Intelligent generation method and system for talent portrait report based on large model

Abstract

The present application relates to the technical field of data processing, in particular to a talent portrait report intelligent generation method and system based on a large model. The method classifies and clusters the original single operation records according to the second micro-service node identifier and the timestamp to obtain a log event cluster; for each combination of the resume project record and the log event cluster, a single item matching deviation value is generated; a bipartite graph matching is performed to obtain the optimal matching data combination of each resume project record; the single item matching deviation value corresponding to the optimal matching is taken as an abnormal index to perform abnormal record screening to obtain a verified resume project; and the verified resume project is sent to a large language model after text splicing in chronological order to generate a portrait report. The present application can effectively identify abnormal resumes lacking underlying log evidence and improve the objectivity of talent portrait generation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for intelligent generation of talent profile reports based on large models. Background Technology

[0002] In the field of enterprise human resource management and technical talent assessment, using Retrieval-Augmented Generation (RAG) technology combined with large language models to generate candidate resume summaries and profile reports has become a standard practice. This technology typically stores candidate resume text along with internal collaborative operation logs (such as code submissions and task check-ins) in a vector database. It then retrieves relevant logs by calculating text similarity, concatenates them, and inputs the results into a large language model for summarization.

[0003] Existing search-enhanced generation systems face an objective data dilution problem when processing both macro-level resumes and micro-level logs—two types of heterogeneous data across different granularities. A single macro-level project declaration in a candidate's resume (such as a six-month business module restructuring) is, in reality, composed of tens of thousands of scattered micro-level logs. When the search system performs arithmetic averaging or simple aggregation based on text similarity, if a candidate's resume declaration contains time misalignments, module packaging, or even false statements, the resulting high-biased value will be diluted by the massive amount of normal daily operation logs in the graph database, which will be used as the denominator in the calculation. This numerical dilution caused by data granularity imbalance makes it impossible for the search system to detect isolated resume records lacking underlying factual basis.

[0004] Furthermore, existing retrieval methods typically rely solely on simple text semantic comparison, neglecting the objective dependencies of enterprise IT architecture. Traditional slicing and splicing mechanisms often fail to maintain the irreversible timeline of business operations when inputting log fragments into large language models. This results in a lack of temporal monotonicity constraints during autoregressive decoding, leading to illusions or the generation of false profiles that severely deviate from actual business evolution patterns. Summary of the Invention

[0005] To address the technical problem that existing technologies cannot effectively identify and filter abnormal resumes lacking underlying log data support, thus generating illusionary or unrealistic talent profile content, the present invention aims to provide a method and system for intelligently generating talent profile reports based on a large model. The specific technical solution adopted is as follows:

[0006] This invention proposes an intelligent generation method for talent profile reports based on a large model, the method comprising:

[0007] Extract the first feature data from each resume project record of the target candidate. The first feature data includes a semantic feature vector, the time interval, and the first microservice node identifier.

[0008] Extract the second feature data from all original single operation records of the target candidate. The second feature data includes a timestamp, a temporary text vector, and a second microservice node identifier. Then, perform secondary classification on the original single operation records according to the second microservice node identifier and the timestamp to obtain multiple log event clusters.

[0009] For any combination of resume project record and any log event cluster, compare the correlation features between the first feature data and the second feature data to obtain the single-item matching deviation value of the data combination; perform bipartite graph matching with resume project record and log event cluster as left and right nodes respectively, and use the edge weight between nodes as the single-item matching deviation value to obtain the optimal matching data combination;

[0010] For all resume project records, the single-item matching deviation value corresponding to the best matching data combination is used as an anomaly indicator to filter out abnormal records and obtain verified resume projects.

[0011] Based on the verified resume items, text is concatenated and profiles are generated in chronological order.

[0012] Furthermore, the method for secondary classification includes:

[0013] The original single operation records with the same second microservice node identifier are grouped into one category to obtain a first-level classification cluster; a time period is set in the time series range, and the original single operation records in the same time period in the first-level classification cluster are grouped into one category to obtain a second-level classification cluster; the second-level classification cluster is used as the log event cluster.

