Science and technology project intelligent matching and pre-scoring system and method based on large model

Abstract

The invention discloses an intelligent science and technology project matching and pre-scoring system based on a large model, and relates to the technical field of project matching evaluation, and the system comprises a data collection module which collects multi-source science and technology project information; the data preprocessing module is used for cleaning, converting and standardizing the data; the large model construction module comprises a feature coding layer, a matching decision layer and a pre-scoring prediction layer; the intelligent matching module outputs a matching degree and similar items; a pre-scoring module; and a user interaction module, a model updating module, a security module and a privacy protection module. The large model fuses multiple technical training, each module has a specific function and algorithm, accurate and efficient operation of the system is guaranteed, and data security privacy is protected. According to the method, a large number of science and technology projects can be efficiently processed, the evaluation period is shortened, the matching accuracy and the pre-scoring reliability are improved, multiple factors are considered, the subjective influence is reduced, the system can be dynamically updated, the knowledge graph is updated, the management monitoring function is perfected, intelligent development of science and technology project management is promoted, and resource configuration is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of project matching and evaluation systems, and in particular to a system and method for intelligent matching and pre-scoring of scientific and technological projects based on a large model. Background Art

[0002] In today's era of rapid technological development, the number of scientific and technological projects has exploded, and how to efficiently manage, evaluate and allocate resources for these projects has become a key issue. Traditional scientific and technological project matching and scoring methods mainly rely on manual operations and have many limitations.

[0003] When manually evaluating scientific and technological projects, faced with a massive amount of project information, the processing efficiency is low. Evaluators need to spend a lot of time reading and analyzing project applications, research reports and other documents, and it is difficult to handle a large number of projects in a short period of time, resulting in a long project evaluation cycle and unable to meet the rapid needs of scientific research and enterprise development in a timely manner. Moreover, manual evaluation is highly subjective, and the knowledge background, experience and personal preferences of different evaluators will affect the evaluation results. The lack of unified and objective standards makes it difficult to guarantee the accuracy and reliability of the evaluation results.

[0004] In addition, traditional methods are unable to cope with complex project relationships and multi-dimensional evaluation factors. Science and technology projects involve multiple fields, multiple technologies, different team capabilities, market demands and other factors. It is difficult for humans to comprehensively and accurately consider the interrelationships between these factors, and it is easy to overlook important information, thus affecting the reasonable matching and accurate scoring of projects.

[0005] The continuous development of artificial intelligence technology, especially the emergence of big model technology, provides new opportunities for solving these problems. Big models can process large amounts of data, automatically learn complex patterns and relationships in data, and have strong natural language processing and analysis capabilities. Therefore, applying big models to the intelligent matching and pre-scoring of scientific and technological projects is expected to overcome the shortcomings of traditional methods, achieve efficient, objective, and accurate project management, improve the efficiency of the allocation of scientific and technological resources, and promote scientific and technological innovation and industrial development. Summary of the invention

[0006] The present invention proposes a large-model-based intelligent matching and pre-scoring system and method for scientific and technological projects to solve the problems mentioned in the above-mentioned prior art.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: a scientific and technological project intelligent matching and pre-scoring system based on a large model, comprising: The data collection module is used to collect multi-source information of scientific and technological projects, including project applications, research reports, patent documents, scientific research personnel resumes, etc. The data sources include but are not limited to government scientific research management platforms, university scientific research databases, corporate R&D archives, and professional academic forums. The collection frequency is dynamically adjusted according to the key nodes of project progress and the dynamic change characteristics of data to ensure the timeliness and integrity of the data; The data preprocessing module performs deep cleaning, structural conversion and semantic alignment on the collected data, extracts key information using entity recognition and relationship extraction technology based on deep learning, corrects data deviations by combining statistical analysis and rule engines, and maps data to a unified semantic framework based on domain ontology, laying a solid foundation for subsequent analysis; The large model construction module builds a large model architecture that integrates knowledge distillation and reinforcement learning, including a project feature encoding layer, a matching decision layer, and a pre-scoring prediction layer. The feature encoding layer uses Transformer-Encoder combined with a graph neural network (GNN) to model the multivariate relationship of projects. The matching decision layer optimizes the matching strategy based on a policy gradient reinforcement learning algorithm. The pre-scoring prediction layer uses ensemble learning to fuse the outputs of multiple base learners. Model training uses large-scale labeled data and simulated data expanded by a generative adversarial network (GAN) to improve model generalization and robustness. The intelligent matching module inputs the preprocessed project data into the large model, generates the project feature vector through the feature coding layer, searches for similar projects based on the dynamic matching strategy of the matching decision layer, and outputs the matching degree and similar project set. The matching degree calculation introduces the calculation formula as follows: ,in For the The weight of the similarity feature, is the corresponding similarity eigenvalue, is the mapping function from feature value to matching contribution value, and n is the number of similarity features; The pre-scoring module pre-scores the project to be evaluated based on the pre-scoring prediction layer of the large model, matches the historical scoring of the project, and the experience rules of domain experts. It outputs the score and detailed explanation. The pre-scoring calculation takes into account the project's innovation degree. , technology maturity T, market demand fit M, resource sustainability R, the formula is: , , , , Where is the weight of each factor, is a correction term based on the distribution of matching item scores; The result evaluation and optimization module collects the actual execution result data of the project, compares the pre-score with the actual score, uses the deviation analysis algorithm to locate the cause of the model prediction deviation, adjusts the model parameters and feature weights based on the feedback data, optimizes the matching and pre-score algorithms, and continuously improves the accuracy of the system; The knowledge graph update module monitors industry trends and scientific research progress in real time, updates knowledge graph nodes and relationships through information extraction and fusion technology, integrates new knowledge into the large model training process, and enhances the model's ability to understand emerging technologies and trends; The system management and monitoring module is responsible for system user authority management, operation log recording and system performance monitoring to ensure the safe and stable operation of the system. It provides a visual monitoring dashboard to display the system operation status and key indicators, so that administrators can find and handle problems in a timely manner.

