A scientific and technological talent recommendation method and system based on multi-agent cooperation

Abstract

This application provides a method and system for recommending scientific and technological talents based on multi-agent collaboration, belonging to the field of multi-agent system technology. The method includes: constructing a user demand profile based on user talent demand information; performing parallel retrieval and fusion of multi-source heterogeneous talent databases through a retrieval and discovery agent; supplementing and verifying candidate information in the preliminary talent candidate list through an information enhancement agent; performing multi-dimensional quantitative evaluation of candidates in standardized talent files based on the user demand profile through a multi-dimensional evaluation agent; and scheduling the execution order and data flow of each agent unit through a collaborative management agent, and recording the entire process information to provide traceability of talent and scientific and technological information. This application can solve the technical problems of inaccurate talent recommendation caused by incomplete talent information integration and excessive reliance on manual processes in existing technologies. By working collaboratively with multiple agents, the efficiency of scientific and technological talent recommendation is improved.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system technology, specifically to a method and system for recommending scientific and technological talents based on multi-agent collaboration. Background Technology

[0002] Technology talent recommendation primarily relies on traditional methods such as talent pool retrieval and recruitment platform screening. Some solutions introduce single algorithm models to assist matching, attempting to improve recommendation efficiency. Traditional talent recommendation processes generally include steps such as demand analysis, information retrieval, screening, and matching ranking. While these methods alleviate the efficiency problem of talent matching to some extent, they still have significant shortcomings and deficiencies when considering the highly specialized, diverse, and results-oriented nature of technology talent. First, the accuracy of demand analysis is low. Most methods rely on simple keyword matching, failing to effectively break down and standardize vague demands, resulting in recommendations that do not accurately match actual needs. Second, talent discovery channels are too limited. Existing technologies are usually confined to internal talent pools or a few external platforms, failing to cover more potential sources of technology talent, such as academic platforms and open-source communities, making it difficult to identify high-quality technology talent with core technical capabilities who are not actively seeking employment. Furthermore, existing methods mostly remain at the level of basic personal information and skill tags, lacking the integration of deeper information such as project achievements and academic contributions, affecting the accuracy and comprehensiveness of talent recommendations. Finally, existing technologies rely excessively on manual screening and evaluation, leading to low efficiency and high costs in the recommendation process.

[0003] In summary, existing technologies suffer from incomplete talent information integration and excessive reliance on manual processes, leading to inaccurate talent recommendations and further impacting the efficiency of science and technology talent recommendation. Summary of the Invention

[0004] This application provides a method and system for recommending scientific and technological talents based on multi-agent collaboration, which addresses the technical problems in existing technologies where incomplete integration of talent information and excessive reliance on manual processes lead to inaccurate talent recommendations, further affecting the efficiency of scientific and technological talent recommendation.

[0005] In view of the above problems, this application provides a method and system for recommending scientific and technological talents based on multi-agent collaboration.

[0006] Firstly, this application provides a method for recommending scientific and technological talents based on multi-agent collaboration. This method is implemented through a system for recommending scientific and technological talents based on multi-agent collaboration. The method includes: based on user talent demand information, a demand analysis agent performs multi-dimensional analysis, extracts demand features according to each dimension, and constructs a user demand profile; using the user demand profile as input constraints, a retrieval and discovery agent performs parallel retrieval and fusion of multi-source heterogeneous talent databases to generate a preliminary talent candidate list; an information enhancement agent supplements and verifies the candidate information in the preliminary talent candidate list to generate standardized talent profiles; a multi-dimensional evaluation agent performs multi-dimensional quantitative evaluation of the candidates in the standardized talent profiles based on the user demand profile, calculates a comprehensive score, and generates an interpretable recommendation report; and a collaborative management agent schedules the execution order and data flow of each agent unit and records the entire process information to provide traceability of talent and technology information.

[0007] Optionally, the semantic analysis module performs multi-dimensional demand analysis on the input demand text information to identify explicit and implicit demand information; attributes are extracted from the explicit and implicit demand information respectively, and the extracted attributes and attribute values ​​are mapped to a preset structured demand profile field library to generate a multi-dimensional structured demand profile; priority weights are configured for each dimension parameter using the weight allocation module, and the weights allocated to each dimension parameter are added to the multi-dimensional structured demand profile to form a weighted user demand profile.

[0008] Optionally, the demand text information is preprocessed, and the multi-dimensional description of talent demand rules in the text information is extracted through the demand rule semantic analysis module to obtain multi-dimensional explicit demand information, including science and technology fields, academic achievement indicators, project experience parameters, and job descriptions; based on the explicit demand information, reasoning analysis is performed to identify implicit demand information with explicit correlation in each dimension.

[0009] Optionally, the influence relationship analysis of the explicit and implicit demand information and demand parameters of each dimension is performed to determine the strength of the influence relationship on the demand for scientific and technological talents; the priority weight of each dimension demand parameter is configured according to the strength of the influence relationship, and the weight is labeled on the multi-dimensional structured demand profile to obtain a weighted user demand profile.

[0010] Optionally, the internal talent pool retrieval module uses vector embedding technology and an approximate nearest neighbor algorithm to perform semantic similarity retrieval within the internal talent pool; the external achievement database discovery module retrieves high-impact achievements and extracts personnel information from academic databases, patent databases, and project databases based on domain keywords; the large model-enhanced search module searches for qualified scientific and technological talent information from open websites; and the multi-source result fusion module performs confidence assessment and deduplication on the retrieval results of each module to generate a preliminary talent candidate list.

[0011] Optionally, based on user demand profiles, demand parameters and priority weights are determined; based on the demand parameters, information for each candidate in the preliminary talent candidate list is determined to be missing, and an enhanced search is performed using a large model based on the missing demand parameters to retrieve publicly available talent information of the candidates; the retrieved talent information is categorized and integrated according to a preset standardized talent profile template to obtain preliminary talent profile information; a multi-source cross-validation mechanism is used to verify the authenticity and consistency of the supplemented preliminary talent profile information; the verification results are confirmed based on priority weights, and the standardized talent profile is generated using the verified preliminary talent profile information.

[0012] Optionally, the authenticity of the published information can be verified through journal and conference paper databases to obtain paper verification results; the authenticity of the project experience can be verified through a research project database; the consistency of the research direction can be verified through the official website of the candidate's institution, the introduction of the research group, and the citation information of peers' academic achievements to obtain academic achievement verification results; the timeline logic of the candidate's research experience and team management experience can be checked through information from multiple sources to obtain scientific and technological experience verification results; the successful verification results and the doubtful verification results from all verification results are integrated and verified with annotations, and the doubtful verification results are sent to the collaborative verification channel for manual verification intervention to generate the final verification result.

[0013] Optionally, a demand mapping relationship is established between the standardized talent files and user demand profiles; based on the demand mapping relationship, demand evaluation is performed in each dimension according to the demand response relationship and priority weight, and a comprehensive score is generated; candidates are ranked according to the comprehensive score, and a large language model is called to analyze and compare the demand parameters of each candidate, generating an interpretable recommendation report for comparing the explicit and implicit matching differences of candidates and the reasons for recommendation.

[0014] Optionally, the system analyzes the developmental needs of users' job skills, performs interactive analysis of the explicit and implicit needs information in the user needs profile to establish a timeline of needs development; inputs the candidate's standardized talent profile and evaluation results from various dimensions into a large language model, understands the time evolution logic of the scientific and technological talent needs goals through the large language model, extracts the time sequence of key events related to time in the candidate's experience and achievements, analyzes the update speed of the candidate's knowledge structure, the evolution trajectory of research direction, and the responsiveness to new technology fields, and predicts the candidate's scientific and technological development evolution trend; based on the timeline alignment algorithm, matches the candidate's scientific and technological development evolution trend with the timeline of needs development to identify the degree of matching of development needs; incorporates the degree of matching of development needs as a weighted item into the multi-dimensional score to obtain a comprehensive score, and outputs structured timeline matching labels and explanatory text.