[0014] Furthermore, the method for obtaining the individual matching deviation value includes:

[0015] For each log event cluster, the average temporary text vector within the cluster is used as the overall temporary text vector, and the average timestamp is used as the center timestamp.

[0016] The single-item matching deviation value is obtained by comparing the vector difference between the semantic feature vector and the overall temporary text vector, comparing the time distance between the time interval and the center timestamp, and comparing the module call levels of the first microservice node identifier and the second microservice node identifier in the system.

[0017] Furthermore, obtaining the individual matching deviation value based on the three comparison results includes:

[0018] The vector difference is the normalized Euclidean distance;

[0019] If the center timestamp is within the time interval, the time distance is set to 0; otherwise, the nearest time distance between the center timestamp and the time interval is taken as the time distance.

[0020] Count the shortest interface call levels for business requests between any two microservice nodes and construct a hierarchical database; determine the module call levels of the first microservice node identifier and the second microservice node identifier in the system within the hierarchical database;

[0021] The individual matching deviation value is obtained by weighted summation of vector differences, time distance, and module call hierarchy.

[0022] Furthermore, after obtaining the log event cluster, it also includes:

[0023] If the number of log event clusters is less than the number of resume project records, the abnormal indicator for each resume project record is set to the preset maximum value; otherwise, the single-item matching deviation value and matching steps are executed to obtain the abnormal indicator.

[0024] Furthermore, the bipartite graph matching algorithm selected is the KM matching algorithm.

[0025] Furthermore, the screening method for the verified resume projects includes:

[0026] If the abnormal indicator is greater than the preset deviation tolerance threshold, the corresponding resume item will be screened out as an abnormal item.

[0027] Furthermore, the method for setting the deviation tolerance threshold includes:

[0028] In the historical evaluation database of former employees of the company, a set of individual matching deviation values ​​for each verified resume item is compiled, and data at a preset percentile position is selected from the set of individual matching deviation values ​​as the deviation tolerance threshold.

[0029] Furthermore, the method for generating the image includes:

[0030] The project text, which is concatenated with the profile generation requirement string, is input into a pre-trained large language model, and the output reasoning text string is used as a talent profile report.

[0031] This invention also proposes an intelligent generation system for talent profile reports based on a large model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the steps of the intelligent generation method for talent profile reports based on a large model.

[0032] The present invention has the following beneficial effects:

[0033] This invention extracts the second feature data from the original single operation record, and aggregates the scattered logs into log event clusters based on the second microservice node identifier and timestamp. This effectively elevates the fragmented logs at the micro-granularity level to an equivalent granularity level comparable to macro-level resume items, eliminating the dimensional barriers when directly comparing cross-granularity data and laying a structured foundation for subsequent deviation quantification. Furthermore, for each pair of resume item records and log event clusters, the association features between the first and second feature data are comprehensively compared from three dimensions: semantics, time, and microservice nodes, generating a single-item matching deviation value. This allows the authenticity of the resume declaration to be objectively quantified from multiple dimensions, avoiding the one-sidedness and unreliability of single-dimensional evaluation. A bipartite graph is constructed using the resume item records and log event clusters as left and right nodes, respectively. Bipartite graph matching is performed using the single-item matching deviation value as the edge weight. A strict one-to-one mapping constraint is applied during the matching process, ensuring that each resume item record can only match one unique log event cluster. Under this exclusive allocation mechanism, genuine resumes with underlying factual support can normally occupy the corresponding log event clusters. However, abnormal resumes lacking evidence cannot reuse the already occupied legitimate event clusters and are forced to be bound to unrelated event clusters with vastly different spatiotemporal characteristics and microservice nodes. Their deviation is exposed in the form of a single-point surge, completely severing the algorithmic path of diluting abnormal values ​​by massive amounts of normal logs, and achieving accurate quantification and exposure of abnormal resumes. Therefore, the single-item matching deviation value corresponding to the optimal matching data combination can be used as an anomaly indicator to filter out abnormal records exceeding the threshold to obtain verified resume items, effectively filtering out the illusion-inducing sources lacking underlying factual support. Furthermore, the retained resume items are concatenated into a profile after being arranged in chronological order, ensuring that the input context received by the large language model strictly follows the temporal monotonicity of business evolution, fundamentally avoiding causal reversal or content illusions caused by temporal disorder. This invention significantly improves the objectivity, authenticity, and temporal rationality of talent profile reports generated by large language models through closed-loop techniques such as secondary classification dimensionality reduction, multi-dimensional feature comparison, bipartite graph exclusive matching, and temporal constraint splicing. Attached Figure Description