[0008] Furthermore, when collecting the resumes of scientific researchers, the data collection module uses natural language processing technology to parse unstructured text, extract key information such as project experience, skills and expertise, and academic achievements, and ensures the accuracy of the information through semantic similarity matching and expert verification.

[0009] Furthermore, in the semantic alignment process of the data preprocessing module, the semantic similarity between words and sentences is calculated based on the pre-trained semantic vector model, and the dynamic programming algorithm is used for optimal matching to solve the problem of semantic heterogeneity of multi-source data.

[0010] Furthermore, the intelligent matching module calculates the similarity of the project domain When using a method based on the semantic vector space model and domain ontology tree, the formula is: ,in Indicates that the project to be evaluated P is the first The semantic vector of a node, Represents the semantic vector of the historical item at the corresponding node, Calculate the similarity function between two semantic vectors (such as cosine similarity), The weight value is determined according to the node depth and association relationship in the domain ontology tree; in the calculation of the overlap of technical keywords When using the algorithm based on keyword co-occurrence network and information entropy, the formula is: ,in Indicates keywords The probability of co-occurrence between the item to be evaluated and the matching item, is the total number of keywords.

[0011] Furthermore, in the pre-scoring module, the evaluation of project innovation is based on patent citation analysis and technology frontier identification technology, the technology maturity is judged by the technology life cycle model and expert experience rules, the market demand fit is determined in combination with the market research report and user demand forecasting model, resource sustainability considers the supply stability and scalability of resources such as funds, manpower and equipment, and the correction items in the pre-scoring calculation are calculated based on the average score deviation of the industry to which the project belongs and the project scale adjustment factor.

[0012] Furthermore, when collecting the actual execution result data of the project, the result evaluation and optimization module connects with the project management system and the financial system to obtain information such as project progress, cost, and quality of results, uses time series analysis and causal inference algorithms to determine the deviation relationship between the pre-scoring and the actual results, adjusts the weight parameters of each feature layer in the large model according to the deviation, and introduces genetic algorithms or particle swarm optimization algorithms to search for the optimal model configuration when optimizing the matching and pre-scoring algorithms.

[0013] Furthermore, when updating the knowledge graph nodes and relationships, the knowledge graph update module uses natural language processing technology to extract entities and relationships from the latest scientific research literature and industry news, and uses entity alignment and relationship fusion algorithms to integrate new information into the existing knowledge graph. In the process of incorporating new knowledge into the large model training, incremental learning technology and knowledge distillation technology are used to enable the model to quickly adapt to new knowledge and reduce the forgetting of original knowledge.

[0014] Furthermore, in the system management and monitoring module, user authority management is based on a role-based access control (RBAC) model, operation log recording uses a distributed log system to ensure data integrity and traceability, system performance monitoring indicators include data collection delay, model inference time, server resource utilization, and a visual monitoring dashboard provides real-time data display and early warning functions. When performance indicators exceed preset thresholds, optimization strategies are automatically triggered or administrators are notified.

[0015] Compared with the prior art, the present invention has the following beneficial effects: In terms of matching efficiency, the present invention can complete the matching analysis of a large number of scientific and technological projects in a short period of time through the rapid processing of multi-source data by a large model. Compared with the traditional manual method, it greatly shortens the project evaluation cycle, enables scientific research resources to be invested in suitable projects more quickly, and improves the overall speed of scientific research work.

[0016] In terms of matching accuracy, the big model comprehensively considers multi-dimensional factors such as project fields, technical keywords, expected results, team information, etc., and calculates the matching degree through precise algorithms, avoiding the subjectivity of manual evaluation. It can find similar project cases more accurately and provide a more reliable reference basis for project evaluation.