[0015] Secondly, this application also provides a science and technology talent recommendation system based on multi-agent collaboration, used to execute a science and technology talent recommendation method based on multi-agent collaboration as described in the first aspect. The system includes: an interaction demand feature extraction module, used to perform multi-dimensional analysis based on user talent demand information through a demand analysis agent, extracting demand features according to each dimension to construct a user demand profile; a parallel retrieval and fusion module, used to take the user demand profile as input constraints, and through a retrieval discovery agent, perform parallel retrieval and fusion of multi-source heterogeneous talent databases to generate a preliminary talent candidate list; a supplementary verification module, used to supplement and verify the candidate information in the preliminary talent candidate list through an information enhancement agent, generating standardized talent profiles; a multi-dimensional quantitative evaluation module, used to perform multi-dimensional quantitative evaluation of candidates in the standardized talent profiles based on the user demand profile through a multi-dimensional evaluation agent, calculating a comprehensive score and generating an interpretable recommendation report; and a multi-agent collaborative management module, used to schedule the execution order and data flow of each agent unit through a collaborative management agent, and record the entire process information to provide traceability of talent and technology information.

[0016] One or more technical solutions provided in this application have at least the following beneficial effects: By analyzing user talent needs information through a demand analysis agent across multiple dimensions and extracting demand features according to each dimension, a user demand profile is constructed. Using this profile as input constraints, a retrieval and discovery agent performs parallel retrieval and fusion of multi-source heterogeneous talent databases to generate a preliminary talent candidate list. An information enhancement agent supplements and verifies the candidate information in this preliminary list, generating standardized talent profiles. A multi-dimensional evaluation agent performs multi-dimensional quantitative evaluation of candidates in the standardized talent profiles based on the user demand profile, calculates a comprehensive score, and generates an interpretable recommendation report. A collaborative management agent schedules the execution order and data flow of each agent unit and records the entire process information for talent and technology information traceability. In other words, through demand analysis, retrieval and discovery, information enhancement, multi-dimensional evaluation, and collaborative management agents, an end-to-end intelligent processing flow from demand analysis to talent recommendation is achieved. This process employs a hierarchical and progressive processing logic, with each agent cooperating through a collaborative mechanism to jointly complete talent discovery, evaluation, and recommendation tasks, improving the comprehensiveness of talent information integration and thus increasing the efficiency of technology talent recommendation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in 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 merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for recommending scientific and technological talents based on multi-agent collaboration, as proposed in this application.

[0019] Figure 2 This is a schematic diagram of the structure of a science and technology talent recommendation system based on multi-agent collaboration according to this application.

[0020] Figure labeling: Interaction requirement feature extraction module 11, parallel retrieval and fusion module 12, supplementary verification module 13, multi-dimensional quantitative evaluation module 14, multi-agent collaborative management module 15. Detailed Implementation

[0021] This application provides a method and system for recommending scientific and technological talents based on multi-agent collaboration. It addresses the technical problems in existing technologies where incomplete talent information integration and excessive reliance on manual processes lead to inaccurate talent recommendations, further impacting the efficiency of talent recommendation. By employing demand analysis agents, retrieval and discovery agents, information enhancement agents, multi-dimensional evaluation agents, and collaborative management agents, an end-to-end intelligent processing flow from demand analysis to talent recommendation is achieved. A hierarchical and progressive processing logic is adopted, with each agent cooperating through a collaborative working mechanism to jointly complete talent discovery, evaluation, and recommendation tasks. This improves the comprehensiveness of talent information integration, thereby enhancing the efficiency of scientific and technological talent recommendation.

[0022] Example 1, as Figure 1 As shown, this application provides a method for recommending scientific and technological talents based on multi-agent collaboration. This method is applied to a system for recommending scientific and technological talents based on multi-agent collaboration, and specifically includes the following steps: Based on user talent demand information, a demand analysis agent performs multi-dimensional analysis, extracts demand features according to each dimension, and constructs a user demand profile.

[0023] Furthermore, this application also includes the following steps: performing multi-dimensional demand analysis on the input demand text information through a semantic analysis module to identify explicit demand information and implicit demand information; extracting attributes from the explicit and implicit demand information respectively, mapping the extracted attributes and attribute values ​​to a preset structured demand profile field library to generate a multi-dimensional structured demand profile; configuring priority weights for each dimension parameter using a weight allocation module, and adding the weights allocated to each dimension parameter to the multi-dimensional structured demand profile to form a weighted user demand profile.

[0024] Furthermore, this application also includes the following steps: preprocessing the demand text information, extracting multi-dimensional descriptions of talent demand rules from the text information through a demand rule semantic analysis module, obtaining multi-dimensional explicit demand information, including science and technology fields, academic achievement indicators, project experience parameters, and job descriptions; and performing reasoning analysis based on the explicit demand information to identify implicit demand information with explicit correlations in each dimension.

[0025] Furthermore, this application also includes the following steps: performing an impact relationship analysis on the explicit and implicit demand information and demand parameters of each dimension to determine the strength of the impact relationship on the demand for scientific and technological talents; configuring the priority weight of each dimension demand parameter according to the strength of the impact relationship, and weighting the multi-dimensional structured demand profile to obtain a weighted user demand profile.

[0026] Specifically, the demand analysis agent receives talent demand information input by users and uses a large model as a semantic understanding engine to transform the talent demand described in natural language into a structured demand profile. The demand analysis agent includes a semantic analysis module, an attribute extraction module, a demand structuring module, and a weight allocation module. The semantic analysis module uses a large language model to identify explicit and implicit demand information in the demand text; the attribute extraction module extracts key attributes and their values ​​for scientific and technological talents, such as their professional field, age requirements, talent level, and ability requirements; the demand structuring module transforms the analysis results into a multi-dimensional structured demand profile; and the weight allocation module assigns initial priorities to different dimensions, such as core technology stack as a hard requirement and project experience as a key evaluation aspect.

[0027] The preprocessing of user talent demand information text includes operations such as removing redundant information, standardizing terminology, and word segmentation to clean and standardize the data in the text. A semantic analysis module uses a large model to extract structured explicit features from the preprocessed demand text, performs word segmentation, and accurately identifies attributes and attribute values ​​in the text, such as technical terms, technology stack names, academic achievement descriptions, job titles, quantities, and time expressions. This includes structured information such as technology stack proficiency, project role, and achievement quantification indicators. Finally, a rule engine and classifier map and populate the extracted attributes and attribute values ​​into pre-defined structured template fields, such as technical fields, project experience parameters, academic achievement indicators, and job competency dimensions, generating a machine-readable, multi-dimensional parameterized explicit demand data table. This provides accurate input for subsequent user demand profiling. Explicit requirements information is multi-dimensional, including the technology field, academic achievement indicators, project experience parameters, job description, academic achievement indicators (such as paper publication level, citation frequency, patent type), project experience parameters (such as project duration, role, technology stack), and job competency dimensions (such as team management size, industry influence, management experience), etc. Explicit requirements are those directly expressed and clearly listed by the user in the requirement text.

[0028] After obtaining complete explicit information, the implicit information within the requirements is analyzed and reasoned about based on a large model and pre-defined rules, thus enriching and making the requirements semantically explicit. For example, for explicit requirement A, the large model is used to analyze its implicit requirements B and C based on pre-defined rules such as age and leadership ability requirements. If B and C do not appear in the explicit requirements, but according to the rules (e.g., if A is required, it usually implies mastery of B), B and C are presumed to be implicit requirements. Implicit requirements are those that are not directly stated but can be obtained through analysis or reasoning using the large model.

[0029] By using attribute extraction technology, specific attributes and attribute values ​​can be extracted from explicit and implicit requirement information. In other words, attribute information with specific meanings can be identified from text, such as skills, job titles, project experience time, etc.