[0034] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1This is a flowchart of a method for intelligently generating talent profile reports based on a large model, provided as an embodiment of the present invention. Detailed Implementation

[0036] This invention addresses a modern enterprise-level IT system architecture, in which microservice architecture is widely adopted. This architecture breaks down a single large business system into multiple independently deployed and running lightweight service units according to domain boundaries. Each service unit focuses on a specific business function domain and interacts across services through lightweight communication protocols (such as HTTP / RESTful API or RPC). In this architecture, the microservice node refers to each independent service instance after the breakdown, possessing its own process space, code repository, and database resources, serving as the actual carrier of enterprise business logic. For example, in an e-commerce business system, it typically includes multiple decoupled microservice nodes such as an authentication service node responsible for user authentication and permission verification, a product service node responsible for product information management and listing, an order service node responsible for transaction order generation and status transition, and an inventory service node responsible for inventory deduction and rollback. The microservice registration name refers to the logically unique identifier that each microservice node registers with a service registry (such as Nacos, Eureka, or Consul) during the startup initialization phase, based on the enterprise's unified service governance specifications. The names are typically encoded using strings with clear business semantics. For example, "user-auth-service" represents the user authentication microservice, "order-manage-service" represents the order management microservice, and "inventory-deduct-service" represents the inventory deduction microservice. The service registry maintains the mapping between the registered names and network addresses of all online microservice nodes, enabling service nodes to address and invoke each other using fixed registered names, without needing to consider the fluctuating IP addresses of specific deployment instances. This set of registered names constitutes a globally unified namespace within the enterprise IT architecture, providing a basic reference system for measuring the topology of cross-service calls.

[0037] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent generation method and system for talent profile reports based on a large model provided by this invention.

[0038] Please see Figure 1 The diagram illustrates a flowchart of a method for intelligently generating talent profile reports based on a large model, according to an embodiment of the present invention. The method includes:

[0039] Step S1: Extract the first feature data from each resume project record of the target candidate. The first feature data includes the semantic feature vector, the time interval, and the identifier of the first microservice node.

[0040] In actual human resources management systems, candidates submit long resumes containing project descriptions spanning different time periods. Therefore, step S1 first requires breaking down the various contents of the long resumes of the target candidates, identifying each resume project record, and further extracting the semantic feature vectors, time intervals, and determining the first microservice node identifier based on the proper nouns contained in the resume project records. These data are then integrated, encapsulated, and stored as the first feature data of each resume project record.

[0041] It should be noted that, in this embodiment of the invention, a long resume text can be divided into multiple independent text paragraphs based on paragraph separators in the established text, with each text paragraph representing a resume item record. This is a simple application of technology in the field of natural language processing; it can also be divided by recognizing separator words, punctuation marks, etc., which will not be elaborated or limited here.

[0042] This invention can pre-obtain the configuration table of the enterprise IT architecture's tracing system. It reads all the independent microservice node names registered in the configuration table, removes duplicates from the collected names to obtain a set of microservice registration names. Each microservice registration name in the set can be assigned a unique identifier. Then, a string matching search is performed between the proper nouns in the resume project record and each microservice registration name in the set. The identifier corresponding to the found microservice registration name is used as the first microservice node identifier. Specifically, if the string matching search fails, and the proper noun refers to a clearly defined system module, such as pure algorithm research or non-R&D management work, a preset null identifier can be assigned to that resume project record. This null identifier ensures that the first feature data has three dimensions, avoiding calculation errors in subsequent steps.

[0043] In one specific implementation of this invention, the text content of each resume entry can be input into a pre-trained text vectorization model. This model encodes the contextual semantics of the sentences and outputs a high-dimensional floating-point array representing the semantic features of the paragraph, serving as the semantic feature vector. The text vectorization model can be a sentence encoding and extraction model based on the Transformer architecture. This model is a well-known and commonly used model in the field of natural language processing, and those skilled in the art can directly select it; further details are omitted.

