Labor management method based on Chinese semantic analysis and multi-agent collaborative decision-making

Abstract

The present application relates to a labor management method based on Chinese semantic analysis and multi-agent collaborative decision-making, belonging to the technical field of intelligent labor management. It includes: receiving labor demand text, generating demand semantic structure through semantic analysis model; distributing demand semantic structure to basic extraction agent, implicit mining agent, scene recognition agent and benchmark calibration agent, respectively extracting basic field, implicit field, scene field and benchmark initial field, and arbitration agent generating multi-dimensional demand feature vector; inputting multi-dimensional demand feature vector into dynamic demand feature model to generate labor demand benchmark, matching personnel dynamic capability portrait to calculate weighted Euclidean distance to obtain matching parameters, and generating dispatching scheme according to matching parameters; collecting actual execution results, updating the weight coefficients and arbitration rules of each field according to the execution deviation, and updating the matching decision threshold. The present application realizes accurate analysis of labor demand and dynamic optimization of dispatching scheme.

Description

Technical Field

[0001] This invention belongs to the field of intelligent employment management technology, specifically involving an employment management method based on Chinese semantic parsing and multi-agent collaborative decision-making. Background Technology

[0002] As enterprises' employment needs become increasingly complex and diversified, traditional employment management methods largely rely on manual processing, resulting in inefficiency, strong subjectivity, and slow response. In recent years, the rapid development of natural language processing and artificial intelligence technologies has provided new technological pathways for the intelligentization of employment management.

[0003] In terms of semantic parsing, existing solutions mainly focus on feature extraction from general texts, lacking specialized parsing models for multi-dimensional entities such as job positions, tasks, time, location, number of people, skills and qualifications, and their interrelationships in employment demand texts. This makes it difficult to transform unstructured employment demand texts into structured, computable semantic structures.

[0004] In terms of multi-agent collaborative decision-making, existing research has applied multi-agent technology to scheduling and decision-making scenarios. However, existing systems are mostly geared towards workshop production or general human resource scenarios. They lack agent division architecture for differentiated tasks such as demand extraction, feature mining, scenario recognition, and benchmark calibration in employment management. Furthermore, they lack a mechanism for arbitration agents to resolve conflicts and weightedly integrate the results extracted by multiple agents.

[0005] In terms of personnel matching and assignment, most existing matching schemes adopt static profiles or single-match strategies. On the one hand, the generation of employment demand benchmarks lacks a dynamic mapping mechanism based on multi-dimensional demand feature vectors. On the other hand, the updating of personnel capability profiles lacks a time decay feedback mechanism based on actual execution results, and the matching decision threshold also lacks adaptive adjustment capabilities based on execution deviations.

[0006] In summary, existing employment management technologies lack deep semantic analysis capabilities for Chinese employment demand texts, lack multi-agent collaboration and conflict arbitration mechanisms for employment management scenarios, and lack dynamic closed-loop optimization capabilities based on execution feedback. Therefore, there is an urgent need for an employment management method that integrates Chinese semantic analysis with multi-agent collaborative decision-making. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides an employment management method based on Chinese semantic parsing and multi-agent collaborative decision-making.

[0008] The objective of this invention can be achieved through the following technical solution: a labor management method based on Chinese semantic parsing and multi-agent collaborative decision-making, comprising: Receive employment demand text, extract job positions, tasks, time, location, number of people, skills and qualification constraints through a semantic parsing model, and generate the demand semantic structure; The semantic structure of the requirements is distributed to the basic extraction agent, the implicit mining agent, the scene recognition agent, and the benchmark calibration agent, which extract basic fields, implicit fields, scene fields, and benchmark initial fields, respectively. The arbitration agent performs conflict resolution and weighted fusion on the fields according to compliance priority, field confidence and historical performance success rate, and outputs a multi-dimensional requirement feature vector. The multidimensional demand feature vector is input into the dynamic demand feature model to generate the labor demand benchmark; the labor demand benchmark is matched and calculated with the dynamic ability profile of personnel to obtain matching parameters, and a work assignment plan is generated based on the matching parameters. Collect actual execution results, update the weight coefficients and arbitration rules of each field according to the execution deviation, and update the matching decision threshold at the same time.