[0030] The extracted attributes and attribute values ​​are mapped to a predefined structured requirement profile field library. This library is a predefined data structure that stores standardized requirement fields. Based on classic models in the human resources field and talent standards for technology companies, a top-level dimension is designed, including hard requirements, core skill stack, projects and experience, academic achievements and accomplishments, leadership and soft skills, and potential and suitability. Each dimension is further decomposed into specific fields that can be filled in. For example, the core skill stack dimension includes a list of programming languages, a list of mainstream frameworks, a list of professional tools, and a list of domain knowledge; the projects and experience dimension includes the type of leading projects, the largest team size, years in relevant industries, and a description of key technical challenges. Based on the attribute type, the attributes and attribute values ​​are filled into the corresponding positions in the predefined structured requirement profile field library, outputting a multi-dimensional structured requirement profile. In other words, the attributes and attribute values ​​extracted from the requirement text are matched to corresponding fields in the field library, identifying which fields in the library correspond to the attributes. After mapping, the attribute information extracted from the text is filled into the corresponding fields. A multi-dimensional structured requirement profile is a structured data object formed by filling all attributes and attribute values ​​into the corresponding positions in the field library according to their semantics. It is usually in JSON format.

[0031] The weighting module conducts in-depth analysis of explicit and implicit requirements, understanding their inherent connections and the actual impact of each requirement on the project. For example, in R&D projects, the impact coefficient for core technology matching is 0.4, and the impact coefficient for team management experience is 0.3. Based on the project type of the current requirement, the module infers the corresponding impact coefficient from the requirement text or specifies it for the user. These coefficients are then normalized so that the sum of all weights is 1, which is the basic weight for each dimension. Combining the strength of requirement descriptive words, such as mastery, proficiency, and priority, the basic weights are fine-tuned according to specific scenarios to obtain the final priority weights for each dimension's requirement parameters. The structured requirement profile is then weighted to obtain a weighted user requirement profile.

[0032] Based on the received user talent demand information, if it is identified as a routine talent recruitment demand, the demand information is pushed to the demand analysis agent according to the preset process. The demand is marked as a first-level processing priority, and the task scheduling process of each agent unit is initiated. The demand reception time, demand type, processing priority, and other information are simultaneously recorded in the information traceability database. After receiving the user demand, the demand analysis agent sequentially calls the semantic analysis module, attribute extraction module, demand structuring module, and weight allocation module to complete the conversion of natural language demand into a structured demand profile. Using a general large model as the semantic understanding engine, the demand text information in the input user talent demand information is deeply analyzed to identify the explicit requirements as: learning direction, background in reinforcement learning and multi-agent collaborative research, more than 5 years of relevant experience, and NeurIPS / ICML top conference papers. The implicit conditions are: young (18-45 years old), top-notch talent (undertaking important national projects, receiving relevant honors and awards), and team leadership ability. The core attributes and attribute value requirements are extracted from the input requirement text information. These core attributes include: professional field (machine learning, reinforcement learning, multi-agent collaboration); work / research experience ≥ 5 years; achievements (high-level publications); conference list (NeurIPS, ICML); ability requirement (team leadership); and talent level (outstanding young talent). The constraints between these attributes are clearly defined, such as achievements requiring high-level publications in specified top-tier conferences, and ability requirements requiring practical team management skills. The extracted key attributes and attribute values ​​are mapped to a pre-defined structured requirement profile field library to generate a multi-dimensional structured requirement profile covering core dimensions such as talent level, professional field, research experience, academic achievements, and leadership ability. Based on the core requirements for attracting outstanding young talents in machine learning, initial priority weights are assigned to each dimension and specific requirement item. The achievement requirement dimension (0.3), research background and professional field dimension (0.25), and ability requirement dimension (0.25) are high-weighted; the experience requirement dimension (0.15) is medium-weighted; and the talent level dimension (0.05) is low-weighted. Within each dimension, specific requirements items are further weighted. For example, in the research background dimension, reinforcement learning and multi-agent collaboration have a weight of 0.5; in the achievement requirement dimension, NeurIPS publications and ICML publications have a weight of 0.5. The requirement analysis agent outputs the generated weighted structured requirement profile to the collaborative management agent. The information collaboration module of the collaborative management agent stores this profile in standardized JSON format and pushes it to the retrieval and discovery agent in real time, while simultaneously recording the requirement parsing process and results in the information traceability database.

[0033] By comprehensively identifying both explicit and implicit information in the requirements, the accuracy of understanding the requirements has been improved, and the recommendation accuracy and efficiency of the talent recommendation system have been significantly optimized, especially when facing complex technology positions with multi-dimensional requirements.

[0034] Using the user demand profile as input constraints, the search and discovery agent performs parallel retrieval and fusion of multi-source heterogeneous talent pools to generate a preliminary talent candidate list.

[0035] Furthermore, this application also includes the following steps: using vector embedding technology and approximate nearest neighbor algorithm in the internal talent pool retrieval module to perform semantic similarity retrieval in the internal talent pool; using the external achievement database discovery module to retrieve high-impact achievements and extract personnel information from academic databases, patent databases, and project databases based on domain keywords; using the large model enhancement search module to search for qualified scientific and technological talent information from open websites; and using the multi-source result fusion module to perform confidence assessment and deduplication on the retrieval results of each module to generate a preliminary talent candidate list.

[0036] Specifically, the retrieval and discovery agent is responsible for the parallel retrieval and fusion discovery of multi-source heterogeneous talent data, aggregating previously scattered and hidden talent signals into a high-quality preliminary candidate list. This includes an internal talent pool retrieval module, an external achievement database discovery module, a large-model-enhanced search module, and a multi-source result fusion module. The internal talent pool retrieval module uses vector embedding technology to map talent features to a high-dimensional semantic space, enabling fast similarity retrieval based on an approximate nearest neighbor algorithm, supporting multi-attribute combination queries and hierarchical filtering mechanisms. The external achievement database discovery module uses domain keywords to retrieve high-impact achievement data from academic paper databases, patent databases, project databases, etc., and selects personnel information from them. The large-model-enhanced search module uses a large model to search for talent information from open websites such as institutional websites, academic platforms, professional recruitment platforms, conference directories, and industry data. The multi-source result fusion module uses a weighted voting fusion strategy to integrate candidate results from different data sources and retrieval methods, generating a preliminary talent candidate list through confidence evaluation and deduplication.

[0037] Vector embedding technology is used to transform user demand profiles and the characteristics of each talent in the internal talent pool into high-dimensional vectors. Vector embedding transforms unstructured text information into numerical vectors in a high-dimensional space using a pre-trained language model. Texts with similar semantics have vectors that are closer together in space. The standardized profiles of each tech talent in the internal talent pool are periodically converted into a comprehensive text description, and corresponding vectors are generated using the vector embedding model and stored in a dedicated vector database. The approximate nearest neighbor algorithm is a highly efficient similarity search algorithm designed specifically for high-dimensional vectors. It can find a batch of candidate vectors most similar to the target vector in a vector database of hundreds of millions of vectors within milliseconds. While not absolutely precise, it greatly balances accuracy and speed. The top-K vectors most similar to Q are searched in the vector index (e.g., K=1000), resulting in a preliminary list of candidate IDs and similarity scores based on semantic similarity. After obtaining the preliminary list, hard criteria from the demand profile are applied for rapid filtering to obtain the internal search results.

[0038] Core domain keywords are extracted from user demand profiles and expanded using synonyms. Based on the expanded keyword list, structured queries are initiated into academic databases, patent databases, and project databases. This not only retrieves the research results themselves but also extracts information such as the names, affiliations, and positions of relevant personnel. Typically, the first author, corresponding author, or project leader are prioritized.