[0044] In one specific implementation of this invention, the Named Entity Recognition (NER) algorithm can be used to scan the current resume project record, extract the natural language time description text contained therein, and then convert the time description text into a standard machine timestamp to form the time interval of the project. That is, the time interval specifically includes a start timestamp and an end timestamp.

[0045] Step S2: Extract the second feature data from all original single operation records of the target candidate. The second feature data includes timestamp, temporary text vector and second microservice node identifier. Then, perform secondary classification on the original single operation records according to the second microservice node identifier and timestamp to obtain multiple log event clusters.

[0046] The underlying code commit or collaborative check-in logs are characterized by extremely high frequency and fragmentation. A macro-level resume project lasting several months often corresponds to thousands of micro-level single-operation logs. Without dimensionality reduction and alignment of the underlying data, subsequent one-to-one comparisons will lose their statistical basis. Therefore, resume project records cannot be directly matched with original single-operation records. It is necessary to first classify and merge the discrete original single-operation records to achieve data dimensionality reduction, elevating the fragmented logs at the micro-granularity level to an equivalent granularity level comparable to the macro-level resume project, thus eliminating the dimensionality barrier when directly comparing cross-granularity data.

[0047] Similar to the resume project record, the first step is to extract the second feature data from each original single operation record. This second feature data includes a timestamp, a temporary text vector, and a second microservice node identifier. The timestamp can be determined by directly querying the occurrence time of the original single operation record. The temporary text vector and semantic feature vector are quantized in the same way; both can be converted into vector form based on existing natural language processing algorithms, using the text annotations attached to the original single operation record. Similarly, the second microservice node identifier can be obtained by extracting the name of the code repository modified in the original single operation record and matching it with the registration names of each microservice in the microservice registration name set.

[0048] After obtaining the second feature data, the original single operation records can be classified and merged based on the second feature data. This embodiment of the invention sets up a two-level classification process, sequentially using the second microservice node identifier and timestamp as classification criteria to obtain the final classification result. That is, within a log event cluster, there should be a set of original single operation records with the same second microservice node identifier and similar timestamps. By limiting the second microservice node identifier, the logs within the cluster are ensured to be highly cohesive in the business space; simultaneously, by limiting the timestamp, the logs within the cluster are ensured to be continuous and consistent in the business evolution cycle. After dual-constraint clustering, the originally randomly scattered micro logs on the timeline and disorganized in terms of module affiliation are aggregated into a macro event set with the same information span and business scope as the resume project. This unifies the benchmark for cross-granularity comparison from the data source, providing statistically comparable objects for subsequent one-to-one mapping.

[0049] Preferably, in some implementations of the embodiments of the present invention, the secondary classification specifically includes:

[0050] Original single operation records with the same second microservice node identifier are grouped into one category to obtain a first-level classification cluster. A time period is set within the time series range. Original single operation records within the same time period in the first-level classification cluster are grouped into another category to obtain a second-level classification cluster. This second-level classification cluster serves as the log event cluster. In a specific implementation of this invention, the length of the time period is set to one month, meaning the time series range is divided in months. If the timestamps of two original single operation records are in the same month, they are considered to belong to the same category.

[0051] Preferably, in some implementations of the embodiments of the present invention, considering that the embodiments of the present invention aim to determine the credibility of resume project records by matching resume project records with log event clusters, and if the number of log event clusters is less than the number of resume project records, it means that the total number of macro-level business projects claimed by the target candidate in the resume may even exceed the total number of macro-level operation events stored in the underlying system after clustering. This quantitative relationship directly violates the objective laws of normal R&D activities and constitutes serious resume fraud. Therefore, after obtaining multiple log event clusters, the number can be directly compared with the resume project records to establish an extreme penalty mechanism, specifically including:

[0052] If the number of log event clusters is less than the number of resume project records, the abnormal indicator for each resume project record is set to the preset maximum value; otherwise, the single-item matching deviation value and matching steps are executed to obtain the abnormal indicator.