[0017] The pre-scoring function provides quantitative evaluation results for projects in advance, combining considerations of multiple factors such as project innovation, technology maturity, market demand fit, and resource sustainability, so that project teams and investors can understand the project potential in advance, which helps to rationally plan resources, optimize project plans, and reduce the risk of project failure.

[0018] The system's dynamic update mechanism ensures that the model can adapt to industry development and the generation of new knowledge in a timely manner, always maintains sensitivity to the latest trends and technologies, and provides evaluation services that keep pace with the times. At the same time, the perfect user interaction and system management functions enhance the user experience and ensure the safe and stable operation of the system. Overall, it brings efficient, accurate, and intelligent solutions to scientific and technological project management, and effectively promotes the progress and development of the scientific and technological field. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic block diagram of a large-model-based intelligent matching and pre-scoring system for scientific and technological projects proposed by the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0022] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0023] Reference Figure 1 :A scientific and technological project intelligent matching and pre-scoring system based on a large model, including: The data collection module is used to collect multi-source information of scientific and technological projects, including project applications, research reports, patent documents, scientific research personnel resumes, etc. The data sources include but are not limited to government scientific research management platforms, university scientific research databases, corporate R&D archives, and professional academic forums. The collection frequency is dynamically adjusted according to the key nodes of project progress and the dynamic change characteristics of data to ensure the timeliness and integrity of the data; The data preprocessing module performs deep cleaning, structural conversion and semantic alignment on the collected data, extracts key information using entity recognition and relationship extraction technology based on deep learning, corrects data deviations by combining statistical analysis and rule engines, and maps data to a unified semantic framework based on domain ontology, laying a solid foundation for subsequent analysis; The large model construction module builds a large model architecture that integrates knowledge distillation and reinforcement learning, including a project feature encoding layer, a matching decision layer, and a pre-scoring prediction layer. The feature encoding layer uses Transformer-Encoder combined with a graph neural network (GNN) to model the multivariate relationship of projects. The matching decision layer optimizes the matching strategy based on a policy gradient reinforcement learning algorithm. The pre-scoring prediction layer uses ensemble learning to fuse the outputs of multiple base learners. Model training uses large-scale labeled data and simulated data expanded by a generative adversarial network (GAN) to improve model generalization and robustness. The intelligent matching module inputs the preprocessed project data into the large model, generates the project feature vector through the feature coding layer, searches for similar projects based on the dynamic matching strategy of the matching decision layer, and outputs the matching degree and similar project set. The matching degree calculation introduces the calculation formula as follows: ,in For the The weight of the similarity feature, is the corresponding similarity eigenvalue, is the mapping function from feature value to matching contribution value, and n is the number of similarity features; The pre-scoring module pre-scores the project to be evaluated based on the pre-scoring prediction layer of the large model, matches the historical scoring of the project, and the experience rules of domain experts. It outputs the score and detailed explanation. The pre-scoring calculation takes into account the project's innovation degree. , technology maturity T, market demand fit M, resource sustainability R, the formula is: , , , , Where is the weight of each factor, is a correction term based on the distribution of matching item scores; The result evaluation and optimization module collects the actual execution result data of the project, compares the pre-score with the actual score, uses the deviation analysis algorithm to locate the cause of the model prediction deviation, adjusts the model parameters and feature weights based on the feedback data, optimizes the matching and pre-score algorithms, and continuously improves the accuracy of the system; The knowledge graph update module monitors industry trends and scientific research progress in real time, updates knowledge graph nodes and relationships through information extraction and fusion technology, integrates new knowledge into the large model training process, and enhances the model's ability to understand emerging technologies and trends; The system management and monitoring module is responsible for system user authority management, operation log recording and system performance monitoring to ensure the safe and stable operation of the system. It provides a visual monitoring dashboard to display the system operation status and key indicators, so that administrators can find and handle problems in a timely manner.

[0024] In the data collection module of the present invention, a series of advanced technical means are used in the collection process of the important information source of the scientific researcher's resume to ensure that the collected information is accurate, comprehensive and highly usable.

[0025] When processing typical unstructured text data such as scientific researchers' resumes, cutting-edge natural language processing technology is used. First, the resume text is carefully preprocessed, including text cleaning to remove noise information (such as formatting tags, irrelevant symbols, etc.), and text segmentation to divide continuous text into word or phrase sequences according to semantic units. Then, deep learning models (such as named entity recognition models based on recurrent neural networks or Transformer architectures) are used for entity recognition to accurately locate key entities related to project experience, skills, expertise, academic achievements, etc. For example, specific project names, technical skill keywords, published paper titles and publications, and obtained patent names and patent numbers can be identified through training models.