[0009] Specifically, the generation of the demand semantic structure is as follows: The semantic parsing model employs a sequence labeling network and a relation classification network. The sequence labeling network uses a BIO labeling strategy to identify the start and end boundaries of each entity in the requirement text. The relation classification network uses each pair of entities as input to determine its relation type, and constructs the requirement semantic structure using the identified entities as nodes and the relation types as edges.

[0010] Specifically, after distributing the aforementioned demand semantic structure to each agent: The basic extraction agent extracts explicitly appearing job positions, tasks, times, locations, number of people, skills, and qualifications as the basic fields using regular expression matching. The implicit mining agent queries a preset implicit feature mapping table based on job type and task description, and outputs work intensity, team collaboration requirements, urgency level and overtime preference as the implicit fields. The scene recognition agent inputs the semantic structure of the demand into the scene classifier and outputs the work environment type, job risk level and seasonality coefficient as the scene field. The benchmarking agent retrieves historical employment data using job type and region as indexes, and outputs the hourly rate, average performance, and rating distribution of similar jobs as the initial fields of the benchmark.

[0011] Specifically, the multidimensional demand feature vector is generated in the following way: The arbitration agent filters the output values ​​of each agent according to compliance priority to construct a compliance candidate set; The product of the field confidence level and the historical performance success rate of each agent is used as the initial weight of the corresponding value; Calculate the average deviation between each candidate value and other values ​​in the compliant candidate set, and perform inverse proportional decay correction on the initial weights based on the average deviations to obtain the final decision weights; The values ​​within the compliance candidate set are weighted and fused based on the final decision weight to generate the multidimensional demand feature vector.

[0012] Specifically, the historical performance success rate is updated as follows: After each fusion of the multidimensional demand feature vector, the arbitration agent compares the fusion result with the actual execution result dimension by dimension. If the execution meets the target, each agent contributing to the fusion result is incremented by one effective count; otherwise, no increment is made. The historical performance success rate of each agent is updated by the ratio of the effective count of each agent to the total number of outputs.

[0013] Specifically, the method for generating the labor demand benchmark is as follows: The multidimensional demand feature vector is input into the dynamic demand feature model, which maps the continuous values ​​of each dimension in the vector to the corresponding job level range, working hour fluctuation range, qualification baseline threshold and scenario adaptation weight, and combines them to generate the employment demand benchmark.

[0014] Specifically, the matching calculation is performed in the following manner: The employment demand benchmark and the personnel dynamic capability profile are mapped to feature vectors in the same multi-dimensional feature space; Calculate the deviation values ​​of the two feature vectors in each dimension, and calculate the weighted Euclidean distance based on the dynamic weight coefficients of each dimension; The initial matching degree is determined based on the reciprocal of the weighted Euclidean distance, and the initial matching degree is normalized to obtain the matching parameters.

[0015] Specifically, the process for generating the dispatch plan is as follows: Arrange all candidates in descending order according to the matching parameters. Starting from the first candidate, verify whether the candidate's qualifications meet the minimum qualification requirements in the employment demand benchmark and whether the available time period overlaps with the employment time in the employment demand benchmark. Select the first candidate who meets both of the above conditions into the dispatch plan, remove the candidate from the candidate list, and repeat the above verification and selection process.

[0016] Specifically, the update method for the dynamic capability profile of personnel is as follows: The dynamic capability profile of the personnel includes: skill proficiency, historical performance score, available time period, geographical location matching degree, and recent fatigue index, and each dimension is stored in vector form; After each work assignment is completed, data on the actual attendance, task completion quality, collaboration evaluation, and working hours of the personnel are obtained. The deviation between the data and the corresponding dimension of the current employment demand benchmark is calculated, and the original profile vector is updated by weighting according to the time decay factor. The time decay factor decreases as the time interval between the current update increases.