[0053] It should be noted that the purpose of the preset maximum value is to set a maximum value so that the verification process in the subsequent step S4 will directly give a failure result. Therefore, it can be directly set to three times the deviation from the tolerance threshold, or it can be marked as an abnormal record in step S4.

[0054] Step S3: For any combination of resume project records and any log event clusters, compare the correlation features between the first feature data and the second feature data to obtain the single-item matching deviation value of the data combination; perform bipartite graph matching with resume project records and log event clusters as left and right nodes respectively, and use the edge weights between nodes as the single-item matching deviation value to obtain the optimal matching data combination.

[0055] After classifying the original single operation records in step S2, dimensionality reduction of discrete data is achieved. At this point, any resume item record can be combined with any log event cluster to obtain multiple data combinations. In this embodiment of the invention, (i, j) can be used to represent the data combination of the i-th resume item record and the j-th log event cluster.

[0056] For any given data combination, the correlation features between the first and second features can be compared. This involves a comprehensive comparison across three dimensions: semantics, time, and microservice nodes, to determine the individual match deviation value between resume project records and log event clusters. This feature allows the authenticity of resume claims to be objectively quantified from multiple dimensions, avoiding the bias and unreliability of single-dimensional evaluation. The larger the individual match deviation value, the greater the overall disconnect between resume project records and log event clusters, and the weaker the match between them.

[0057] Because the single-item matching deviation value can represent the matching cost of data combinations, a bipartite graph matching method can be used. Resume project records and log event clusters are treated as left and right nodes respectively, and the edge weights between nodes are set as the single-item matching deviation value. Optimal matching data combinations can be obtained through matching optimization. This optimal matching data combination is a one-to-one mapping relationship, with each resume project record corresponding to a data combination result. Under this exclusive allocation mechanism, genuine resumes with underlying factual support can normally occupy the corresponding log event clusters. Abnormal resumes lacking evidence, unable to reuse already occupied legitimate event clusters, are forced to bind to unrelated event clusters with vastly different spatiotemporal characteristics and microservice node differences. Their deviation is exposed in the form of a single-point surge, completely cutting off the computational path where abnormal values ​​are diluted by massive amounts of normal logs. This achieves accurate quantification and exposure of abnormal resumes. The larger the single-item matching deviation value of the final optimal matching data combination, the more likely the corresponding resume project record is an abnormal resume lacking evidence. Such records will be deleted in subsequent verification steps and will not be used as a reference for profiling.

[0058] Preferably, in some implementations of the embodiments of the present invention, each dimension of the first feature data and the second feature data is compared separately, and then a comprehensive analysis is performed to obtain a single-item matching deviation value, specifically including:

[0059] First, considering that a log event cluster is a collection of original single operation records, containing multiple sets of second feature data, where the second microservice node identifier is the same, but the data in the other two dimensions may differ, it is necessary to perform statistical quantification separately. For each log event cluster, the average temporary text vector within the cluster can be used as the overall temporary text vector, and the average timestamp can be used as the center timestamp. This enables the statistical analysis of the overall data.

[0060] It should be noted that all temporary text vectors are vectorized using the same pre-trained vectorization model, therefore their dimensions are uniform, and the overall temporary text vector can be obtained directly by averaging. The average timestamp can be calculated directly by averaging the absolute timestamps over the time series to obtain the center timestamp.

[0061] After unifying the data as described above, for any data combination, the vector difference between the semantic feature vector and the overall temporary text vector can be compared. The larger the vector difference, the more severe the semantic deviation in the business content description between the two types of data in the data combination. The time distance between the stated time interval and the central timestamp can also be compared. A larger time distance indicates a greater gap between the operation time recorded in the system logs and the development time claimed in the resume, and a greater mismatch between the two. Finally, the module call levels of the first and second microservice node identifiers in the system can be compared. A larger difference in this level comparison indicates a more distant relationship between the actually modified code module in the logs and the core module claimed in the resume in terms of business path. The single-item matching deviation value is then obtained based on these three comparison results.

