the system enhances the semantic understanding ability and improves the accuracy and reliability of the matching by generating background context and verifying its consistency. These improvements together help enterprises respond quickly and optimize decisions in the complex market. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the overall process of the tender information precise matching and pushing system of the present invention;

[0031] Figure 2 is the processing steps of the enterprise portrait construction module of the present invention;

[0032] Figure 3It is the semantic vector generation process of the tender information processing module of the present invention;

[0033] Figure 4 It is the game and learning process of the competition-aware recommendation module of the present invention;

[0034] Figure 5 It is the dynamic process of enterprise intention modeling and weight adjustment of the present invention. Detailed implementation manners

[0035] The following combines the attached Figure 1 ~the attached Figure 5 to describe the detailed implementation manners of the present invention.

[0036] In the field of tender information matching and pushing, traditional methods often only focus on the superficial matching degree between enterprises and tender projects, such as keyword coincidence or industry relevance. Although this method is simple and intuitive, it cannot reflect the actual competition dynamics in the market - even if a project highly matches an enterprise, if the competitors are strong, the winning probability may be greatly reduced. Therefore, simply relying on the matching degree to push information often leads to low efficiency in the investment of enterprise resources. To solve this problem, the present invention introduces the group game theory, regards enterprises as game participants in the competitive environment, and comprehensively evaluates the winning potential by analyzing enterprise capabilities, project characteristics and the behaviors of potential opponents, so as to push more strategically valuable tender information.

[0037] As Figure 1 shown is the overall flowchart of the tender information precise matching and pushing system based on artificial intelligence of the present invention, which shows the complete process from enterprise information collection to final recommendation, including key steps such as data processing, semantic matching and competition analysis.

[0038] The more specific modules are as follows:

[0039] Enterprise portrait construction module, which is responsible for collecting multi-dimensional business feature information of enterprises, including the industry they belong to (such as construction, IT), main business (such as engineering contracting, software development), service area (such as East China region), historical bidding records (such as participation times, amount) and winning bid cases (such as the number and type of successful projects). Based on this information, using the deep semantic embedding technology, the enterprise features are transformed into a high-dimensional semantic portrait vector e. This vector not only captures the explicit features of the enterprise, but also can reflect its potential capabilities through semantic relationships. As Figure 2 shown is the flowchart of the enterprise portrait construction module of the present invention, which details the process of collecting multi-dimensional information of enterprises and generating the semantic portrait vector, including the integration and coding of industry, business, region and historical data.

[0040] Tendering information processing module, which grabs announcement data from multiple tendering platforms in real time, such as procurement websites or industry vertical websites. Natural language processing is performed on each tendering text, including word segmentation, noise removal (such as removing irrelevant punctuation), entity recognition (such as extracting project type, budget), and finally a multi-dimensional semantic vector t is generated. This vector condenses the core requirements and features of the tendering project, facilitating subsequent matching calculations. For example, Figure 3 Shown is the flowchart of the tendering information processing module of the present invention, demonstrating the complete steps from multi-source data grabbing to generating the tendering semantic vector, including processing procedures such as word segmentation, noise removal, and entity recognition.

[0041] Semantic matching module, which calculates the similarity between the enterprise semantic portrait vector e and the tendering semantic vector t, and the system generates a correlation score. This score reflects the degree of fit between the enterprise's capabilities and the project requirements. The calculation process relies on a deep semantic embedding model to ensure that the similarity is not only based on literal matching but also can understand deep semantic associations.

[0042] Competition-aware recommendation module, which is the core of the system of the present invention and simulates the market competition environment through the group game theory. The specific steps include:

[0043] Regarding multiple enterprises participating in the same tender as game participants, and each enterprise has its own capability portrait. Taking the enterprise's capability portrait, tendering project characteristics, and the behavior patterns of potential competitors (such as historical participation frequency, bidding strategies) as inputs, a dynamic game model is constructed.

[0044] Adopting a multi-agent reinforcement learning algorithm to simulate the competition process of enterprises in tendering and predict the expected winning bid revenue of enterprises in each tender. The revenue not only depends on the matching degree but is also closely related to the competition intensity.

[0045] Sorting and pushing module, which comprehensively considers the correlation score s and the expected winning bid revenue r to sort all tendering information. The sorting result preferentially pushes those projects that are highly matched with the enterprise and have high winning bid potential, ensuring that the enterprise can focus its resources on the optimal opportunities. For example, Figure 4 Shown is the flowchart of the competition-aware recommendation module of the present invention, which details the process of simulating competition based on the group game theory and multi-agent reinforcement learning, including model construction, probability adjustment, and revenue calculation.