[0026] After extracting key information, a method combining semantic similarity matching and expert verification is used to ensure the accuracy of the information. Semantic similarity matching is based on pre-trained semantic vector models (such as Word2Vec, BERT, etc.), which converts the extracted project experience, skills and expertise into semantic vector representations, and then calculates the similarity between different vectors. For example, the similarity of the project experience claimed by the researchers is compared with the known standard project descriptions to determine the rationality and authenticity of their descriptions. For academic achievements, similarity matching is performed with literature information in authoritative academic databases (such as Web of Science, CNKI, etc.) to verify the existence and relevance of the achievements.

[0027] Senior experts and scholars in the field were invited to participate in the expert verification phase. Experts manually reviewed the collected information based on their rich experience and professional knowledge. They carefully evaluated the actual value of project experience, the actual level of skills and expertise, and the influence of academic achievements, and further verified and confirmed any questionable or unclear information. This dual guarantee mechanism of combining semantic similarity matching with expert verification effectively improves the accuracy of the collected scientific research personnel's resume information, providing a solid and reliable data foundation for subsequent project intelligent matching and pre-scoring.

[0028] In the data preprocessing module of the present invention, semantic alignment processing is a key operation, which plays an important role in solving the semantic heterogeneity problem of multi-source data.

[0029] In the semantic alignment process, we make full use of the pre-trained semantic vector model. This model is trained on large-scale text data and can map words and sentences into a high-dimensional semantic space, so that each word and sentence has a corresponding semantic vector representation. In this way, calculating the semantic similarity between words and sentences becomes more accurate and efficient. For example, for similar concepts from different data sources, although their expressions may be different, in the semantic vector space, the vector distance between words and sentences with similar meanings is closer.

[0030] In order to further achieve accurate semantic alignment, a dynamic programming algorithm is used for optimal matching. The dynamic programming algorithm comprehensively considers multiple factors such as the semantic similarity of words and sentences and their positional relationship in the text. It constructs the optimal substructure and recursively solves the overall optimal matching solution, which can effectively handle complex semantic relationships and avoid local optimal solutions.

[0031] Through this approach based on the combination of pre-trained semantic vector models and dynamic programming algorithms, the system can accurately identify semantically identical or similar content in different data sources, thereby successfully solving the problem of semantic heterogeneity in multi-source data and providing a high-quality, semantically unified data foundation for subsequent data fusion and analysis, ensuring the accuracy and reliability of data processing by the entire system.

[0032] In the present invention, the intelligent matching module calculates the similarity of the project domain When using a method based on the semantic vector space model and domain ontology tree, the formula is: ,in Indicates that the project to be evaluated P is the first The semantic vector of a node, Represents the semantic vector of the historical item at the corresponding node, Calculate the similarity function between two semantic vectors (such as cosine similarity), The weight value is determined according to the node depth and association relationship in the domain ontology tree; in the calculation of the overlap of technical keywords When using the algorithm based on keyword co-occurrence network and information entropy, the formula is: ,in Indicates keywords The probability of co-occurrence between the item to be evaluated and the matching item, is the total number of keywords.

[0033] In the pre-scoring module of the present invention, the evaluation of project innovation adopts a combination of advanced patent citation analysis and cutting-edge technology identification technology. In terms of patent citation analysis, a large-scale patent citation relationship network database is constructed. By deeply mining the citation relationship between the patents involved in the project to be evaluated and other patents, the complex network analysis algorithm is used to calculate the patent's influence indicators, such as the number of citations, citation breadth, citation depth, etc., and combined with the technical innovation weight of the cited patent, the project's innovation degree is comprehensively evaluated. At the same time, a deep learning algorithm is used to analyze academic literature and industry reports in the field of cutting-edge technology, identify emerging technology trends and key technology breakthroughs, establish a technology frontier knowledge base, and further determine the project's innovation level by matching and comparing the project technology with the cutting-edge technology in the knowledge base.

[0034] The judgment of technology maturity is based on a complete technology life cycle model and rich expert experience rules. The technology life cycle model covers the stages of technology introduction, growth, maturity and decline. By analyzing multi-dimensional indicators such as the application time of the technology used in the project in the market, the speed of technology diffusion, and the frequency of technology improvement, the life cycle stage is determined, and then the technology maturity is evaluated. Expert experience rules are constructed by gathering the experience and knowledge of senior experts in the industry on the maturity judgment of different technology fields, and an expert rule base is constructed. For example, for certain specific technologies, detailed maturity judgment rules are formulated based on factors such as the degree to which their key performance indicators meet industry standards, technical stability, and scalability, which complement and verify the evaluation results of the technology life cycle model to ensure the accuracy of the technology maturity evaluation.

[0035] The determination of market demand fit integrates market research reports and advanced user demand forecasting models. Market research reports collect market data from multiple channels, including market size, market growth rate, market segmentation structure, competitor analysis and other information, and use statistical analysis and market trend forecasting algorithms to understand the overall market demand situation and potential opportunities. The user demand forecasting model uses user behavior modeling technology based on big data analysis. By collecting and analyzing user search behavior, purchase behavior, feedback and other data in related fields, it explores users' potential needs and preference change trends and builds a user demand portrait. Combine the overall market demand with individual user needs to accurately evaluate the degree of fit between the project and market demand.