[0017] Specifically, the process of updating the weight coefficients and arbitration rules of each field, and simultaneously updating the matching decision threshold, is as follows: The actual execution results are compared with the expected target dimension by dimension, including basic fields, implicit fields, scenario fields, and baseline initial fields, and the deviation vector is calculated for each dimension. If the magnitude of the deviation vector exceeds the preset tolerance range, the weight adjustment amount is allocated according to the proportion of the deviation value of each dimension and superimposed on the current weight coefficient of the corresponding field. At the same time, the matching decision threshold is adjusted according to the statistical results of the deviation vector of each field. The conflict resolution parameters inside the arbitration intelligent body are dynamically adjusted according to the historical performance success rate of each intelligent agent, including the initial weight allocation of the output value of each intelligent agent and the priority offset when adjudicating conflicts.

[0018] The beneficial effects of this invention are: (1) This invention performs sequence labeling and relation classification on employment demand text through a semantic parsing model. It constructs a demand semantic structure with entities as nodes and relation types as edges. It can accurately extract multi-dimensional entities and their interrelationships from unstructured Chinese employment demand text, thereby improving the accuracy and structuring of employment demand parsing.

[0019] (2) This invention achieves effective integration of multi-source heterogeneous features by dividing and coordinating the basic extraction agent, the implicit mining agent, the scene recognition agent and the benchmark calibration agent, and by having the arbitration agent resolve conflicts and perform weighted fusion based on compliance priority, field confidence and historical performance success rate, thereby improving the comprehensiveness and accuracy of labor demand feature extraction.

[0020] (3) This invention generates a labor demand benchmark through a dynamic demand feature model, performs matching calculations by combining a personnel capability profile that is dynamically updated by a time decay factor, and updates the field weight coefficients, matching decision thresholds and arbitration rules based on the actual execution results, forming a closed-loop iterative optimization mechanism, which improves the matching accuracy and adaptability of the dispatching scheme. Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating the employment management method based on Chinese semantic parsing and multi-agent collaborative decision-making of the present invention. Figure 2This is a decision logic diagram of multi-agent arbitration fusion and feature space matching in this invention. Detailed Implementation

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0024] Please see Figure 1-2 Employment management methods based on Chinese semantic parsing and multi-agent collaborative decision-making include: Receive employment demand text, extract job positions, tasks, time, location, number of people, skills and qualification constraints through a semantic parsing model, and generate the demand semantic structure; The semantic structure of the requirements is distributed to the basic extraction agent, the implicit mining agent, the scene recognition agent, and the benchmark calibration agent, which extract basic fields, implicit fields, scene fields, and benchmark initial fields, respectively. The arbitration agent performs conflict resolution and weighted fusion on the fields according to compliance priority, field confidence and historical performance success rate, and outputs a multi-dimensional requirement feature vector. The multidimensional demand feature vector is input into the dynamic demand feature model to generate the labor demand benchmark; the labor demand benchmark is matched and calculated with the dynamic ability profile of personnel to obtain matching parameters, and a work assignment plan is generated based on the matching parameters. Collect actual execution results, update the weight coefficients and arbitration rules of each field according to the execution deviation, and update the matching decision threshold at the same time.

[0025] Example 1 This example illustrates a temporary workforce scenario at a manufacturing company. The company recently received a batch of urgent orders, but its existing staff was unable to complete the production tasks within the agreed timeframe, necessitating the recruitment of temporary workers from outside. The company's workforce management personnel input the workforce requirement text through the system interface, and the system initiates the intelligent workforce management process upon receiving the text.

[0026] A specific employment requirement document reads: "Project A urgently needs M electricians holding certificate C, with working hours from date T1 to date T2, and working location in region L." Upon receiving the aforementioned employment request text, the semantic parsing model is activated. The semantic parsing model comprises a sequence labeling network and a relation classification network.