[0046] Among them, the core of the present invention is that the system regards all potential enterprises that may participate in a certain tendering project as independent game participants, and each enterprise participates in the competition according to its own capability portrait and historical behavior pattern. The decision of the enterprise is simplified into two choices: participating in a certain tendering project or choosing not to participate. And the expected return of each enterprise under different choices is quantified through a carefully designed revenue function.

[0047] Specifically, let the set of all participating enterprises be N, and the strategy of each enterprise i be represented by a i where a i = 1 indicates participation, and a i = 0 indicates non - participation. The matching degree between enterprise i and the tender project is represented by s i , and this value is calculated from the tender requirements and the enterprise's ability profile through semantic similarity analysis.

[0048] Based on this, the revenue function r i is defined as:

[0049]

[0050] The core of this function is that the revenue not only depends on the enterprise's own matching degree but also is affected by the competition intensity of other participating enterprises. The competition intensity is reflected by ∑ j≠i a j s j .

[0051] To determine a stable competition pattern, the system introduces the concept of Nash equilibrium and finds a state where no party has an incentive to unilaterally deviate by iteratively adjusting the strategies of each enterprise. In the implementation process, the system first estimates the initial participation probability of each enterprise based on historical bidding data, and then gradually adjusts these probabilities through a simulated game process.

[0052] For example, for a specific tender project, the system initializes the participation probability, then calculates the expected revenue of each enterprise according to the current strategy combination, and adjusts the probability in the direction of the strategy with higher revenue. This process is repeated until the participation probabilities of all enterprises tend to be stable. Finally, the system uses the results of equilibrium analysis to estimate the competition intensity and combines the enterprise's ability characteristics to calculate its expected revenue in this tender project, providing a basis for the recommendation decision.

[0053] Furthermore, multi - agent reinforcement learning dynamically optimizes the decision - making strategies of enterprises to adapt to the changing competition environment in the tender market. In this system, each enterprise is modeled as an agent. The system defines the state of the agent according to the characteristics of the current tender project (such as project type, budget scale), the participation of competitors, and the enterprise's own historical performance. The decision of the agent is also simplified to two actions: participation or non - participation, and the reward function follows the revenue calculation method in the group game, that is:

[0054]

[0055] To enable the agent to learn the optimal strategy, the system adopts the Q-learning algorithm. Each agent maintains a Q-table to record the expected long-term rewards for taking different actions in different states. The update of the Q-table follows the following rules:

[0056]

[0057] Among them, s represents the current state, s' represents the next state, a represents the current action, r represents the immediate reward, α is the learning rate, and γ is the discount factor.

[0058] In the training phase, the system uses historical bidding data to simulate multiple rounds of games. In each round, each agent selects an action according to the current Q-table, and the environment calculates the reward and updates the state based on the actions of all agents. Through multiple rounds of iteration, the agent gradually learns the optimal strategy in different competitive scenarios. After training is completed, the agent can make better decisions based on the learned knowledge in future bidding projects.

[0059] To improve the learning efficiency, an exploration-exploitation mechanism can also be introduced, such as the ε-greedy strategy. This mechanism ensures that the agent can still try new strategies while using the existing knowledge by setting a certain probability of random exploration. In addition, considering the dynamics of the bidding market, the system will regularly retrain the model with the latest data to maintain the effectiveness and adaptability of the strategy.

[0060] The above technical solutions can be understood through a specific embodiment:

[0061] Suppose there is an enterprise A that mainly engages in infrastructure construction and is particularly good at transportation system engineering. Recently, two bidding projects have emerged in the market. One is a smart transportation system construction project with a large scale, attracting many powerful enterprises to participate; the other is a road maintenance project in a certain area with a small scale and relatively few competitors. Enterprise A has a certain degree of match for both projects, but has limited resources and can only focus on one of them. Faced with this situation, the system recommends the optimal choice for Enterprise A by analyzing the competitive pattern and expected benefits.

[0062] First, the system regards each tender project as a competitive scenario, and both Company A and other potential competitors are regarded as participants. Each company needs to decide whether to invest resources to participate in the tender, and the benefits depend not only on the matching degree between the company itself and the project requirements, but also on the intensity of competition from other participants. Taking the intelligent transportation system project as an example, assume that the matching degree of Company A is 0.8, while the matching degrees of the main competitors, Company B and Company C, are 0.9 and 0.7 respectively. If both Company B and C choose to participate, then the competition faced by Company A will be very intense. The system quantifies this impact through a benefit function, and the formula takes into account the dilution effect of the total matching degree of all participants on individual benefits. After calculation, the expected benefit of Company A in this project is approximately 0.31, reflecting that the competition has weakened its return.