[0039] The structured user demand profile is input into a large model, which generates a dataset suitable for general search engines. A web crawler framework is used to execute the generated search commands, obtaining an initial list of web pages. Then, the large model is used again to read and understand the content of the web pages, extracting structured talent information and determining its match with the user's needs, thereby obtaining information on qualified technology professionals.

[0040] A multi-source results fusion module employs a weighted voting fusion strategy to assess the confidence of search results from the internal talent pool retrieval module, the external results database discovery module, and the large model augmentation search module. Based on the authority of the data source itself, the freshness of the information, and the reliability of the extraction process, a base confidence score is assigned to each candidate record returned by each module. The same candidate may be discovered by multiple modules, which is treated as votes from different review panels, but each review panel (data source) has a different weight (confidence score), and the final score is determined by the weighted vote count. For example, the internal module returned 50 people, the external results database returned 30, and the large model search returned 20. After deduplication, it was found that 10 people appeared in both the internal and external databases, and 5 people appeared in both the external database and the open network. Weights are set as follows: Internal 0.5, External 0.35, Open Network 0.15. For candidate A, who appears in both internal and external databases, the internal score is 0.9 and the external score is 0.8, so the overall score is 0.5*0.9 + 0.35*0.8 = 0.73. For candidate B, who only appears in the Open Network, the score is 0.95, so the overall score is 0.15*0.95 = 0.1425.

[0041] By deduplicating data, each candidate is ensured to appear only once, preventing the same person from being repeatedly recommended by multiple modules. The specific process includes: standardizing and cleaning all names; using a rule-based and model-based entity linking system, clustering and deduplication are performed using auxiliary information (such as institution, email, personal homepage, and unique academic ID). For example, determining whether J. Smith from a paper and John Smith from GitHub are the same person. All candidates are sorted from highest to lowest based on their comprehensive score, and the top N are selected to generate a preliminary talent candidate list. Simultaneously, data is retrieved from multiple heterogeneous data sources, expanding the range of candidate sources. Through semantic similarity retrieval and large-scale model-enhanced search, candidates that accurately match the user's needs are precisely matched.

[0042] First, vector embedding technology is used to map the professional fields, academic experience, research experience, achievements, honors and awards, and ability characteristics of all talents in the internal talent pool to a high-dimensional semantic space, generating talent feature vectors. Then, based on the approximate nearest neighbor algorithm, similarity retrieval is performed in the high-dimensional semantic space using the core keywords of the structured demand profile as search conditions. Simultaneously, multi-attribute combination queries and hierarchical filtering mechanisms are enabled to return candidates selected from the internal talent pool. Domain keywords, such as machine learning, reinforcement learning, and multi-agent collaboration, and achievement requirements, such as high-level papers, are extracted from the structured demand profile. Searches are then conducted using conference paper databases, journal paper databases, and science and technology project databases. Researchers who have published papers in reinforcement learning and multi-agent collaboration as core authors are selected from the conference paper database. Researchers who have led or been core participants in the highest-level machine learning projects and hold team management roles are selected from the science and technology project database, retrieving information on several candidates with high-impact achievements. The structured demand profile is converted into precise search commands recognizable by a large model. This general model is then invoked to search open websites such as the official websites of AI labs at globally renowned universities, international research institutions, professional academic platforms, directories of top AI conferences, and industry technical communities. The search retrieves information on young scholars with backgrounds in reinforcement learning and multi-agent collaborative research, over 5 years of relevant experience, and team leadership experience. Several candidate candidates are identified. First, the confidence level of all candidates retrieved from the three modules is assessed. A confidence score is assigned to each candidate based on the authority of the data source: 100% for internal talent pools, 95% for paper databases, 90% for university / research institution websites, and 75% for open platforms. Then, deduplication is performed, removing duplicate candidates with essentially identical names, affiliations, research directions, and core achievements. The search discovery agent outputs a preliminary candidate list to the collaborative management agent. The information collaboration module standardizes and stores the list before immediately pushing it to the information enhancement agent.

[0043] The information-enhanced intelligent agent supplements and verifies the candidate information in the preliminary talent candidate list, generating standardized talent profiles.

[0044] Furthermore, this application also includes the following steps: determining demand parameters and priority weights based on user demand profiles; determining the missing information of each candidate in the preliminary talent candidate list based on the demand parameters, and performing enhanced search by calling a large model based on the missing demand parameters to retrieve publicly available talent information of the candidates; classifying and integrating the retrieved talent information according to a preset standardized talent profile template to obtain preliminary talent profile information; verifying the authenticity and consistency of the supplemented preliminary talent profile information using a multi-source cross-validation mechanism; confirming the verification results based on priority weights, and generating the standardized talent profile using the verified preliminary talent profile information.

[0045] Furthermore, this application also includes the following steps: verifying the authenticity of the published paper information through journal paper databases and conference paper databases to obtain paper verification results; verifying the authenticity of the project experience through research project databases, and verifying the consistency of the research direction through the official website of the candidate's institution, the introduction of the research group, and the citation information of peer academic achievements to obtain academic achievement verification results; verifying the timeline logic consistency of the candidate's research experience and team management experience through information from multiple sources to obtain scientific and technological experience verification results; integrating the successful verification results and the questionable verification results from all verification results, making verification annotations, and sending the questionable verification results to the collaborative verification channel for manual verification intervention to generate the final verification result.

[0046] Specifically, the information-enhancing agent is responsible for verifying, supplementing, and improving the retrieved talent information, including an information supplementation module and an information verification module. The information supplementation module further searches for publicly available resume information, academic achievements, project experience, honors and awards, academic influence, etc., of candidates through a large model, and integrates the information to form a standardized talent profile; the information verification module adopts a multi-source cross-validation mechanism, such as verifying the consistency of research directions through information on authors of academic papers, and checking the timeline logical consistency of information from multiple sources.

[0047] Extract all requirement parameters (i.e., the dimensions to be examined) and the priority weight of each parameter from the user requirement profile. Requirement parameters are the specific requirement dimensions in the user requirement profile; priority weights are the importance weights of each requirement parameter in the requirement profile, typically a value between 0 and 1, with the sum of all weights equal to 1. Based on the requirement parameters, determine the missing information for each candidate in the preliminary talent candidate list. Iterate through each candidate, checking if their existing information contains the information required by the requirement parameters; if not, mark it as missing. Construct search queries for missing requirement parameters. Use a large model to generate more intelligent search instructions. The large model can generate multiple search queries based on the candidate's known information (such as name, company, and position) and the missing parameters, and may directly extract relevant information from the search results. Call the large model's API, design a prompt, and use the generated query to call the search engine. After obtaining the search results, use the large model again to extract information from the search results, determine if it contains the required information, and extract it to obtain the publicly available supplementary information for each candidate.

[0048] The existing and supplementary information of candidates is combined to form a complete but potentially disorganized information set. A standardized talent profile template is a predefined structured data template used to uniformly store talent information, typically including fields such as basic personal information, educational background, work experience, project experience, skills, and academic achievements. Using this pre-defined standardized talent profile template to integrate all retrieved information, all candidate information is organized into a consistent and standardized format, resulting in preliminary talent profile information.

[0049] Multi-source cross-validation is a method that combines multiple data sources to verify the authenticity and consistency of talent information. Different data sources cross-validate each other, effectively reducing the possibility of errors or inconsistencies. Authentic information leaves a traceable mark on multiple credible channels, while false or erroneous information often struggles to maintain consistency across all channels. Therefore, multi-source cross-validation is more reasonable. Authenticity verification confirms whether information (such as a paper or project) objectively exists and is not fabricated by the candidate. Consistency verification confirms whether descriptions of the same fact from different sources are consistent and whether there are logical contradictions between different experiences of the candidate, especially timeline conflicts.