[0027] 1. Perform entity recognition Sequence labeling networks employ a BIO (Browser-Initial-Output) labeling strategy to sequentially label employment demand text. In the BIO strategy, B represents the start of an entity, I represents the interior of an entity, and O represents non-entity parts. The sequence labeling network labels each word in the text with one of these three types of tags. Taking the qualification entity in "Project A urgently needs M electricians holding certificate C" as an example, when the word "certificate" is scanned, the sequence labeling network recognizes it as the start of the qualification entity and labels it as B-QUAL. The following "C" is labeled as I-QUAL, indicating that the character belongs to the interior of the qualification entity. By determining the complete boundary of "certificate C" in this way, the location and classification of each entity are completed.

[0028] 2. Relationship Classification The relation classification network takes each pair of entities as input and determines the type of relationship between them. Specifically, the network extracts contextual features from the entity pairs, including the semantic vectors of each entity, the positional interval between the two entities in the text, and the dependency path features between the two entities. These features are concatenated and input into a fully connected layer. After passing through a softmax function, the probability distribution of the entity pair belonging to various relation labels is calculated, and the relation label with the highest probability is selected as the relation type determination result for the entity pair. Taking the entity pair "Project A" and "Electrician" as an example, the network extracts the semantic vectors of "Project A" and "Electrician," obtains the positional interval between them in the text, and extracts the dependency path features obtained from dependency parsing. These features are concatenated and input into a fully connected layer. After softmax calculation, the network outputs that the relationship "Job-Skill" has the highest probability, so the relationship of the entity pair is determined to be a job-skill relationship, that is, the job requires the skill of an electrician.

[0029] For "Electrician" and "Certificate C", the relationship with the highest output probability is a skill-qualification relationship, meaning that the qualification requirement for this skill is holding Certificate C. Similarly, "Project A" and "M people" are identified as a position-number relationship, meaning that the number of people required for this position is M; "Project A" and "Time T1 to Time T2" ​​are identified as a position-time relationship, meaning the working hours for this position; and "Project A" and "Region L" are identified as a position-location relationship, meaning the work location for this position.

[0030] Using identified entities as nodes and relationships between entities as edges, a demand semantic structure is constructed. The semantic structure construction module in the system creates an empty directed graph structure, adds nodes to the directed graph according to the previously identified entities, and labels each node with an entity type, including job title nodes, number of employees nodes, qualification nodes, skill nodes, time nodes, and location nodes. Then, directed edges are added to the directed graph according to the relationship types output by the aforementioned relationship classification network, with each directed edge containing a relationship type label. This generates a demand semantic structure containing the aforementioned entity nodes and relationship edges. This demand semantic structure fully represents all the key information in the employment demand text and their interrelationships, and is output to subsequent agents in the form of a standardized graph structure as input data for multi-agent extraction in subsequent steps.

[0031] Example 2 This embodiment builds upon the generation result of the semantic structure of the above requirements, and further explains the specific process of multi-agent extraction and arbitration fusion.

[0032] The system simultaneously distributes the requirement semantic structure constructed in Example 1 to four agents: the basic extraction agent, the implicit mining agent, the scene recognition agent, and the benchmark calibration agent.

[0033] 1. Extraction of basic fields After receiving the semantic structure of the requirements, the basic extraction agent extracts explicitly appearing fields using regular expression matching. Specifically, the basic extraction agent has multiple built-in regular expression rules, such as "Project[AZ]" for matching job titles, "[X] people" for matching number of employees, "Electrician|Welder|Fitter" for matching skills, "Certificate[AZ]" for matching qualifications, "Date T[0-9] to Date T[0-9]" for matching time, and "Region [AZ]" for matching location. The basic extraction agent sequentially matches the text content of each entity node in the semantic structure of the requirements with the above regular expressions. If a match is successful, the text value of that entity is extracted as the value of the corresponding field. Taking "Project A" as an example, the regular expression "Project [AZ]" matches successfully, and "Project A" is extracted as the value of the job field; "X people" matches the number of people regular expression, and X is extracted as the number of people; the other fields are processed in the same way, and finally the basic field set is obtained, including job "Project A", task "electrician", time "T1 to T2", location "region L", number of people X, skill "electrician", and qualification "certificate C".

