Deep learning-based bid inviting and purchasing service capability intelligent matching method
By using deep learning technology to intelligently match bidding and procurement service capabilities, and by generating dynamic scores using pre-trained language models and attention mechanisms, the problem of inaccurate matching in existing technologies is solved, and more efficient bidding and procurement service recommendations are achieved.
Patent Information
- Application Number
- CN202511676449.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-02-10
AI Technical Summary
In existing technologies, keyword-based bidding and procurement service matching methods cannot accurately identify the service provider's deep semantic capabilities and their suitability for bidding needs, which affects the accuracy and effectiveness of the recommended list.
A deep learning-based approach is adopted, which uses a pre-trained language model to perform deep semantic encoding on bidding requirements and service provider capability documents. Combined with the real-time operational status of service providers and business rules, a dynamic matching score is generated, and an interpretable report is generated through an attention mechanism to build an optimization closed loop.
It improves the accuracy and effectiveness of the recommended list, accurately captures the deep semantic fit between bidding needs and service provider capabilities, corrects deviations in the matching process, and continuously improves matching performance through user feedback.
Smart Images
Figure CN121501980A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of bidding and procurement service matching technology, and specifically relates to an intelligent matching method for bidding and procurement service capabilities based on deep learning. Background Technology
[0002] In the current bidding and procurement service matching process, most of the work involves building a keyword library based on industry experience, and assessing the degree of fit between service providers and bidding needs by statistically analyzing the frequency, position and co-occurrence relationship of keywords. However, the above-mentioned matching methods rely on the surface form of words, which differs from the inherent semantic richness of natural language. When faced with complex, novel, or descriptive requirements, the matching results are prone to bias and it is difficult to accurately identify the root causes of subtle differences in service providers' experience or professional capabilities in similar projects. As a result, the matching results cannot fully reflect the degree of semantic fit between the service providers' actual capabilities and the bidding requirements at a deeper level, which in turn affects the accuracy and effectiveness of the subsequent recommended list. Summary of the Invention
[0003] This application provides an intelligent matching method for bidding and procurement service capabilities based on deep learning, which effectively solves the problem in existing technologies where the accuracy and effectiveness of recommended lists are affected by the reliance on superficial keyword matching and the difference between it and the inherent semantic richness of natural language. This method can improve the accuracy and effectiveness of recommended lists.
[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application provides a deep learning-based intelligent matching method for bidding and procurement service capabilities, including: Obtain bidding requirements documents and service provider capability documents, and use a pre-trained language model to perform deep semantic encoding on the bidding requirements documents and service provider capability documents to generate a comprehensive semantic matching score.
[0005] The system integrates the real-time operational status of service providers and dynamically adjusts the comprehensive semantic matching score based on the real-time operational status of the service providers to generate a dynamic matching score.
[0006] The dynamic matching score is validated for compliance and business rules are applied to generate the final recommendation ranking.
[0007] An interpretability description is generated for the final recommendation ranking based on an attention mechanism, forming a matching result report.
[0008] Collect user feedback data on the matching result report.
[0009] The pre-trained language model is fine-tuned using the feedback dataset to complete the model optimization loop.
[0010] Furthermore, a pre-trained language model is used to perform deep semantic encoding on the bidding requirements document and the service provider capability document to generate a comprehensive semantic matching score, including: The bidding requirements document is preprocessed and input into a pre-trained language model to obtain a requirement semantic vector; the service provider capability document is preprocessed and input into the pre-trained language model to generate a service provider capability semantic vector.
[0011] Calculate the mean cosine similarity between the demand semantic vector and the service provider capability semantic vector to obtain the preliminary semantic matching degree; obtain the description text of similar projects in the past of the service provider, input the description text of similar projects into the pre-trained language model to generate historical project semantic vectors, and calculate the similarity between the bidding demand semantic vector and the historical project semantic vectors to obtain the historical behavior corroboration score.
[0012] The preliminary semantic matching score is weighted and fused with the historical behavior corroboration score to generate a comprehensive semantic matching score.
[0013] Furthermore, the system integrates the real-time operational status of service providers and dynamically adjusts the comprehensive semantic matching score based on this status to generate a dynamic matching score, including: Obtain the current project load rate data of the service provider through the application programming interface.