[0050] By accessing journal and conference paper databases via API, key identifiers such as title, author list, journal / conference name, year, volume, issue, page number, and DOI are extracted from each paper in the supplemented preliminary talent profile information. Using this information as search criteria, precise searches are performed in the journal and conference paper databases to obtain paper verification results, including verified, questionable, and unverified. Verified means a completely matching record was found in the authoritative database; questionable means a similar record was found but with discrepancies; unverified means no paper was found in any of the major databases.

[0051] Based on the research project experience provided by the candidate, records of the project are searched from the research project database to verify the candidate's role and contributions. Additionally, the official website of the candidate's institution and the introduction of their research group can provide evidence of the projects or research areas the candidate participated in. Member introductions from the candidate's institution's official website, laboratory, or research group homepage are crawled and parsed; their academic collaboration network is analyzed (through co-authorship analysis); and the topics of cited literature for their achievements are traced. The semantic similarity of the claimed research direction in the candidate's profile with the descriptions from the above sources is compared to obtain the verification results of their academic achievements.

[0052] It incorporates a set of common-sense time logic rules, such as: at the same point in time, a person cannot simultaneously pursue a full-time degree at two universities; the total number of years of work experience should not significantly exceed their age minus the number of years of study; and the career advancement sequence should follow conventional rules and cannot jump backwards. It extracts all timestamped events from the supplemented preliminary talent profile information, such as education experience, work experience, and project duration, arranges them on a timeline, automatically runs rule checks, and obtains the verification results for scientific and technological experience.

[0053] The outputs of all automated validators are aggregated, and a verification annotation is added next to each piece of original information. If the verification is successful, the information is marked as successfully verified; if the verification is questionable, the information is marked as questionable. Questionable verification results are sent to the collaborative verification channel for further verification by human reviewers. Human reviewers examine and compare the information sources and details. Based on all the verification results, successfully verified information is confirmed as the final valid data, and questionable information is marked, generating the final verification result.

[0054] Read the priority weights from the user requirement profile. For verification results, adopt a differentiated tolerance strategy: high-weight requirement parameters, such as core technology stack and key project experience, must have verified information; otherwise, the candidate may be marked as unreliable in this key dimension, severely impacting their score. Medium- and low-weight requirement parameters are allowed a small amount of insufficient or questionable / pending information, but this will be noted in the final report for decision-makers' reference. Only verified information is entered into the new container according to the standardized talent profile template. All questionable or unverified information will be removed or moved to a separate unverified information section. The final verified standardized talent profile—that is, the set of information whose key fields have been verified or manually determined to be reliable—will only be adopted in the final process.

[0055] After receiving the preliminary list of talent candidates, the information augmentation agent sequentially calls the information supplementation module and the information verification module. For example, to improve and verify the information of 25 candidates, the specific process is as follows: Addressing the issue of missing information for each candidate in the preliminary list, a large model is invoked for augmented search, retrieving publicly available information such as the candidate's resume, published papers, research projects, research experience, academic honors, and social positions. The retrieved information is then categorized and integrated according to a pre-set standardized talent profile template. A multi-source cross-validation mechanism is employed to verify the authenticity and consistency of the supplemented talent information. Specifically, this involves verifying the authenticity of published papers through journal and conference paper databases, verifying the authenticity of project experience through a research project database, verifying the consistency of research directions through the candidate's institution's official website, research group introduction, and peer citations, and checking the timeline logic of the candidate's research experience and team management experience through multiple sources. Questionable information is prominently marked and can also be output for manual review. After information verification and removal of candidates with questionable information, the remaining candidates undergo comprehensive information supplementation and standardized integration to generate standardized talent profiles. These profiles are then output to the collaborative management intelligent agent, stored by the information collaboration module, and pushed to the multi-dimensional evaluation intelligent agent. By configuring the weights of the required parameters and supplementing missing candidate information, the capabilities of each candidate are comprehensively evaluated, ensuring more accurate recommendations. Multi-source cross-validation effectively reduces the spread of erroneous information, ensuring that the final recommended talent information is more authentic and reliable.

[0056] The multidimensional evaluation agent performs a multidimensional quantitative evaluation of candidates in the standardized talent profile based on user needs profiles, calculates a comprehensive score, and generates an interpretable recommendation report.

[0057] Furthermore, this application also includes the following steps: establishing a demand mapping relationship between the standardized talent profile and the user demand profile; based on the demand mapping relationship, evaluating the demand in each dimension according to the demand response relationship and priority weight, and generating a comprehensive score; ranking the candidates according to the comprehensive score, and calling a large language model to analyze and compare the demand parameters of each candidate, generating an interpretable recommendation report for comparing the explicit and implicit matching differences of the candidates and the reasons for recommendation.

[0058] Furthermore, this application also includes the following steps: analyzing the development needs of users' job skills, conducting interactive analysis of the explicit and implicit needs information in the user needs profile to establish a developmental needs timeline chain; inputting the candidate's standardized talent profile and evaluation results of each dimension into a large language model, understanding the time evolution logic of the scientific and technological talent needs goals through the large language model, extracting the time sequence of key events related to time in the candidate's experience and achievements, analyzing the candidate's knowledge structure update speed, research direction evolution trajectory, and responsiveness to new technology fields, and predicting the candidate's scientific and technological development evolution trend; matching the candidate's scientific and technological development evolution trend with the needs development timeline chain based on a timeline alignment algorithm to identify the development needs matching degree; adding the development needs matching degree as a weighted item to the multi-dimensional score to obtain a comprehensive score, and outputting structured matching labels and explanatory text.

[0059] Specifically, the multidimensional assessment agent is used to quantitatively evaluate and intelligently rank candidate talents. It calculates and evaluates talents from different dimensions. First, it calculates the hard match degree, based on explicit requirement rules, to assess the degree of compliance with conditions such as professional field, age, and ability requirements. Second, it calculates the professional skill similarity based on semantic similarity and weighted professional matching algorithms. Furthermore, it considers evaluating professional proficiency and innovation ability through the quality and impact of achievements. A weighted summation algorithm is used to comprehensively calculate the evaluation results of each dimension according to the weights configured in the demand profile, generating a comprehensive score for each candidate. Recommendation reasons are automatically generated based on a large language model, highlighting key matching points and conducting a comparative analysis of differentiated advantages. The talent ranking list is integrated with standardized talent profiles to generate a standardized recommendation report and output it to the user.

[0060] A demand mapping relationship is established between structured, standardized talent profiles and user demand profiles, defining how the evaluation system should compare and measure each piece of information in the profile with each parameter in the demand profile. For complex or implicit demands, the mapping rules may point to combinations of multiple fields in the profile, such as the direction of recently published papers, the open-source technology trends they follow, and the nature of pre-research projects they have participated in. For each candidate, the degree to which they meet user demands is assessed based on the characteristics (such as skills and experience) in their standardized profile. The overall score is the result of a weighted sum of the scores for each demand dimension according to the weights assigned to the structured demand profile. Candidates are ranked according to their overall scores. The candidate with the highest overall score is ranked first and recommended to the user first.