[0036] Resource sustainability comprehensively considers the supply stability and scalability of resources such as funds, manpower, and equipment. In terms of funds, analyze the rationality of project budget planning, the stability of funding sources, and the efficiency of fund use, predict the funding needs and supply of the project at different stages, and assess the risk of a broken funding chain. In terms of manpower, evaluate the professional skills distribution, staff turnover, and training and development plans of project team members to ensure that the team has the ability to continuously advance the project. In terms of equipment, examine the advancement of the equipment, maintenance costs, update cycles, and the matching degree between the equipment and the technical requirements of the project. At the same time, consider the scalability of resources, that is, the difficulty of obtaining additional resources during the development of the project, such as whether it can attract new investment, recruit suitable talents, upgrade or expand equipment, etc. By comprehensively weighing these factors, accurately judge the sustainability of project resources.

[0037] The correction item in the pre-score calculation is based on the average score deviation of the industry to which the project belongs and the project scale adjustment factor. First, collect a large number of historical project score data, conduct statistical analysis according to industry classification, calculate the average score of each industry and the standard deviation of the score, and determine the average score deviation range of the industry to which the project belongs. The project scale adjustment factor takes into account factors such as the project's capital investment scale, team size, and expected results and influence, and is calculated by establishing a multi-factor weighted model. For example, for large projects with large capital investment, large team size, and expected results with wide influence, the scale adjustment factor is larger, and the correction effect on the pre-score is more significant, so that the pre-score results are more in line with the actual situation of the project.

[0038] The result evaluation and optimization module in the present invention demonstrates a high degree of systematicity and efficiency in the data collection process. It establishes a deep docking mechanism with the project management system and the financial system to ensure the acquisition of comprehensive and accurate project actual execution result data.

[0039] In the process of interacting with the project management system, detailed information on the project progress is obtained through carefully designed data interfaces and data collection protocols. This includes not only the conventional task completion percentage, the actual time of key milestones compared with the planned time, but also covers multi-dimensional progress data such as the actual time distribution of project activities, dynamic tracking of resource allocation and actual use, and the connection efficiency between project stages. With the help of advanced work breakdown structure (WBS) visualization tools, project tasks are refined layer by layer, and the floating time and total time difference of activities on the critical path are calculated in combination with the critical path method (CPM), so as to accurately identify key tasks and potential delay risk points that may affect the overall progress of the project.

[0040] At the same time, the connection with the financial system enables comprehensive control of project cost data. The actual expenditure details of the budget allocated to each cost item (such as labor cost, equipment procurement cost, raw material cost, R&D cost, etc.) and the cost accumulation at different stages are extracted in detail. The cost-benefit analysis method is used to deeply evaluate the cost control effect, including the specific reasons for cost overruns or savings, such as resource waste, rationality of budget adjustments, and the relationship between cost changes and project scope changes.

[0041] A strict and multi-dimensional evaluation index system has been established for the collection of project results quality information. From the test results of product performance indicators, we not only pay attention to whether the final product meets the predetermined performance standards, but also collect long-term monitoring data on the stability and reliability of the product under different environmental conditions. In terms of the acceptance of technological innovation results, a professional technical review team is organized to conduct a detailed evaluation of the degree of technological innovation, the applicable value of innovative results, and the potential role in promoting the development of industry technology based on industry standards and the expected innovation goals of the project. At the same time, we actively collect user feedback on project results in actual application scenarios, including user satisfaction surveys, actual use effect evaluations, market acceptance and other data, to comprehensively measure the quality of project results from the perspective of user experience and market feedback.

[0042] When determining the deviation relationship between the pre-scoring and the actual results, a combination of advanced time series analysis and causal inference algorithms was used. In terms of time series analysis, multivariate time series models, such as vector autoregression (VAR) models or dynamic factor models (DFM), were constructed for data such as project progress, cost, and quality of results. By analyzing the trend, seasonality, and periodicity characteristics of time series data, as well as the dynamic correlation between different variables, the trend of changes in key indicators during project execution is predicted and compared with the predicted values ​​of the pre-scoring model at the corresponding time point. Using causal inference algorithms, such as the causal discovery algorithm based on Bayesian networks or the structural equation model (SEM), the potential causal relationship between the input features of the pre-scoring model (such as project fields, technical keywords, team capabilities, etc.) and the actual execution results is deeply explored. Identify which features have a key impact on the results during project execution, as well as the direction and intensity of this impact, so as to accurately locate the possible sources of deviation in the pre-scoring model.