[0014] Based on the preset mapping rules between load rate and order acceptance capacity, a capacity adaptation adjustment factor is generated based on the current project load rate data.
[0015] The comprehensive semantic matching score is corrected using the aforementioned capacity adaptability adjustment factor to generate a dynamic matching score.
[0016] Furthermore, the mapping rule between load rate and order acceptance capacity is as follows: establish a correspondence table between load rate ranges and priority adjustment ranges, and map the current project load rate to a specific capacity adaptability adjustment factor by querying this table.
[0017] Furthermore, the dynamic matching score undergoes compliance verification and business rules are applied to generate the final recommendation ranking, including: Obtain the service provider's qualification and compliance status, perform a logical AND operation between the dynamic matching score and the service provider's qualification and compliance status, filter out non-compliant service providers, and generate a list of compliant candidates.
[0018] The service providers in the compliance candidate list are sorted in descending order according to the dynamic matching score to generate a preliminary ranking sequence.
[0019] The business rules that prioritize supporting specific types of service providers are applied in the initial ranking sequence to generate the final recommended ranking.
[0020] Furthermore, based on the attention mechanism, an interpretability description is generated for the final recommendation ranking, forming a matching result report, including: For each service provider in the final recommended ranking, an attention mechanism is used to analyze the relevance between its capability document and the bidding requirements document, generating key matching criteria.
[0021] The key matching criteria are converted into natural language descriptions to form readable explanatory text.
[0022] The readable explanatory text is combined with the final recommendation ranking to generate a matching result report.
[0023] Furthermore, the attention mechanism analysis specifically involves extracting the attention weight distribution during the encoding process of the pre-trained language model and identifying the text segment with the highest attention weight as the key matching basis.
[0024] Furthermore, user feedback data on the matching result report is collected, including: Acquire user interactions with service providers in the matching results report as implicit feedback signals.
[0025] The implicit feedback signal is associated with and stored in conjunction with the semantic vector generated in this matching task.
[0026] Integrate all associated stored feedback data to construct a feedback dataset for model optimization.
[0027] Furthermore, the pre-trained language model is fine-tuned using the feedback dataset to complete the model optimization loop, including: Extract semantic vectors and corresponding implicit user feedback signals from the feedback dataset.
[0028] Supervised fine-tuning training samples are constructed using extracted semantic vectors and feedback signals.
[0029] The pre-trained language model is then fine-tuned using the training samples to optimize the model parameters.
[0030] Furthermore, the pre-trained language model is a BERT model.
[0031] Secondly, this application provides a deep learning-based intelligent matching system for bidding and procurement service capabilities, including: Document Acquisition Module: Acquires bidding requirement documents and service provider capability documents, and uses a pre-trained language model to perform deep semantic encoding on the bidding requirement documents and service provider capability documents to generate a comprehensive semantic matching score.
[0032] Semantic encoding scoring module: Integrates the real-time operational status of service providers, dynamically adjusts the comprehensive semantic matching score based on the real-time operational status of service providers, and generates a dynamic matching score.
[0033] Operational Status Adjustment Module: Performs compliance verification on the dynamic matching score and applies business rules to generate the final recommendation ranking.
[0034] Compliance verification ranking module: Based on the attention mechanism, it generates an interpretability description for the final recommended ranking and forms a matching result report.
[0035] Interpretable report generation module: collects user feedback data on the matching result report.
[0036] Feedback fine-tuning and optimization module: Uses the feedback dataset to fine-tune the pre-trained language model, completing the model optimization loop.
[0037] Thirdly, this application provides a deep learning-based intelligent matching device for bidding and procurement service capabilities, which includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program to implement the steps of the deep learning-based intelligent matching method for bidding and procurement service capabilities.
[0038] Fourthly, this application provides a storage medium storing computer program instructions, which are read and executed by a processor to perform the steps of a deep learning-based intelligent matching method for bidding and procurement service capabilities.
[0039] Fifthly, this application provides a computer program product, including a computer program or instructions, wherein when the computer program or instructions are executed by a processor, the steps of implementing a deep learning-based intelligent matching method for bidding and procurement service capabilities are provided.