[0061] After receiving the structured demand profile and candidate talent files, the multidimensional evaluation agent sequentially calls the multidimensional evaluation module, the demand fit comprehensive calculation module, the interpretability recommendation module, and the result output module to complete the quantitative evaluation of talents and generate a recommendation report. According to the dimensions of the structured demand profile, each candidate is quantitatively scored in multiple dimensions using a 10-point scale. Higher scores indicate higher matching and fit. Specific evaluation rules are as follows: Based on the candidate's professional title, recognition within the field, and talent titles, the evaluation assesses whether they meet the criteria for outstanding young talents. Candidates with a young talent title and who are core researchers in their institution receive 10 points; those with associate senior professional titles or above in universities / research institutions and significant influence in their field receive 8 points; and those with intermediate professional titles and outstanding achievements receive 6 points. The semantic similarity algorithm calculates the fit between the research direction and reinforcement learning and multi-agent collaboration. A core research direction of reinforcement learning + multi-agent collaboration with in-depth research receives 10 points; focusing on one direction with interdisciplinary research receives 8 points; and research directions related but not in core areas receive 4 points. Scoring is based on the number of years of relevant work / research experience and the degree of matching of experience fields. 10 points are awarded for 5 years or more and full matching of reinforcement learning and multi-agent collaboration fields. 8 points are awarded for more than 5 years but with some overlap in fields. 6 points are awarded for 3-5 years and core field matching. 0 points are awarded for less than 3 years. A comprehensive score is awarded based on academic publications, including author contributions, number of papers, journal / conference ranking, and citation count. Papers published as first / corresponding author are weighted at 1.0, others at 0.2. Top journals and conferences in the field are weighted at 1.0, others at 0.2. Highly cited papers are weighted at 1.0, others at 0.2. The score for each candidate's top 5 papers published in the past 10 years is calculated accordingly. Other criteria include: 10 points for being the first inventor with at least two authorized invention patents in reinforcement learning or multi-agent collaboration and achieving technology transfer; 8 points for having one authorized invention patent with transfer potential; 5 points for having a utility model patent or published invention patent application; 10 points for being the chief editor of a machine learning-related academic monograph with high industry recognition; 7 points for being an associate editor of a related monograph; and 4 points for participating in monograph writing. The overall score is calculated by weighting the sub-dimensions within the achievement requirements. The evaluation is based on factors such as team management experience and team achievements. A candidate who has led a research team for at least 5 years with top-tier conference publications / approved projects receives 10 points; a candidate who has led a team for at least 3 years with significant achievements receives 8 points; and a candidate with team leadership experience and tangible results receives 5 points. A weighted summation algorithm is used to comprehensively calculate the candidate's overall score by combining the scores from each dimension according to the pre-defined weights in the structured requirements profile.

[0062] This analysis examines the developmental needs of users' job skills, focusing on the deep characteristics related to the evolution of job roles and the iteration of capabilities over time. It not only focuses on current skill requirements but also on the required skill levels, technical depth, and role positioning at different future stages. The analysis involves interactive analysis of explicit and implicit needs stored in the user needs profile to identify developmental needs. Through this interactive analysis, a timeline of job requirements is constructed, describing the dynamic evolution of job requirements over time. This timeline consists of several consecutive development stages, each including elements such as a time window, core tasks, required skill levels, and expected outcomes.

[0063] Extract all time-stamped key events from the candidate's standardized talent profile, including the start and end dates of their educational experience, the start and end dates of each job change, the duration of each project they participated in, the publication date of each paper, the authorization date of each patent, and records of each promotion. Arrange these events in chronological order to form a timeline of the candidate's key events.

[0064] By utilizing a large language model and based on each candidate's standardized talent profile, personalized recommendation reasons are automatically generated, highlighting core matching points and outstanding abilities. Furthermore, a comparative analysis of the differentiated advantages of the top 10 candidates with the highest comprehensive scores is conducted to clarify each candidate's research direction characteristics, core achievements, team management highlights, and the fit between their research directions and the needs of the candidates.

[0065] Large language models, leveraging their powerful semantic understanding and reasoning capabilities, perform analyses of knowledge structure update speed, research direction evolution trajectory, and responsiveness in new technology fields. By examining changes in the candidate's technology stack, programming languages, development tools, and algorithmic frameworks used at different times, they calculate the frequency and depth of their technological updates. For example, a large language model identifies which technologies a candidate primarily used during a given period and which new technologies they switched to or added in the next, thus determining the pace of their technological updates. By analyzing changes in the topics of the candidate's published papers, the areas of projects they participated in, and the directions of their patent applications over different time periods, they depict the evolution path of their research interests. For example, they identify whether a candidate's research direction gradually expands from a single field to related fields, completely crosses boundaries from one field to another, or focuses on a specific sub-field for a long period. They also identify emerging technological waves that have occurred during the candidate's career, such as the rise of cloud computing, the explosion of deep learning, and the emergence of large language models. Then, they analyze how long it took for the candidate to produce relevant outputs after the emergence of these new technologies, such as publishing related papers, participating in related projects, and applying for related patents. The length of the time interval directly reflects the candidate's technological acumen and learning and transformation capabilities. By comprehensively inferring the candidate's possible future development trajectory, a structured description of their technological development and evolution trend is formed, which includes both a summary of past trajectories and a prediction of future trends.

[0066] The candidate's technological development trend is matched with the timeline of demand development. Each stage in the demand development timeline and each key node in the candidate's technological development trend is transformed into a multi-dimensional feature vector, including technical depth, ability level, task type, and degree of innovation, ensuring a comprehensive portrayal of the essential characteristics of each stage or node. A time-series alignment algorithm is used to match these two sequences, handling any scaling or offset between them on the timeline to find the most reasonable correspondence. This not only compares the similarity of current abilities but also focuses on the consistency of development rhythm, i.e., whether the candidate's ability leap coincides with the upgrading of job requirements at specific points in time.

[0067] After aligning the candidate's performance, the following calculations are performed: Stage Coverage Rate (what proportion of the key stages in the demand development timeline are covered by the candidate's development trajectory); Synchronicity Coefficient (the degree of alignment between the candidate's skill advancement and the job requirement upgrade timeline; for example, if the job requires management skills in the second year, and the candidate has team management experience from the end of the first year to the second year, the synchronicity is high); and Lead-or-Lag Index (whether the candidate's development is ahead of, in sync with, or lagging behind the job requirements; too much lead-or-lagging may mean the candidate is prone to leaving due to a lack of challenge, while too much lagging may mean they cannot keep up with the job requirements). The Stage Coverage Rate, Synchronicity Coefficient, and Lead-or-Lag Index are then combined to generate a normalized development-demand matching score, ranging from 0 to 1. A higher score indicates a better alignment between the development trend and the demand evolution path.

[0068] The degree of alignment between development needs and performance requirements is added as a new weighting factor to the existing multi-dimensional scoring. A weight value for each development dimension is preset based on the nature of the position. For example, for core positions requiring long-term development, the weight can be set to 25% to 30%; for short-term execution positions, the weight can be appropriately reduced to 10%. The updated comprehensive score is calculated according to the new weight. The formula is: the new comprehensive score equals the original comprehensive score multiplied by one minus the development weight, plus the degree of alignment between development needs and performance requirements multiplied by the development weight. Therefore, the degree of alignment between development needs and performance requirements becomes a crucial factor influencing the final ranking.

[0069] Based on the matching score and specific findings during the alignment process, structured temporal matching labels are automatically generated. For example, if the matching score is higher than 0.85 and the synchronization between stages is good, it is marked as highly synchronous development; if the matching score is moderate but the candidate's development is significantly faster than the requirements, it is marked as advanced development; if the matching score is moderate but the candidate's development is slower than the requirements, it is marked as lagging development; if the matching path is unique but can meet the requirements through cross-disciplinary experience, it is marked as cross-disciplinary adaptation. A large language model is invoked to generate explanatory text in natural language based on the detailed results of the alignment process, describing how the candidate's development trend matches the evolution path of the job requirements, specifically identifying which stages show a high degree of fit, which stages may have risks or gaps, and the potential impact of this matching on future work performance.

[0070] Candidates are ranked from highest to lowest based on their overall scores. The top-scoring candidates are selected to form the final recommendation list. The final recommendation list is then compared and analyzed with the corresponding standardized talent files, personalized recommendation reasons, and differentiated advantages. Based on the actual needs of talent introduction, a standardized talent recommendation report is generated. The report includes the core recommendation list, detailed candidate files across all dimensions, personalized recommendation reasons, and talent introduction suitability suggestions.

[0071] By coordinating the execution order and data flow of each intelligent agent unit through intelligent agent management, and recording the entire process information, talent and technology information traceability is provided.