[0034] 2. Extraction of Hidden Fields After receiving the semantic structure of the demand, the implicit feature mining agent queries a pre-built implicit feature mapping table based on the job type and task description. The implicit feature mapping table is a pre-constructed two-dimensional table structure, where each row corresponds to a combination of job type and task description, and each column corresponds to an implicit feature dimension, including workload coefficient, teamwork requirement, urgency level, and overtime preference. This mapping table is pre-annotated based on historical employment data and industry expert knowledge. The implicit feature mining agent extracts the text values ​​of the job entity "Project A" and the skill entity "Electrician" from the demand semantic structure. Using the job type of "Project A" and the task description of "Electrician" as a joint index, it locates the corresponding row in the implicit feature mapping table and reads the values ​​of the four columns—work intensity coefficient, teamwork requirement, urgency level, and overtime preference—from that row, outputting them as implicit fields. For example, if the query yields a workload coefficient of "High," a teamwork requirement of "Medium," an urgency level of "Urgent," and an overtime preference of "Acceptable overtime," then these four values ​​are output as implicit fields.

[0035] 3. Scene Field Extraction After receiving the semantic structure of the demand, the scene recognition agent inputs it into the scene classifier. The scene classifier is a pre-trained classification model containing an input layer, a hidden layer, and an output layer. The input layer receives a feature vector formed by concatenating the semantic vectors of each entity node in the demand semantic structure. The hidden layer performs a non-linear transformation on the input features. The output layer calculates the probability distribution of each scene label using the softmax function and selects the label with the highest probability as the classification result. The output dimensions of the scene classifier include three types of labels: work environment type, job risk level, and seasonality coefficient. Taking the current input as an example, the location entity in the demand semantic structure is "Region L", the job entity is "Project A", and the skill entity is "Electrician". The scene classifier extracts the semantic vectors of the above entities, concatenates them, and inputs them into the model. The output work environment type is "Factory Workshop" (highest probability), job risk level is "Medium" (highest probability), and seasonality coefficient is "S" (S is a constant greater than zero, representing the seasonal impact factor under this scene). These three values ​​are output as scene fields.

[0036] 4. Extraction of initial benchmark fields After receiving the demand semantic structure, the benchmarking agent retrieves historical employment databases indexed by job type and region. The historical employment database stores complete records of all previous employment tasks, with each record containing fields such as job type, work region, hourly rate, performance score, and employee rating. The benchmarking agent extracts the job type of the job entity "Project A" and the region name of the location entity "Region L" from the demand semantic structure. Using "Job Type = Electrician" and "Region = Region L" as joint search criteria, it queries all records in the historical employment database that meet these criteria. For the retrieved record set, it calculates the median hourly rate, the average performance score, and the distribution ratio of each rating (Excellent / Good / Satisfactory / Unsatisfactory). These three statistical values ​​are used as the hourly rate benchmark, the average performance score benchmark, and the rating distribution benchmark, respectively, and combined as the initial benchmark fields.

[0037] 5. Arbitration Integration After each of the four agents completes its extraction, the basic fields, implicit fields, scenario fields, and baseline initial fields are simultaneously input into the arbitration agent.