[0063] In contrast, the competition environment for the road maintenance project is much milder. Assume that the matching degree of Company A is 0.7, while the matching degree of the only competitor, Company D, is 0.5, and no other companies participate. The system calculates that the expected benefit of Company A in this project is approximately 0.47, which is obviously higher than that of the intelligent transportation system project. This shows that although the matching degree of the road maintenance project is slightly lower, the lower competition intensity makes it a more attractive option.

[0064] To further optimize the decision-making, the system not only relies on static analysis but also introduces a dynamic learning mechanism. By simulating multiple rounds of competition, the system allows multiple companies to try to participate or withdraw from multiple tender projects and adjusts strategies according to the benefits. For example, in the intelligent transportation system project, if the number of competitors is large and the benefits of Company A remain low, the system will gradually tend to avoid similar high-competition projects; while in the road maintenance project, due to the consistently high benefits, the system will strengthen the tendency to participate. This method enables Company A to find the best strategy in the complex and changing market. Such a strategy is applied to multiple companies and multiple tender projects simultaneously, thus forming a better balance in the system, making the recommendation of the tender result consider both the benefits and the success rate.

[0065] In another embodiment, the present invention also lays an analysis foundation by collecting the behavioral data of companies on the platform. The enterprise behavior and signal collection module is responsible for recording various types of interaction information of companies within the system, and this information can include any one or more of the company's search history, browsing records, and click behaviors. For example, a company may enter keywords such as "industrial automation equipment" in the search bar, or frequently browse tender announcements related to intelligent manufacturing, or even click to view the specific requirements of certain projects. These behavioral data provide valuable clues for the system, reflecting the company's current business interests or potential needs.

[0066] In a further embodiment, in order to deeply understand the needs of enterprises, such as Figure 5As shown, the present invention introduces an enterprise intention modeling module. This module uses a variational autoencoder to process the collected enterprise behavior data and external business signals, generating an implicit latent distribution of enterprise intentions. The advantage of the variational autoencoder is that it can compress complex high-dimensional data into a low-dimensional representation form, capturing the core features of enterprise interests. For example, an enterprise may be interested in multiple fields simultaneously. The variational autoencoder generates a compact vector by analyzing the patterns in the behavior data, reflecting the main intention direction of the enterprise. In addition, the interests of enterprises often change over time. Therefore, the system combines time series analysis techniques and analyzes the changes in the implicit latent distribution over time through a long short-term memory network to identify the dynamic trends of enterprise intentions. For example, if an enterprise has gradually increased its searches and views on "renewable energy" projects over a certain period of time, the long short-term memory network can predict through the evolution of time series data that the enterprise's future focus may further concentrate on this field. This dynamic analysis ensures that the system's perception of enterprise intentions always remains real-time and forward-looking.

[0067] Based on the changing trends of enterprise intentions, the system of the present invention can flexibly adjust the recommendation strategy to ensure that the pushed tender information highly matches the latest needs of the enterprise.

[0068] The feature weight dynamic adjustment module achieves this goal by dynamically adjusting the feature weights in the semantic matching module. Specifically, when the system identifies that an enterprise's interest in a certain emerging field has increased, it will increase the feature weights related to this field, and vice versa, reduce the weights of irrelevant features, thereby improving the recommendation accuracy for the potential interest direction of the enterprise. For example, an enterprise was originally focused on traditional manufacturing, but recently began to explore the new energy equipment field. The system will automatically increase the weights of new energy-related features, making the recommendation results more inclined to push tender information for new energy projects. This dynamic adjustment mechanism not only improves the accuracy of recommendations but also enhances the system's adaptability to changes in enterprise needs.

[0069] In another embodiment, the present invention also pays attention to the problem of incomplete information. The tender information generation module solves this problem by inputting the tender announcements with missing structures into a generator and generating background context information related to the tender announcements based on industry corpora and historical announcement data. The generator uses rich external data to infer the missing details. For example, for a simple announcement "Smart city construction project", the generator may supplement information such as "The project may involve Internet of Things technology and data platform construction". These supplementary contents make the semantics of the announcement more complete, facilitating subsequent matching and recommendation by the system.

[0070] However, the generated content must maintain consistency with the original announcement to avoid deviating from the topic or introducing incorrect information. The generated content verification module uses a convolutional neural network or a BERT model as a discriminator, receives the original tender announcement and the background context information generated by the generator, and outputs a consistency score. This score reflects the semantic fit between the generated content and the original announcement. For example, if the generated content is irrelevant to the announcement topic, the discriminator will output a low score, and the system will reject the content or require the generator to readjust. Through this verification mechanism, the system ensures the reliability and practicality of the generated information.