[0040] The beneficial effects of this application are: This application generates a comprehensive score by using a pre-trained language model for deep semantic encoding, adjusts the score based on the service provider's real-time operational status, generates a ranking after compliance verification and business rules, and uses an attention mechanism to generate descriptions and provide feedback for fine-tuning the model. This completes the matching and optimization process, effectively solving the problem in existing technologies where the accuracy and effectiveness of the recommended list are affected by the reliance on superficial keyword matching and the difference between it and the inherent semantic richness of natural language. It can accurately capture the deep semantic fit between bidding requirements and service provider capabilities, correct deviations that may occur during the matching process, clearly distinguish the subtle differences in service providers' capabilities, improve the accuracy and effectiveness of the recommended list, and continuously improve the overall matching performance through the optimization loop formed by user feedback.
[0041] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description and the accompanying drawings. Attached Figure Description
[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0043] Figure 1 A flowchart illustrating the deep learning-based intelligent matching method for bidding and procurement service capabilities of this application is shown. Detailed Implementation
[0044] To address the issues raised in the background technology, this application employs a pre-trained language model with deep semantic encoding to generate a comprehensive score, adjusts the score based on the service provider's real-time operational status, generates a ranking after compliance verification and business rules, and uses an attention mechanism to generate descriptions and provide feedback for fine-tuning the model, thus constructing a complete matching and optimization process and improving the accuracy and effectiveness of the recommended list.
[0045] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0046] In some embodiments, such as Figure 1 As shown, this application provides a deep learning-based intelligent matching method for bidding and procurement service capabilities, including: S1. Obtain the bidding requirements document and the service provider's capability document, and use a pre-trained language model to perform deep semantic encoding on the bidding requirements document and the service provider's capability document to generate a comprehensive semantic matching score.
[0047] S2. Integrate the real-time operational status of service providers and dynamically adjust the comprehensive semantic matching score based on the real-time operational status of service providers to generate a dynamic matching score.
[0048] S3. Perform compliance verification on the dynamic matching score and apply business rules to generate the final recommendation ranking.
[0049] S4. Based on the attention mechanism, generate an interpretability description for the final recommendation ranking and form a matching result report.
[0050] S5. Collect user feedback data on the matching result report.
[0051] S6. Use the feedback dataset to fine-tune the pre-trained language model, completing the model optimization loop.
[0052] In some embodiments, S1 utilizes a pre-trained language model to perform deep semantic encoding on the bidding requirements document and the service provider capability document, generating a comprehensive semantic matching score, including: S11. Preprocess the bidding requirements document and input it into the pre-trained language model to obtain the requirement semantic vector; preprocess the service provider capability document and input it into the pre-trained language model to generate the service provider capability semantic vector; wherein, the pre-trained language model is the BERT model.
[0053] The tender requirements document is a formal document issued by the tendering party to clarify the procurement requirements of the project, including project requirements, qualification requirements, technical standards, etc.
[0054] First, the bidding requirements document is preprocessed, including text denoising, word segmentation, and standardization. Then, the obtained text is input into a pre-trained BERT model. The model captures the semantic context of the text through a multi-layer Transformer structure and outputs a requirement semantic vector. , Each vector in the document corresponds to a semantic fragment of the requirement document, reflecting the semantic features of that semantic fragment.
[0055] Service provider capability documentation refers to documents submitted or made public by service providers to showcase their service capabilities, including service scope, qualifications and certifications, technical strength, and contract fulfillment capabilities.
[0056] Similarly, the same preprocessing operations are performed on the service provider capability documents, and then the resulting text is input into the same BERT model to generate service provider capability semantic vectors. , Each vector in the document corresponds to a semantic fragment of the capability document, reflecting the semantic features of a certain capability of the service provider.
[0057] S12. Calculate the mean cosine similarity between the demand semantic vector and the service provider capability semantic vector to obtain the preliminary semantic matching degree; obtain the description text of similar projects in the past of the service provider, input the description text of similar projects into the pre-trained language model to generate historical project semantic vectors, and calculate the similarity between the bidding demand semantic vector and the historical project semantic vectors to obtain the historical behavior corroboration score.
[0058] calculate and The mean cosine similarity is used to obtain the preliminary semantic matching degree. .