[0072] Specifically, after the multi-dimensional evaluation agent outputs a standardized talent recommendation report to the collaborative management agent, the information collaboration module sends the report to the system's interactive interface, allowing users to view, download, filter, and precisely search for candidate information online. Simultaneously, the collaborative management agent aggregates all processing information from the entire talent recommendation process into an information traceability database, forming a complete and traceable processing record for subsequent tracing, verification, and secondary evaluation. If there is a need for further adjustments to the recommendation results, feedback can be provided instantly through the system interface. Based on this feedback, the collaborative management agent will quickly reschedule the corresponding agent for further precise processing, adjusting priorities and weights until a talent recommendation result that meets the actual talent recruitment needs is output.

[0073] In summary, the science and technology talent recommendation method based on multi-agent collaboration provided in this application has the following beneficial effects: By analyzing user talent needs information through a demand analysis agent across multiple dimensions and extracting demand features according to each dimension, a user demand profile is constructed. Using this profile as input constraints, a retrieval and discovery agent performs parallel retrieval and fusion of multi-source heterogeneous talent databases to generate a preliminary talent candidate list. An information enhancement agent supplements and verifies the candidate information in this preliminary list, generating standardized talent profiles. A multi-dimensional evaluation agent performs multi-dimensional quantitative evaluation of candidates in the standardized talent profiles based on the user demand profile, calculates a comprehensive score, and generates an interpretable recommendation report. A collaborative management agent schedules the execution order and data flow of each agent unit and records the entire process information for talent and technology information traceability. In other words, through demand analysis, retrieval and discovery, information enhancement, multi-dimensional evaluation, and collaborative management agents, an end-to-end intelligent processing flow from demand analysis to talent recommendation is achieved. This process employs a hierarchical and progressive processing logic, with each agent cooperating through a collaborative mechanism to jointly complete talent discovery, evaluation, and recommendation tasks, improving the comprehensiveness of talent information integration and thus increasing the efficiency of technology talent recommendation.

[0074] Example 2: Based on the same inventive concept as the multi-agent collaborative technology talent recommendation method in Example 1, this application also provides a multi-agent collaborative technology talent recommendation system. Please refer to the appendix. Figure 2 The aforementioned science and technology talent recommendation system based on multi-agent collaboration includes: The interactive demand feature extraction module 11 is used to perform multi-dimensional analysis based on user talent demand information through a demand analysis agent, extract demand features according to each dimension, and construct a user demand profile. The parallel retrieval and fusion module 12 is used to take the user demand profile as input constraints, perform parallel retrieval and fusion of multi-source heterogeneous talent databases through a retrieval discovery agent, and generate a preliminary talent candidate list. The supplementary verification module 13 is used to supplement and verify the candidate information in the preliminary talent candidate list through an information enhancement agent, and generate standardized talent profiles. The multi-dimensional quantitative evaluation module 14 is used to perform multi-dimensional quantitative evaluation of candidates in the standardized talent profiles based on the user demand profile through a multi-dimensional evaluation agent, calculate a comprehensive score, and generate an interpretable recommendation report. The multi-agent collaborative management module 15 is used to schedule the execution order and data flow of each agent unit through a collaborative management agent, and record the entire process information to provide talent and technology information traceability.

[0075] Furthermore, the interaction requirement feature extraction module 11 in the aforementioned multi-agent collaborative technology talent recommendation system is also used for: performing multi-dimensional requirement analysis on the input requirement text information through a semantic analysis module to identify explicit requirement information and implicit requirement information; extracting attributes from the explicit and implicit requirement information respectively, mapping the extracted attributes and attribute values ​​to a preset structured requirement profile field library to generate a multi-dimensional structured requirement profile; configuring priority weights for each dimension parameter using a weight allocation module, and adding the weights allocated to each dimension parameter to the multi-dimensional structured requirement profile to form a weighted user requirement profile.

[0076] Furthermore, the interaction demand feature extraction module 11 in the aforementioned multi-agent collaborative science and technology talent recommendation system is also used for: preprocessing the demand text information, extracting multi-dimensional descriptions of talent demand rules from the text information through the demand rule semantic analysis module, obtaining multi-dimensional explicit demand information, including science and technology fields, academic achievement indicators, project experience parameters, and job descriptions; and performing reasoning analysis based on the explicit demand information to identify implicit demand information with explicit correlations in each dimension.

[0077] Furthermore, the interaction demand feature extraction module 11 in the multi-agent collaborative science and technology talent recommendation system is also used to: perform influence relationship analysis on the explicit demand information and implicit demand information of each dimension of demand parameters to determine the strength of the influence relationship on the demand for science and technology talents; configure the priority weight of each dimension of demand parameters according to the strength of the influence relationship, and perform weight labeling on the multi-dimensional structured demand profile to obtain a weighted user demand profile.

[0078] Furthermore, the parallel retrieval and fusion module 12 in the aforementioned multi-agent collaborative science and technology talent recommendation system is also used for: performing semantic similarity retrieval in the internal talent pool using vector embedding technology and approximate nearest neighbor algorithm through the internal talent pool retrieval module; retrieving high-impact achievements and extracting personnel information from academic databases, patent databases, and project databases based on domain keywords through the external achievement database discovery module; searching for qualified science and technology talent information from open websites through the large model enhancement search module; and using the multi-source result fusion module to perform confidence assessment and deduplication on the retrieval results of each module to generate a preliminary talent candidate list.

[0079] Furthermore, the supplementary verification module 13 in the aforementioned multi-agent collaborative science and technology talent recommendation system is also used for: determining demand parameters and priority weights based on user demand profiles; determining the missing information of each candidate in the preliminary talent candidate list based on the demand parameters, and performing enhanced search by calling a large model based on the missing demand parameters to retrieve publicly available talent information of the candidates; classifying and integrating the retrieved talent information according to a preset standardized talent profile template to obtain preliminary talent profile information; verifying the authenticity and consistency of the supplemented preliminary talent profile information using a multi-source cross-validation mechanism; confirming the verification results based on priority weights, and generating the standardized talent profile using the confirmed preliminary talent profile information.

[0080] Furthermore, the supplementary verification module 13 in the aforementioned multi-agent collaborative science and technology talent recommendation system is also used to: verify the authenticity of paper publication information through journal paper databases and conference paper databases to obtain paper verification results; verify the authenticity of project experience through research project databases; verify the consistency of research direction through the official website of the candidate's institution, the introduction of the research group, and the citation information of peers' academic achievements to obtain academic achievement verification results; verify the timeline logic consistency of the candidate's research experience and team management experience through information from multiple sources to obtain science and technology experience verification results; integrate the successful verification results and the questionable verification results from all verification results, perform verification annotations, and send the questionable verification results to the collaborative verification channel for manual verification intervention to generate the final verification result.

[0081] Furthermore, the multi-dimensional quantitative evaluation module 14 in the aforementioned multi-agent collaborative science and technology talent recommendation system is also used for: establishing a demand mapping relationship between the standardized talent profile and the user demand profile; evaluating each dimension of demand based on the demand mapping relationship, according to the demand response relationship and priority weight, and generating a comprehensive score; ranking candidates according to the comprehensive score, and calling a large language model to analyze and compare the demand parameters of each candidate, generating an interpretable recommendation report for comparing the explicit and implicit matching differences of candidates and the reasons for recommendation.

[0082] Furthermore, the multi-dimensional quantitative evaluation module 14 in the aforementioned multi-agent collaborative science and technology talent recommendation system is also used for: analyzing the development needs of users' job skills; conducting interactive analysis of the developmental needs characteristics of explicit and implicit needs information in the user's needs profile; establishing a needs development timeline chain; inputting the candidate's standardized talent file and the evaluation results of each dimension into a large language model; understanding the time evolution logic of science and technology talent needs goals through the large language model; extracting the time sequence of key events related to time in the candidate's experience and achievements; analyzing the candidate's knowledge structure update speed, research direction evolution trajectory, and responsiveness to new technology fields; predicting the candidate's science and technology development evolution trend; matching the candidate's science and technology development evolution trend with the needs development timeline chain based on a timeline alignment algorithm; identifying the development needs matching degree; adding the development needs matching degree as a weighted item to the multi-dimensional score to obtain a comprehensive score; and outputting structured matching labels and explanatory text.