[0043] In the process of adjusting the weight parameters of each feature layer in the large model according to the deviation, the key features that have a greater impact on the pre-scoring are first determined based on the deviation analysis results. For those features that show a discrepancy with the pre-scoring expectations in the actual results and have a significant impact on the project results, such as the project team collaboration efficiency, which has a significant impact on the progress and quality of the results in the actual project, but has a low weight in the pre-scoring model, the system will increase its weight in the model in a targeted manner. At the same time, for those features whose impact is overestimated or underestimated in the pre-scoring, such as certain technical keywords that do not have a key impact on innovation as expected in the actual project, their weights are appropriately reduced.

[0044] When optimizing the matching and pre-scoring algorithms, a genetic algorithm (GA) or a particle swarm optimization algorithm (PSO) is introduced to search for the optimal model configuration. Taking the genetic algorithm as an example, the parameters of the pre-scoring model (such as feature weights, model hyperparameters, etc.) are encoded as chromosomes, and new model parameter combinations are continuously generated by simulating the selection, crossover, and mutation operations in the biological evolution process. The accuracy and stability of the model corresponding to each parameter combination in predicting the actual project results are evaluated according to the fitness function, and individuals with higher fitness are selected for the next generation of evolution. After multiple rounds of iterations, the optimal or near-optimal model configuration parameters are gradually converged, thereby improving the predictive ability and generalization performance of the pre-scoring model for different types of scientific and technological projects, enabling the model to more accurately reflect the actual situation of the project and provide more reliable support for project decision-making.

[0045] The knowledge graph update module of the present invention plays a key role in maintaining the timeliness and integrity of the knowledge graph, and its operation process covers multiple complex and sophisticated technical links.

[0046] When updating the nodes and relationships of the knowledge graph, advanced natural language processing technology is used to extract valuable entities and relationships from multi-source text data such as the latest scientific research literature and industry news. For scientific research literature, deep learning models (such as pre-trained models based on the Transformer architecture) are used for semantic understanding and information extraction. First, the text is preprocessed on the literature, including word segmentation, part-of-speech tagging, named entity recognition and other operations to identify key entities, such as new technology names, scientific research institutions, experts and scholars, research results, etc. Then, through syntactic analysis and semantic parsing technology, the semantic relationships between entities are mined, such as the inheritance relationship between technologies, the relationship between institutions and research results, and the relationship between scholars and research fields. For industry news, a combination of rule-based and machine learning methods is used to quickly capture dynamic information related to scientific and technological projects, and extract entities such as corporate dynamics, market trends, policy and regulatory changes and their relationships.

[0047] In the entity alignment phase, entity alignment technology based on semantic vector space model and graph matching algorithm is used. The semantic similarity between entities extracted from new data and entities in the existing knowledge graph is calculated. By representing entities as low-dimensional semantic vectors, the similarity between entities is measured using metrics such as cosine similarity and Euclidean distance. For entities with high similarity, the graph matching algorithm is further used to consider their neighbor node relationships and path information in the knowledge graph structure to determine whether they are the same entity or highly related entities, so as to achieve accurate alignment. The relationship fusion algorithm is based on logical reasoning and probabilistic statistical models to check the consistency and fuse the newly extracted relationships with the existing relationships in the knowledge graph. For conflicting or inconsistent relationships, rule-based reasoning systems and data-driven probabilistic models (such as Bayesian networks) are used for analysis and adjudication to retain the most reasonable and credible relationship information.

[0048] In the process of incorporating new knowledge into the training of the large model, incremental learning technology and knowledge distillation technology are used to ensure efficient model updating and knowledge inheritance. Incremental learning technology enables the model to learn and adjust only new knowledge based on existing knowledge, avoiding the huge consumption of computing resources caused by repeated training. By designing an efficient incremental learning algorithm, the model can quickly identify key information in new knowledge, integrate it into the existing knowledge system, and update the parameters and internal representation of the model. At the same time, knowledge distillation technology is introduced to solve the problem of forgetting the original knowledge that may occur in the process of learning new knowledge. The large model is used as the teacher model, and a smaller student model is trained to imitate the behavior of the teacher model. When training the student model, not only new data is used, but also the prediction results of the teacher model on the original data are used as additional supervision information, so that the student model can retain the original knowledge contained in the teacher model while learning new knowledge, thereby realizing the model's rapid adaptation to new knowledge and effective retention of original knowledge.

[0049] The system management and monitoring module of the present invention plays a vital role in ensuring the safe, stable and efficient operation of the entire system, and its functional implementation involves deep integration and application of multiple key technical fields.