[0059] Obtain descriptions of similar past projects from the service provider. These descriptions represent projects previously undertaken by the service provider that are similar to the current bidding project in terms of business scope, service content, and technical requirements. They include information such as project background, implementation content, and deliverables.
[0060] After preprocessing the description text of similar past projects from service providers, the data is input into the BERT model to generate semantic vectors for historical projects. ,calculate and The cosine similarity is used to obtain a score for historical behavior evidence. .
[0061] S13. The preliminary semantic matching score is weighted and fused with the historical behavior evidence score to generate a comprehensive semantic matching score.
[0062] Calculate the comprehensive semantic matching score , Among them, the analytic hierarchy process (AHP) can be used to determine and weight , Specifically, the target layer determines the weights of S1 and S2. , The criteria layer consists of project type complexity, requirement clarity, and relevance to historical experience; the solution layer consists of S1 and S2. Using a 1-9 scale, the importance of each indicator in the criteria layer to the target layer and the importance of each score in the solution layer to each indicator in the criteria layer are compared to establish a judgment matrix. The largest eigenvalue and corresponding eigenvector of the matrix are calculated to obtain the initial weights. Finally, a consistency check is performed. When the consistency ratio CR... The test passed at 0.1, and the result was obtained. and .
[0063] In some embodiments, S2 integrates the real-time operational status of the service provider and dynamically adjusts the comprehensive semantic matching score based on the real-time operational status of the service provider to generate a dynamic matching score, including: S21. Obtain the current project load rate data of the service provider through the application programming interface.
[0064] The current project load factor represents the ratio of the number of projects currently undertaken by the service provider to its maximum capacity.
[0065] A connection is established with the service provider's project management system through a pre-defined application programming interface (API), and a data retrieval request is sent to the project management system. The API returns the service provider's current project load rate data according to the request. .
[0066] S22. Based on the preset load rate and order acceptance capacity mapping rules, generate a capacity adaptation adjustment factor based on the current project load rate data; wherein, the load rate and order acceptance capacity mapping rules are specifically as follows: establish a correspondence table between load rate ranges and priority adjustment ranges, and map the current project load rate to a specific capacity adaptation adjustment factor by querying the table.
[0067] The pre-defined load rate and order acceptance capacity mapping rules establish a correspondence between load rate ranges and priority adjustment increments. For example, a load rate range of [0, 30%] corresponds to a priority adjustment increment of +10%, a load rate range of [30%, 70%] corresponds to a priority adjustment increment of 0%, and a load rate range of [70%, 100%] corresponds to a priority adjustment increment of [10%]. The load rate range and priority adjustment range can be adjusted according to the business scenario.
[0068] Querying the corresponding relationship based on Determine the corresponding priority adjustment range within the given load factor range, and then convert this priority adjustment range into a capacity adaptability adjustment factor. For example, a priority adjustment of +10% corresponds to .
[0069] S23. Use the capacity adaptability adjustment factor to correct the comprehensive semantic matching score and generate a dynamic matching score.
[0070] use right Make corrections to obtain a dynamic matching score. Reference formula: ; It can reflect the semantic matching degree between bidding requirements and service provider capabilities, and also take into account the service provider's ability to accept orders.
[0071] In some embodiments, S3 performs compliance checks on the dynamic matching scores and applies business rules to generate the final recommendation ranking, including: S31. Obtain the service provider's qualification and compliance status, perform a logical AND operation between the dynamic matching score and the service provider's qualification and compliance status, filter out non-compliant service providers, and generate a list of compliant candidates.
[0072] Obtain the service provider's qualification compliance status C. The service provider's qualification compliance status C is determined based on industry qualification standards, policy requirements, etc. C=1 means that the service provider's qualification is compliant, and C=0 means that the service provider's qualification is not compliant.
[0073] Perform a logical AND operation on each service provider: when C=1, retain that service provider and its corresponding... When C=0, the service provider is filtered out.
[0074] All qualified and compliant service providers were screened out, and their information was integrated to generate a list of compliant candidate providers. .
[0075] S32. Sort the service providers in the compliance candidate list in descending order according to the dynamic matching score to generate a preliminary ranking sequence.