[0071] To further improve the quality of the generated content, industry domain constraints are introduced during the training of the generator. The specific method is to add a domain consistency loss during training, prompting the generated content to contain industry keywords and conform to the terms and specifications of a specific industry. This constraint ensures that the generated information is not only rich in content but also highly professional. For example, in the energy industry, the generated content may use terms such as "power generation efficiency" and "energy conversion rate", while in the construction industry, it tends to mention expressions such as "construction period" and "material specifications". The domain consistency loss promotes the generator to output context information that conforms to industry norms by calculating the occurrence frequency of industry keywords and weighting their importance.

[0072] Through the collaborative operation of the above modules, the system of the present invention realizes the accurate capture of the enterprise's intention and the intelligent enhancement of tender information, and finally provides high-quality recommendation services for users. This technical solution not only improves the accuracy of matching but also helps enterprises respond quickly to opportunities in a complex market environment through dynamic adjustment and information supplementation.

[0073] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An accurate matching and pushing system for tender information based on artificial intelligence, characterized in that Including: An enterprise portrait construction module, which is used to collect enterprise information and construct an enterprise semantic portrait vector; A tender information processing module, which is used to collect tender announcement information and construct a tender semantic vector; A semantic matching module, which is used to calculate the semantic similarity between the enterprise semantic portrait vector and the tender semantic vector, and generate a correlation score between the tender project and the enterprise; A competition-aware recommendation module, which is used to construct a competition-aware model based on group game, regard multiple enterprises as game participants, construct a dynamic game model, use a multi-agent reinforcement learning algorithm to simulate the tender competition process, and output the expected winning bid revenue of each enterprise in each tender; A sorting and pushing module, which is used to sort the tender information based on the comprehensive correlation score and the expected winning bid revenue and push it to the enterprise. The sorting methods include any one or more of the following: sorting by correlation score, sorting by expected winning bid revenue, and sorting by comprehensive score; where the comprehensive score is the weighted average of the correlation score and the expected winning bid revenue.

2. The precise matching and pushing system for tendering information based on artificial intelligence according to claim 1, characterized in that, The dynamic game model calculates the expected winning bid revenue r of enterprise i using the revenue function i , and the calculation formula of the revenue function is as follows: where s i represents the relevance score of the enterprise to the tender project, a j represents the participation decision of the enterprise. When the value is 1, it means participation; when the value is 0, it means non-participation.

3. The artificial intelligence-based tender information precise matching and pushing system according to claim 1, characterized in that The semantic similarity calculation uses cosine similarity to measure the closeness between the enterprise semantic portrait vector and the tender semantic vector. The calculation formula is: where, e represents the enterprise semantic portrait vector, t represents the tender semantic vector, e·t represents the dot product of the two vectors, ∥e∥ represents the Euclidean norm of the enterprise semantic portrait vector, and ∥t∥ represents the Euclidean norm of the tender semantic vector.

4. The precise matching and pushing system for tendering information based on artificial intelligence according to claim 1, characterized in that, The enterprise information includes any one or more of the following: industry, main business, service area, historical bidding records, winning bid cases.

5. The accurate matching and pushing system for tender information based on artificial intelligence according to claim 1, characterized in that, The system further includes: An enterprise behavior and signal collection module, which is used to collect the behavior data of the enterprise in the system, including any one or more of the following: the search history, browsing records, and click behaviors of the enterprise.

6. The artificial intelligence-based tender information precise matching and pushing system according to claim 5, wherein, The system further includes: An enterprise intention modeling module, which uses a variational autoencoder to process the collected enterprise behavior data and external business signals, generates an implicit latent distribution of the enterprise intention, and combines time series analysis techniques to analyze the change of the implicit latent distribution over time through a long short-term memory network to identify the changing trend of the enterprise intention.

7. The artificial intelligence-based tender information precise matching and pushing system according to claim 6, characterized in that, The system further includes: A feature weight dynamic adjustment module, which dynamically adjusts the feature weights in the semantic matching module based on the identified changing trend of the enterprise intention, and improves the recommendation accuracy of the potential interest direction of the enterprise by increasing or decreasing the feature weights related to the emerging interests of the enterprise.

8. The precise matching and pushing system for tendering information based on artificial intelligence according to claim 1, characterized in that The system further includes: A tender information generation module, which is used to input the tender announcement with missing structure into a generator, and generate background context information related to the tender announcement according to industry corpus and historical announcement data.

9. The system for precise matching and pushing of tender information based on artificial intelligence according to claim 8, wherein, The system further includes: A generated content verification module, which uses a convolutional neural network or a BERT model as a discriminator, receives the original tender announcement and the background context information generated by the generator, and outputs a consistency score.

10. The artificial intelligence-based tender information precise matching and pushing system according to claim 9, characterized in that During the training process of the generator, industry domain constraints are introduced. By adding domain consistency loss to the training of the generator, it is urged that the generated content contains industry keywords and conforms to the terms and specifications of a specific industry.

Images