[0062] Furthermore, the individual matching deviation value is obtained based on three comparison results, including:

[0063] The vector difference is the normalized Euclidean distance. It should be noted that because the semantic feature vector and the overall temporary text vector have the same dimension, a distance scalar can be directly obtained using the Euclidean distance calculation formula. In a specific implementation of this invention, to facilitate the quantification of individual matching deviation values, the obtained Euclidean distance can be further normalized. The normalization method can be range standardization. The maximum and minimum values ​​used can be obtained by consulting historical records. Furthermore, the obtained maximum and minimum values ​​can be used as data boundaries; that is, data with values ​​less than the minimum value are directly set to 0, and data with values ​​greater than the maximum value are directly set to 1. The specific range standardization formula is a well-known technique among those skilled in the art and will not be elaborated upon here.

[0064] If the center timestamp is within the given time interval, the time distance is set to 0; otherwise, the nearest time distance between the center timestamp and the given time interval is used as the time distance. It should be noted that the nearest time distance is the distance between the center timestamp and the nearest boundary of the time interval. For example, if the center timestamp is before the given time interval, the distance between the center timestamp and the start timestamp of the time interval is used as the time distance. Similarly, in the specific implementation of this embodiment, normalization can be performed using range standardization, eliminating dimensions while achieving effective quantization.

[0065] We calculate the shortest interface call depth between any two microservice nodes and construct a hierarchical database. This call depth reflects the coupling relationship between business modules. For example, if the front-end interface module directly calls the user authentication module, the call depth is 1; if it needs to be forwarded through a gateway, the depth increases accordingly. The final result is a two-dimensional hierarchical database, which can be stored as a matrix. The element in the nth row and mth column represents the shortest interface call depth between the nth and mth microservice nodes.

[0066] The module call levels of the first microservice node identifier and the second microservice node identifier in the system are determined in the hierarchical database. A larger module call level value indicates a longer correlation path in the actual business logic of the two microservices within the system, and a lower business relevance. In a specific implementation of this invention, the module call level can be further processed using a normalization method. A maximum-value normalization method can be used, where the largest element in the hierarchical database is used as the denominator, and the data to be normalized is used as the numerator to achieve normalization.

[0067] The individual matching deviation value is obtained by weighted summation of vector differences, time distance, and module call hierarchy.

[0068] In one specific implementation of this invention, the semantic weight of vector difference is set to 0.4, the temporal weight of time distance is set to 0.4, and the topological weight of module call level is set to 0.2, meaning the sum of the three weights is 1. Furthermore, the semantic weight and temporal weight are considered to be of equal importance, and both are more important than the topological weight. It should be noted that because the first and second microservice node identifiers may be preset null values, and therefore have no corresponding content in the hierarchical database (since they represent additional content without reference information), the topological weight can be directly set to 0, and the remaining two weights can be set to 0.5.

[0069] Preferably, in some implementations of the embodiments of the present invention, the bipartite graph matching algorithm is selected as the KM matching algorithm.

[0070] It should be noted that traditional KM matching algorithms require the number of nodes on both sides to be equal during execution. However, in the scenario of this embodiment, the number of resume item records and the number of log event clusters may not be equal. Therefore, when executing the KM matching algorithm, the number of nodes on both sides can be ensured to be equal by filling in virtual nodes, i.e., filling in virtual resume item record nodes (because the matching process will only begin if the number of resume item records is less than or equal to the number of log event clusters). The edge weights of the virtual nodes can be set to a preset maximum value. This ensures that when the KM algorithm solves the global minimization objective function, any scheme that matches virtual nodes to real event clusters will incur extremely high costs. Under the premise of satisfying the one-to-one constraint, the algorithm will prioritize assigning real event clusters to real resume nodes. Only after all real nodes have completed matching will the remaining event clusters be forced to be assigned to virtual nodes. Since the matching results of virtual nodes do not participate in subsequent calculations, this filling scheme ensures that the KM algorithm can run without affecting the objectivity of the matching deviation obtained by real resume nodes.

[0071] The final result of the KM algorithm matching is the optimal one-to-one matching data combination for each resume project record.

[0072] Step S4: For all resume item records, use the single-item matching deviation value corresponding to the best matching data combination as an anomaly indicator to filter out abnormal records and obtain verified resume items.