[0076] right All service providers are categorized according to their respective service providers. Sort the data from largest to smallest to obtain a preliminary ranking sequence. Among them, the same The service providers can be sorted in a second sort by the first letter of their names or by their qualification level. The second sorting rules can be set according to business needs.
[0077] S33. Apply the business rules that prioritize supporting specific types of service providers in the initial ranking sequence to generate the final recommended ranking.
[0078] Establish business rules that prioritize support for specific types of service providers, such as micro and small enterprises, local enterprises, and high-tech enterprises. These business rules are set based on policy guidance and regional development needs.
[0079] The criteria for determining the beneficiaries of support should be clearly defined. For example, micro and small enterprises should provide proof of their micro and small enterprise qualifications, and high-tech enterprises should have high-tech enterprise certification.
[0080] Traversal The system identifies service providers that meet the support criteria and adjusts their rankings accordingly. For example, it might elevate eligible service providers to a fixed position based on their existing ranking, or set a minimum ranking threshold for them. After the adjustments are complete, a final recommended ranking is generated. .
[0081] In some embodiments, the attention-based mechanism in S4 generates an interpretable description for the final recommendation ranking, forming a matching result report, including: S41. For each service provider in the final recommendation ranking, use an attention mechanism to analyze the relevance between its capability document and the bidding requirement document, and generate key matching criteria. Specifically, the attention mechanism analysis involves: extracting the attention weight distribution during the encoding process of the pre-trained language model, and identifying the text segment with the highest attention weight as the key matching criteria.
[0082] right For each service provider, the attention weight distribution W of the service provider's capability document and bidding requirements document during the BERT model encoding process is retrieved. The text segment with the highest attention weight is selected and used as the key matching point P between the service provider and the bidding requirements. P contains the key text segment on the requirement side and the corresponding matching text segment on the capability side. For example, the text segment "has ISO9001 certification" in the service provider's capability document and the text segment "has obtained ISO9001 certification" in the bidding requirements document.
[0083] S42. Transform key matching criteria into natural language descriptions to form readable explanatory text.
[0084] Using the framework of "key demand points + capability matching points + matching conclusions", the key text fragments on the demand side in P are taken as key demand points, and the corresponding matching text fragments on the capability side are taken as capability matching points. These are then substituted into the framework, organized into natural language, and supplemented with logical connectors and degree descriptive words to form a readable explanatory text T for the service provider.
[0085] S43. Combine the readable explanatory text with the final recommendation ranking to generate a matching results report.
[0086] The matching results report includes the report title, generation time, and final recommendation ranking. T is a readable explanatory text for each service provider.
[0087] Final Recommendation Ranking Embed the report according to the ranking order, and embed the corresponding T for each service provider in the final recommended ranking. Under the ranking entries, add the report title and generation time to obtain the matching result report R.
[0088] In some embodiments, the user feedback data collected in S5 regarding the matching result report includes: S51. Obtain user interactions with service providers in the matching results report as implicit feedback signals.
[0089] User interaction types include user actions to... The user can view, favorite, compare, or ignore service providers. Different actions reflect different attitudes towards the matching results. For example, clicking to view indicates that the user is interested in the service provider, while ignoring it indicates that the user is not interested in the service provider.
[0090] The operation log recording function captures user interactions in real time while browsing the matching results report (R), and marks each interaction, such as clicking to view being marked as... The collection is marked as Ignore the marker as This forms implicit feedback signals F from users to various service providers.
[0091] S52. Associate and store the implicit feedback signal with the semantic vector generated in this matching task.
[0092] This matching task is assigned a unique task ID, each service provider within this task is assigned a unique service provider ID, and the bidding requirements for this task are assigned a unique requirement ID.
[0093] F represents the final recommendation ranking when the user views the current matching results report R. The interaction operation tags generated by each service provider in the process mean that each F corresponds to a specific service provider within this matching task. At this point, using "task ID + service provider ID" as the association key, the F of the same service provider is linked to its corresponding service provider capability semantic vector. Semantic vectors of historical projects The association uses "task ID + requirement ID" as the association key to link the semantic vector of the requirements for this task. It is linked with the associated data of all service providers to form a "task ID-requirement ID-" structure. -Service Provider ID- - The "-F" complete association chain stores the associated data in a dedicated database.
[0094] S53. Integrate all associated stored feedback data to construct a feedback dataset for model optimization.