[0050] User rights management is based on the mature and widely used role-based access control (RBAC) model. It ensures the security and confidentiality of system data and functions through a carefully designed user role system, rights allocation strategy and strict access control mechanism. First, the system defines a series of clear user roles, such as system administrator, ordinary user, project leader, expert reviewer, etc. Each role is assigned a specific set of rights. These rights are accurate to each functional module and operation level of the system. For example, the system administrator has the highest rights such as system configuration, user management, data backup and recovery, while ordinary users only have basic rights such as viewing project information and submitting project applications. Rights allocation is strictly set based on the user's job responsibilities and business needs to ensure that users can only access their authorized resources and perform corresponding operations. In the process of access control implementation, when a user logs in to the system, the identity authentication mechanism (such as username / password authentication, digital certificate authentication or multi-factor authentication) first verifies the user's identity. Once the identity authentication is passed, the system dynamically loads the corresponding permission list according to the user's role, and performs real-time permission checks when the user performs each operation. This user rights management method based on the RBAC model not only effectively prevents unauthorized access and data leakage risks, but also greatly simplifies the complexity of rights management and improves system management efficiency.

[0051] The operation log records use a distributed log system. This choice ensures the integrity, consistency and traceability of the system operation data, and provides strong support for the system's operation and maintenance management, security audit and troubleshooting. The distributed log system is based on an advanced distributed architecture design and has significant advantages such as high reliability, high scalability and high performance. In the log collection stage, by deploying log collection agents in various key nodes and business modules of the system, various operation events can be captured in real time, including user login, data query, data modification, model training and evaluation operations. The collected log data is transmitted to the distributed storage nodes in real time for centralized storage to ensure that the data will not be lost due to single point failures. During the storage process, technical means such as data sharding, redundant backup and consistent hashing algorithm are used to ensure the integrity and availability of log data in a distributed environment. At the same time, in order to facilitate query and analysis, log data is structured and stored in a certain format (such as JSON format), and an efficient indexing mechanism is established to support fast log retrieval and data mining. Through operation log records, system administrators can trace in detail the executor, operation time, operation content and operation results of each operation, so as to promptly discover potential security issues, system failures or abnormal operation behaviors, and take appropriate measures to deal with them.

[0052] System performance monitoring indicators cover multiple key aspects such as data collection delay, model inference time, and server resource utilization. These indicators fully reflect the system's operating status and performance. The data collection delay indicator is used to measure the time interval for the system to obtain data from the data source and transmit it to the processing module. The data collection delay is monitored in real time by setting timestamps at the data collection point and the data processing entrance and calculating the difference between the two. The model inference time focuses on the time required for the model to process the input data and generate output results. This indicator is crucial for evaluating the operating efficiency and response speed of the model. During the model inference process, the system calculates the inference time by accurately timing the start time and end time of the model and compares it with the preset performance threshold. The server resource utilization indicator monitors the server's CPU usage, memory occupancy, disk I / O read and write speed, and network bandwidth usage. These resource usage data are obtained in real time through the system's underlying performance monitoring interface and tools (such as sar, top and other commands in the Linux system or professional performance monitoring software). As the front-end display interface of the system management and monitoring module, the visual monitoring dashboard provides administrators with an intuitive and real-time view of the system's operating status. The dashboard uses data visualization technology (such as bar charts, line charts, dashboards, etc.) to present system performance monitoring indicators in an intuitive and easy-to-understand way, allowing administrators to understand the system's operating status at a glance. At the same time, the dashboard also has an early warning function. When any performance indicator exceeds the preset threshold, the system will automatically trigger the early warning mechanism and promptly notify the administrator in a variety of ways (such as pop-up warning boxes, sending emails, SMS notifications, etc.). After receiving the early warning notification, the administrator can quickly locate the root cause of the problem based on the detailed performance data and system log information provided by the dashboard, and adopt corresponding optimization strategies, such as adjusting server resource configuration, optimizing model algorithms, optimizing data collection processes, etc., to ensure that the system always runs at the best performance state.