[0083] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The method and specific example of a technology talent recommendation system based on multi-agent collaboration in the foregoing embodiment one are also applicable to the technology talent recommendation system based on multi-agent collaboration in this embodiment. Through the foregoing detailed description of a technology talent recommendation method based on multi-agent collaboration, those skilled in the art can clearly understand the technology talent recommendation system based on multi-agent collaboration in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0084] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0085] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for recommending scientific and technological talents based on multi-agent collaboration, characterized in that, include: Based on user talent demand information, the demand analysis agent performs multi-dimensional analysis, extracts demand features according to each dimension, and constructs a user demand profile. Using the user demand profile as input constraints, the search and discovery agent performs parallel retrieval and fusion of multi-source heterogeneous talent pools to generate a preliminary talent candidate list. The information-enhanced intelligent agent supplements and verifies the candidate information in the preliminary talent candidate list to generate standardized talent profiles; The multidimensional evaluation agent performs multidimensional quantitative evaluation of candidates in the standardized talent files based on user demand profiles, calculates comprehensive scores, and generates interpretable recommendation reports. By coordinating the execution order and data flow of each intelligent agent unit through intelligent agent management, and recording the entire process information, talent and technology information traceability is provided.

2. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 1, characterized in that, The construction of user demand profiles includes: The semantic analysis module performs multi-dimensional requirement analysis on the input requirement text information to identify explicit and implicit requirement information. The explicit and implicit demand information are extracted separately, and the extracted attributes and attribute values ​​are mapped to a preset structured demand profile field library to generate a multi-dimensional structured demand profile. The weight allocation module is used to configure priority weights for parameters of each dimension, and the weights allocated to the parameters of each dimension are added to the multi-dimensional structured demand profile to form a weighted user demand profile.

3. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 2, characterized in that, Identify explicit and implicit requirements, including: The textual information of the requirements is preprocessed, and the multi-dimensional description of talent demand rules in the textual information is extracted through the semantic analysis module of the requirements rules to obtain multi-dimensional explicit demand information, including the field of science and technology, academic achievement indicators, project experience parameters, and job descriptions. Based on the explicit demand information, inference analysis is performed to identify implicit demand information with explicit correlations in each dimension.

4. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 3, characterized in that, The weight allocation module is used to configure priority weights for parameters of each dimension, and the weights assigned to each dimension parameter are added to the multi-dimensional structured requirement profile to form a weighted user requirement profile, including: An impact analysis of the explicit and implicit demand information on the demand parameters of each dimension was conducted to determine the strength of the impact on the demand for scientific and technological talents. Priority weights for each dimension of demand parameters are configured according to the strength of the influence relationship, and the multi-dimensional structured demand profile is weighted to obtain a weighted user demand profile.

5. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 1, characterized in that, Generate a preliminary talent candidate list, including: The internal talent pool retrieval module employs vector embedding technology and an approximate nearest neighbor algorithm to perform semantic similarity retrieval within the internal talent pool. The module discovers high-impact achievements and extracts personnel information from academic databases, patent databases, and project databases based on domain keywords. The search module is enhanced by using a large model to search for information on qualified science and technology talents from open websites; The multi-source result fusion module is used to evaluate the confidence level and remove duplicates from the search results of each module, generating a preliminary list of talent candidates.

6. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 1, characterized in that, Generate standardized talent profiles, including: Based on user demand profiles, determine demand parameters and priority weights; Based on the demand parameters, the information of each candidate in the preliminary talent candidate list is determined to be missing, and the large model is called to perform enhanced search based on the missing demand parameters to retrieve the publicly available talent information of the candidates. According to the pre-set standardized talent file template, the retrieved talent information is classified and integrated to obtain preliminary talent file information; A multi-source cross-validation mechanism was used to verify the authenticity and consistency of the supplemented preliminary talent file information; The verification results are confirmed based on priority weights, and the standardized talent file is generated using the preliminary talent file information that has passed the verification.

7. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 6, characterized in that, include: The authenticity of paper publication information is verified through journal and conference paper databases to obtain paper verification results; The authenticity of the project experience is verified by the research project database, and the consistency of the research direction is verified by the official website of the candidate's institution, the introduction of the research group, and the citation information of peers' academic achievements, so as to obtain the academic achievement verification results. By verifying the timeline consistency of candidates' research experience and team management experience through information from multiple sources, we can obtain verification results of their scientific and technological experience. Integrate all successful and questionable verification results, add verification annotations, and send the questionable verification results to the collaborative verification channel for manual verification intervention to generate the final verification result.

8. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 4, characterized in that, A multi-dimensional evaluation agent performs a quantitative assessment of candidates in the standardized talent profile based on user needs profiles, calculates a comprehensive score, and generates an interpretable recommendation report, including: Establish a demand mapping relationship between the standardized talent files and user demand profiles; Based on the aforementioned demand mapping relationship, demand evaluation is performed in each dimension according to the demand response relationship and priority weight, and a comprehensive score is generated. Candidates are ranked based on their comprehensive scores, and a large language model is used to analyze and compare the requirements parameters of each candidate, generating an interpretable recommendation report to compare the explicit and implicit matching differences of the candidates and the reasons for the recommendation.

9. The method for recommending scientific and technological talents based on multi-agent collaboration according to claim 8, characterized in that, The evaluation process involves a multi-dimensional assessment agent that performs a quantitative evaluation of candidates in the standardized talent profile based on user needs profiles, calculating a comprehensive score, and also includes: Analyze the development needs of users' job skills, conduct interactive analysis of the developmental needs characteristics of explicit and implicit needs information in the user needs profile, and establish a timeline chain of needs development. The standardized talent profiles of candidates and the evaluation results of each dimension are input into the big language model. The big language model is used to understand the time evolution logic of the demand goals for scientific and technological talents, extract the time sequence of key events related to time in the candidate's experience and achievements, analyze the updating speed of the candidate's knowledge structure, the evolution trajectory of the research direction, and the responsiveness to new technology fields, and predict the candidate's scientific and technological development trend. Based on the time-series alignment algorithm, the technological development evolution trend of candidates is matched with the time-series chain of demand development to identify the degree of matching of development needs. The degree of matching the development needs is added as a weighting factor to the multi-dimensional score to obtain a comprehensive score, and structured time-series matching tags and explanatory text are output.

10. A technology talent recommendation system based on multi-agent collaboration, characterized in that, The steps for implementing the method for recommending scientific and technological talents based on multi-agent collaboration as described in any one of claims 1 to 9, wherein the system for recommending scientific and technological talents based on multi-agent collaboration comprises: The interaction requirement feature extraction module is used to analyze user talent requirement information from multiple dimensions through a requirement analysis intelligent agent, extract requirement features according to each dimension, and build a user requirement profile. The parallel retrieval and fusion module is used to take the user demand profile as input constraints, and perform parallel retrieval and fusion of multi-source heterogeneous talent pools through the retrieval and discovery agent to generate a preliminary talent candidate list. The supplementary verification module is used to supplement and verify the candidate information in the preliminary talent candidate list through an information-enhanced intelligent agent, and generate standardized talent profiles. The multi-dimensional quantitative evaluation module is used to conduct multi-dimensional quantitative evaluation of candidates in the standardized talent file based on user demand profiles through a multi-dimensional evaluation intelligent agent, calculate a comprehensive score, and generate an interpretable recommendation report. The multi-agent collaborative management module is used to schedule the execution order and data flow of each agent unit through collaborative management agents, and record the entire process information to provide talent and technology information traceability.

Images