[0095] The construction cycle of the feedback dataset is determined. At each construction cycle node, all associated and stored feedback data is extracted from a dedicated database. The extracted feedback data is cleaned and integrated into a feedback dataset according to a preset format. .
[0096] In some embodiments, S6, fine-tuning the pre-trained language model using the feedback dataset to complete the model optimization loop, includes: S61. Extract semantic vectors and corresponding implicit user feedback signals from the feedback dataset.
[0097] Define the data extraction requirements, such as extracting the feedback dataset. Valid data within a recent construction cycle, or records with clearly defined operation markers extracted from implicit feedback signals F.
[0098] from Select the semantic vectors of the data that meet the extraction requirements. Service Provider Capability Semantic Vector And implicit feedback signal F, and perform format conversion on the extracted data, such as converting semantic vectors from matrix form to tensor form that the model can recognize, and converting operation tags in F into numerical labels, for example... Corresponding to numerical label 1, Corresponding to numerical labels such as 0, the converted semantic vector data and feedback signal labels are obtained; among them, It can be used as supplementary data, and its inclusion can be selected based on the need for fine-tuning.
[0099] S62. Construct supervised fine-tuned training samples using the extracted semantic vectors and feedback signals.
[0100] Each training sample contains input features and labels, where the input features are... and The combination of semantic vectors and feedback signal labels reflects the matching results between bidding requirements and service provider capabilities. The labels are the corresponding numerical labels after F transformation, reflecting the user's preference for the matching results between bidding requirements and service provider capabilities. The semantic vectors are paired with the feedback signal labels to form a single training sample, and finally divided into training sets according to a preset ratio. Validation set and test set .
[0101] S63. Use training samples to perform supervised fine-tuning of the pre-trained language model and optimize the model parameters.
[0102] Load the pre-trained BERT model used by S11, freeze some parameters in the lower layers, and only unfreeze the parameters of the top layers and semantic matching-related layers. As input, the error between the predicted value and the sample label is calculated using the cross-entropy loss function. The trainable parameters are iteratively updated using the Adam optimizer. After each training round, the error is calculated using... Validate the model performance; if the validation error increases, stop training to avoid overfitting. Finally, use... Evaluate the performance of the fine-tuned model, output the optimized model parameters, and complete the pre-trained model optimization.
[0103] In some embodiments, this application provides a deep learning-based intelligent matching system for bidding and procurement service capabilities, comprising: Document Acquisition Module: Acquires bidding requirement documents and service provider capability documents, and uses a pre-trained language model to perform deep semantic encoding on the bidding requirement documents and service provider capability documents to generate a comprehensive semantic matching score.
[0104] Semantic encoding scoring module: Integrates the real-time operational status of service providers, dynamically adjusts the comprehensive semantic matching score based on the real-time operational status of service providers, and generates a dynamic matching score.
[0105] Operational Status Adjustment Module: Performs compliance verification on dynamic matching scores and applies business rules to generate the final recommendation ranking.
[0106] Compliance verification ranking module: Based on the attention mechanism, it generates an interpretable description for the final recommendation ranking and forms a matching result report.
[0107] Interpretable report generation module: Collects user feedback data on matching result reports.
[0108] Feedback fine-tuning and optimization module: Uses the feedback dataset to fine-tune the pre-trained language model, completing the model optimization loop.
[0109] In some embodiments, this application provides a deep learning-based intelligent matching device for bidding and procurement service capabilities, which includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program to implement the steps of the deep learning-based intelligent matching method for bidding and procurement service capabilities.
[0110] In some embodiments, this application provides a storage medium storing computer program instructions, which are read and executed by a processor to perform the steps of a deep learning-based intelligent matching method for bidding and procurement service capabilities.
[0111] Fifthly, this application provides a computer program product, including a computer program or instructions, wherein when the computer program or instructions are executed by a processor, the steps of implementing a deep learning-based intelligent matching method for bidding and procurement service capabilities are provided.
[0112] It should be noted that, in this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0113] In this application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.
[0114] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0115] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.