[0073] As described in step S3, for each resume item record, the larger the single-item matching deviation value corresponding to the final optimal matching data combination, the lower the reference value of that resume item record, and the more likely it is to be abnormal resume information, requiring screening. Therefore, the single-item matching deviation value corresponding to the optimal matching data combination can be used as an anomaly indicator to filter out abnormal records and obtain verified resume items. That is, verified resume items are considered valid resume content and can be used to generate talent profile reports.

[0074] It should be noted that after the talent profile report generation command is issued, the management system should immediately stop the data flow of all log event clusters and release the memory resources they occupy. At the same time, all macro-level resume item records belonging to the target candidate in the graph database should be loaded into the working memory to form a list to be processed, and then an anomaly screening operation should be performed to obtain the set of retained records used for profile generation.

[0075] Preferably, in some implementations of the present invention, if the abnormal indicator is greater than a preset deviation tolerance threshold, the corresponding resume item is screened out as an abnormal item. The method for setting the deviation tolerance threshold includes:

[0076] In the historical evaluation database of former employees of the company, a set of individual matching deviation values ​​is compiled for each verified resume item. Data at a preset percentile position is selected from this set of individual matching deviation values ​​as the deviation tolerance threshold. In one specific implementation of this invention, the preset percentile is set to 95%, representing the maximum tolerance limit for deviations in the vast majority of genuine resumes throughout the company's history.

[0077] Step S5: Based on the verified resume items, perform text splicing and profile generation in chronological order.

[0078] Large language models are essentially autoregressive models that perform probabilistic predictions. If unordered text fragments are directly input, the model is prone to generating summaries with reversed timelines or causal relationships. Therefore, this embodiment of the invention further extracts the start timestamps of verified resume items. A conventional array sorting function can be called to perform an ascending sort operation on all record objects in the retained record set based on the size of the start timestamp. This ascending sorting mechanism, in essence, forces the input text sequence to strictly follow the irreversible evolution timeline of the candidate's professional ability growth in reality. After completing the ascending sort, the system extracts the original long text content within each verified resume item record in ascending order of array indices starting from 0. The system uses line breaks as delimiters to concatenate the extracted long text content into a long text of the target candidate's resume with monotonically increasing time characteristics. At this point, because the long text has undergone anomaly removal and time-series rearrangement in the previous steps, its content fragments possess a strict causal logical relationship.

[0079] Preferably, in some implementations of the embodiments of the present invention, the method for generating the portrait includes:

[0080] The project text, concatenated with the user profile generation requirement string, is input into a pre-trained large language model. The output inference text string serves as the talent profile report. The user profile generation requirement string is the string corresponding to the user profile generation command, such as "Please summarize the candidate's core technology evolution and capability profile based on the following experiences." The concatenated project text and the user profile generation requirement string can be combined to generate a complete prompt context. The system establishes a communication connection with the large language model interface and sends this complete prompt context as the input sequence to the large language model. After receiving the prompt, which is subject to strict fact filtering and monotonic temporal constraints, the large language model performs natural language inference and decoding. The system receives the final text string output by the large language model and sends it as the final talent profile report to the front-end interface for display. It should be noted that this large language model is a built-in model in the enterprise management system. This embodiment of the invention does not optimize or change this model; existing modules can be directly used on the basis of the existing system.

[0081] In summary, this invention extracts first feature data for each candidate's resume item record, including a semantic feature vector, the time interval, and a first microservice node identifier; it extracts second feature data for all original single-operation records, including a timestamp, a temporary text vector, and a second microservice node identifier, and performs secondary classification clustering based on the second microservice node identifier and the timestamp to obtain log event clusters; for each combination of resume item record and log event cluster, it compares the correlation features between the first and second feature data to generate a single-item matching deviation value; it performs bipartite graph matching with the resume item record as the left node and the log event cluster as the right node to obtain the optimal matching data combination for each resume item record; it uses the single-item matching deviation value corresponding to the optimal match as an anomaly indicator to perform abnormal record filtering to obtain verified resume items; and it concatenates the verified resume items in chronological order and sends them to a large language model to generate a profile report. This invention can effectively identify abnormal resumes lacking underlying log evidence, improving the objectivity of talent profile generation.