[0053] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A scientific and technological project intelligent matching and pre-scoring system based on a large model, characterized in that: include: The data collection module is used to collect multi-source information of scientific and technological projects, including project applications, research reports, patent documents, and scientific research personnel resumes. The data sources include government scientific research management platforms, university scientific research databases, corporate R&D archives, and professional academic forums. The collection frequency is dynamically adjusted based on the key nodes of project progress and the dynamic change characteristics of data; The data preprocessing module performs deep cleaning, structural conversion and semantic alignment on the collected data, extracts key information using deep learning-based entity recognition and relationship extraction technology, corrects data deviations in combination with statistical analysis and rule engines, and maps data to a unified semantic framework based on domain ontology; The large model construction module builds a large model architecture that integrates knowledge distillation and reinforcement learning, including a project feature encoding layer, a matching decision layer, and a pre-scoring prediction layer. The feature encoding layer uses Transformer-Encoder combined with a graph neural network (GNN) to model the multivariate relationship of projects. The matching decision layer optimizes the matching strategy based on a policy gradient reinforcement learning algorithm. The pre-scoring prediction layer uses ensemble learning to fuse the outputs of multiple base learners. Model training uses large-scale labeled data and simulated data expanded by a generative adversarial network (GAN) to improve model generalization and robustness. The intelligent matching module inputs the preprocessed project data into the large model, generates the project feature vector through the feature coding layer, searches for similar projects based on the dynamic matching strategy of the matching decision layer, and outputs the matching degree and similar project set. The matching degree calculation introduces the calculation formula as follows: ,in For the The weight of the similarity feature, is the corresponding similarity eigenvalue, is the mapping function from feature value to matching contribution value, and n is the number of similarity features; The pre-scoring module pre-scores the project to be evaluated based on the pre-scoring prediction layer of the large model combined with the historical scoring of the matching project, outputs the score and detailed explanation, and the pre-scoring calculation takes into account the project's innovation , technology maturity T, market demand fit M, resource sustainability R, the formula is: , , , , Where is the weight of each factor, is a correction term based on the distribution of matching item scores; The result evaluation and optimization module collects the actual execution result data of the project, compares the pre-score with the actual score, uses the deviation analysis algorithm to locate the cause of the model prediction deviation, adjusts the model parameters and feature weights based on the feedback data, optimizes the matching and pre-score algorithms, and continuously improves the accuracy of the system; The knowledge graph update module monitors industry trends and scientific research progress in real time, updates knowledge graph nodes and relationships through information extraction and fusion technology, integrates new knowledge into the large model training process, and enhances the model's ability to understand emerging technologies and trends; The system management and monitoring module is responsible for system user authority management, operation log recording and system performance monitoring, and provides a visual monitoring dashboard to display system operation status and indicators.

2. According to the large model-based intelligent matching and pre-scoring system for scientific and technological projects according to claim 1, it is characterized in that: When collecting the resumes of scientific researchers, the data collection module uses natural language processing technology to parse unstructured text, extract project experience, skills and expertise, and academic achievements information, and match them through semantic similarity.

3. The large-model-based intelligent matching and pre-scoring system for scientific and technological projects according to claim 1, characterized in that: In the semantic alignment process of the data preprocessing module, the semantic similarity between words and sentences is calculated based on the pre-trained semantic vector model, and the dynamic programming algorithm is used for matching.

4. The large-model-based intelligent matching and pre-scoring system for scientific and technological projects according to claim 1 is characterized in that: The intelligent matching module calculates the similarity of the project domain When using the method based on the semantic vector space model and domain ontology tree, the formula is: ,in Indicates that the project to be evaluated P is the first The semantic vector of a node, Represents the semantic vector of the historical item at the corresponding node, is the similarity calculation function of two semantic vectors, which is calculated using the cosine similarity formula. The weight value is determined according to the node depth and association relationship in the domain ontology tree; in the calculation of the overlap of technical keywords When using the algorithm based on keyword co-occurrence network and information entropy, the formula is: ,in Indicates keywords The probability of co-occurrence between the item to be evaluated and the matching item, is the total number of keywords.

5. The large-model-based intelligent matching and pre-scoring system for scientific and technological projects according to claim 1 is characterized in that: In the pre-scoring module, the evaluation of project innovation is based on patent citation analysis and technology frontier identification technology, the technology maturity is judged through the technology life cycle model, the market demand fit is determined in combination with the market research report and the user demand forecasting model, and resource sustainability considers the supply stability and scalability of financial, human and equipment resources. The correction items in the pre-scoring calculation are calculated based on the average score deviation of the industry to which the project belongs and the project scale adjustment factor.

6. The large-model-based intelligent matching and pre-scoring system for scientific and technological projects according to claim 1 is characterized in that: When collecting the actual execution result data of the project, the result evaluation and optimization module connects with the project management system and the financial system to obtain the project progress, cost, and result quality information, uses time series analysis and causal inference algorithms to determine the deviation relationship between the pre-scoring and the actual results, adjusts the weight parameters of each feature layer in the large model according to the deviation, and introduces genetic algorithms or particle swarm optimization algorithms to search for model configurations when optimizing matching and pre-scoring algorithms.

7. The large-model-based intelligent matching and pre-scoring system for scientific and technological projects according to claim 1 is characterized in that: When updating the knowledge graph nodes and relationships, the knowledge graph update module uses natural language processing technology to extract entities and relationships from scientific research literature and industry news, and adopts entity alignment and relationship fusion algorithms to integrate new information into the existing knowledge graph. In the process of incorporating new knowledge into the large model training, incremental learning technology and knowledge distillation technology are used to make the model adapt to new knowledge and reduce the forgetting of original knowledge.

8. The large-model-based intelligent matching and pre-scoring system for scientific and technological projects according to claim 1 is characterized in that: In the system management and monitoring module, user authority management is based on the role-based access control RBAC model, operation log recording uses a distributed log system to ensure data integrity and traceability, system performance monitoring indicators include data collection delay, model inference time, server resource utilization, and a visual monitoring dashboard provides real-time data display and early warning functions. When performance indicators exceed preset thresholds, optimization strategies are automatically triggered or administrators are notified.

Images