[0116] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A deep learning-based intelligent matching method for bidding and procurement service capabilities, characterized in that, include: Obtain bidding requirement documents and service provider capability documents, and use a pre-trained language model to perform deep semantic encoding on the bidding requirement documents and service provider capability documents to generate a comprehensive semantic matching score. Integrate the real-time operational status of service providers, and dynamically adjust the comprehensive semantic matching score based on the real-time operational status of the service providers to generate a dynamic matching score; The dynamic matching score is validated for compliance and business rules are applied to generate the final recommendation ranking; An interpretable description is generated for the final recommendation ranking based on an attention mechanism, forming a matching result report; Collect user feedback data on the matching result report; The pre-trained language model is fine-tuned using the feedback dataset to complete the model optimization loop.
2. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 1, characterized in that, Using a pre-trained language model, deep semantic encoding is performed on the bidding requirements document and the service provider capability document to generate a comprehensive semantic matching score, including: The bidding requirements document is preprocessed and input into the pre-trained language model to obtain the requirement semantic vector; the service provider capability document is preprocessed and input into the pre-trained language model to generate the service provider capability semantic vector. Calculate the mean cosine similarity between the demand semantic vector and the service provider capability semantic vector to obtain the preliminary semantic matching degree; obtain the description text of similar projects in the past of the service provider, input the description text of similar projects into the pre-trained language model to generate historical project semantic vectors, and calculate the similarity between the bidding demand semantic vector and the historical project semantic vectors to obtain the historical behavior corroboration score. The preliminary semantic matching score is weighted and fused with the historical behavior corroboration score to generate a comprehensive semantic matching score.
3. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 1, characterized in that, Integrate the real-time operational status of service providers, and dynamically adjust the comprehensive semantic matching score based on the real-time operational status of the service providers to generate a dynamic matching score, including: Obtain the current project load rate data of the service provider through the application programming interface; Based on the preset mapping rules between load rate and order acceptance capacity, a capacity adaptation adjustment factor is generated based on the current project load rate data; The comprehensive semantic matching score is corrected using the aforementioned capacity adaptability adjustment factor to generate a dynamic matching score.
4. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 3, characterized in that, The specific rules for mapping load rate to order acceptance capacity are as follows: establish a correspondence table between load rate ranges and priority adjustment ranges, and map the current project load rate to a specific capacity adaptability adjustment factor by querying this table.
5. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 1, characterized in that, The dynamic matching score is subjected to compliance verification and business rules are applied to generate the final recommendation ranking, including: Obtain the service provider's qualification and compliance status, perform a logical AND operation between the dynamic matching score and the service provider's qualification and compliance status, filter out non-compliant service providers, and generate a compliant candidate list; The service providers in the compliance candidate list are sorted in descending order according to the dynamic matching score to generate a preliminary ranking sequence; The business rules that prioritize supporting specific types of service providers are applied in the initial ranking sequence to generate the final recommended ranking.
6. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 1, characterized in that, An interpretable description is generated for the final recommendation ranking based on an attention mechanism, forming a matching result report, including: For each service provider in the final recommended ranking, an attention mechanism is used to analyze the relevance between its capability document and the bidding requirements document, generating key matching criteria. The key matching criteria are transformed into natural language descriptions to form readable explanatory text; The readable explanatory text is combined with the final recommendation ranking to generate a matching result report.
7. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 6, characterized in that, The specific analysis using the attention mechanism involves extracting the distribution of attention weights during the encoding process of the pre-trained language model and identifying the text segment with the highest attention weight as the key matching basis.
8. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 1, characterized in that, Collect user feedback data on the matching result report, including: Acquire user interactions with service providers in the matching results report as implicit feedback signals; The implicit feedback signal is associated with and stored in conjunction with the semantic vector generated in this matching task; Integrate all associated stored feedback data to construct a feedback dataset for model optimization.
9. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 1, characterized in that, The pre-trained language model is fine-tuned using the feedback dataset to complete the model optimization loop, including: Extract semantic vectors and corresponding implicit user feedback signals from the feedback dataset; Supervised fine-tuning training samples are constructed using extracted semantic vectors and feedback signals; The pre-trained language model is then fine-tuned using the training samples to optimize the model parameters.
10. The intelligent matching method for bidding and procurement service capabilities based on deep learning according to claim 1, characterized in that, The pre-trained language model is the BERT model.