[0082] Based on the same inventive concept, this invention also proposes a large-model-based intelligent generation system for talent profile reports, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the steps of the large-model-based intelligent generation method for talent profile reports.

[0083] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0084] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for intelligently generating talent profile reports based on a large model, characterized in that, The method includes: Extract the first feature data from each resume project record of the target candidate. The first feature data includes a semantic feature vector, the time interval, and the first microservice node identifier. Extract the second feature data from all original single operation records of the target candidate. The second feature data includes a timestamp, a temporary text vector, and a second microservice node identifier. Then, perform secondary classification on the original single operation records according to the second microservice node identifier and the timestamp to obtain multiple log event clusters. For any combination of resume project record and any log event cluster, compare the correlation features between the first feature data and the second feature data to obtain the single-item matching deviation value of the data combination; perform bipartite graph matching with resume project record and log event cluster as left and right nodes respectively, and use the edge weight between nodes as the single-item matching deviation value to obtain the optimal matching data combination; For all resume project records, the single-item matching deviation value corresponding to the best matching data combination is used as an anomaly indicator to filter out abnormal records and obtain verified resume projects. Based on the verified resume items, text is concatenated and profiles are generated in chronological order.

2. The intelligent generation method for talent profile reports based on a large model according to claim 1, characterized in that, The methods for secondary classification include: The original single operation records with the same second microservice node identifier are grouped into one category to obtain a first-level classification cluster; a time period is set in the time series range, and the original single operation records in the same time period in the first-level classification cluster are grouped into one category to obtain a second-level classification cluster; the second-level classification cluster is used as the log event cluster.

3. The intelligent generation method for talent profile reports based on a large model according to claim 1, characterized in that, The method for obtaining the single-item matching deviation value includes: For each log event cluster, the average temporary text vector within the cluster is used as the overall temporary text vector, and the average timestamp is used as the center timestamp. The single-item matching deviation value is obtained by comparing the vector difference between the semantic feature vector and the overall temporary text vector, comparing the time distance between the time interval and the center timestamp, and comparing the module call levels of the first microservice node identifier and the second microservice node identifier in the system.

4. The intelligent generation method for talent profile reports based on a large model according to claim 3, characterized in that, The process of obtaining the single-item matching deviation value based on three comparison results includes: The vector difference is the normalized Euclidean distance; If the center timestamp is within the time interval, the time distance is set to 0; otherwise, the nearest time distance between the center timestamp and the time interval is taken as the time distance. Count the shortest interface call levels of business requests between any two microservice nodes and construct a hierarchical database; determine the module call levels of the first microservice node identifier and the second microservice node identifier in the system in the hierarchical database; The individual matching deviation value is obtained by weighted summation of vector differences, time distance, and module call levels.

5. The intelligent generation method for talent profile reports based on a large model according to claim 1, characterized in that, After obtaining the log event cluster, it also includes: If the number of log event clusters is less than the number of resume project records, then the abnormal indicator for each resume project record is set to the preset maximum value; otherwise, the single-item matching deviation value and matching steps are executed to obtain the abnormal indicator.

6. The intelligent generation method for talent profile reports based on a large model according to claim 1, characterized in that, The bipartite graph matching algorithm selected is the KM matching algorithm.

7. The intelligent generation method for talent profile reports based on a large model according to claim 1, characterized in that, The screening methods for the verified resume projects include: If the abnormal indicator is greater than the preset deviation tolerance threshold, the corresponding resume item will be screened out as an abnormal item.

8. The intelligent generation method for talent profile reports based on a large model according to claim 7, characterized in that, The method for setting the deviation tolerance threshold includes: In the historical evaluation database of former employees of the company, a set of individual matching deviation values ​​for each verified resume item is compiled, and data at a preset percentile position is selected from the set of individual matching deviation values ​​as the deviation tolerance threshold.

9. The intelligent generation method for talent profile reports based on a large model according to claim 1, characterized in that, The method for generating the portrait includes: The project text, which is concatenated with the profile generation requirement string, is input into a pre-trained large language model, and the output reasoning text string is used as a talent profile report.

10. A talent profile report intelligent generation system based on a large model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent generation method for talent profile reports based on a large model as described in any one of claims 1 to 9.

Images