[0038] 5.1 Consistency Check and Conflict Marking The arbitration agent first performs a consistency check on the extraction results of the same field by each agent. Taking the "number of people" field as an example, the basic extraction agent directly extracts the number of people X from the requirement text, the implicit mining agent infers the required number of people (X+Y) from the mapping table based on the "high" work intensity coefficient (where Y is the additional number of people offset corresponding to the work intensity coefficient), the scene recognition agent does not involve the extraction of the number of people (output is empty), and the benchmark calibration agent obtains the typical number of people range for this position from (X-A1) to (X+A2) based on historical data statistics, and the recommended value is (X-A3) (where A1, A2, and A3 are the lower limit offset, upper limit offset, and recommended value offset obtained from historical data statistics, respectively). Thus, the "number of people" field has four different extraction results: X, (X+Y), empty value, and (X-A3). The arbitration agent marks this situation as a conflict, and the corresponding conflict field is recorded as "number of people", with the conflict value being one of the four values ​​mentioned above.

[0039] 5.2 Compliance Priority Screening Subsequently, the arbitration agent filters based on compliance priority. Compliance priority is ranked from highest to lowest according to mandatory legal provisions, internal company policies, and industry recommended standards. The "Number of Employees" field does not involve mandatory legal provisions or internal company policies, therefore it does not trigger elimination. Taking the "Qualification" field as an example, the basic extraction agent extracts the qualification "Certificate C," the implicit mining agent queries the mapping table based on the job type "Electrician" and obtains the qualification "Certificate C or Certificate D," the benchmarking agent statistically analyzes historical data to obtain the qualification requirement for similar positions as "Certificate C," and the scene recognition agent matches the qualification requirement from the scene rule base based on the job risk level "Medium" as "Certificate C." Among the four extraction results, "Certificate D" exists. However, the internal company policy stipulates that "electrician positions involving medium or higher risk levels must hold Certificate C." Therefore, "Certificate D" violates the highest priority rule, and the arbitration agent eliminates "Certificate D." The remaining value, "Certificate C," participates in subsequent fusion.

[0040] 5.3 Weighted Summation and Fusion After the elimination is completed, the arbitration agent executes an inverse weighted correction mechanism based on group bias. Each agent's output for each field carries a confidence value Q, and the historical performance success rate R of each agent is maintained by the system in real time.

[0041] For a candidate value V of the conflict field, the arbitrating agent first calculates its initial fusion weight W: W=Q×R Where Q is the field confidence level of the agent corresponding to this value, and R is the historical performance success rate of the agent.

[0042] Subsequently, the arbitration agent calculates the average absolute deviation (Davg) of the value V relative to other values ​​within the compliance candidate set. Taking the "number of people" field as an example, assuming the set of values ​​within the compliance candidate set is {X, X+Y, X-A3}, the average absolute deviation of the value X is calculated as follows: Davg=(∣X (X+Y)∣+∣X (X A3)∣) / 2 The arbitral agent determines the inverse weighted coefficient A of the deviation based on the mean absolute deviation Davg:

[0043] Where L is a preset adjustment factor, and the value range of A is between (0,1], and decreases as the mean absolute deviation Davg increases.

[0044] Next, the arbitrary agent multiplies the initial fusion weight W by the bias inverse weighting coefficient A to obtain the final decision weight Wf: Wf = W × A Finally, the arbitration agent performs a weighted aggregation of the values ​​within the compliance candidate set based on the final decision weights. Using the "number of people" field example above, the aggregated number of people is: Vfinal=(Wf1×X+Wf2×(X+Y)+Wf3×(X A3)) / (Wf1+Wf2+Wf3) The calculated result is rounded down and used as the final feature value of the field.

[0045] Example 3 This embodiment builds upon the above-mentioned multi-dimensional demand feature vector generation results, and further explains the specific process of generating the labor demand benchmark, matching calculation, and dispatching scheme.

[0046] The system inputs the multidimensional demand feature vector generated by the arbitration agent in Example 2 into the dynamic demand feature model.

[0047] 1. Generation of labor demand benchmarks The dynamic demand feature model is a pre-trained mapping model based on historical employment data, consisting of an input layer, multiple hidden layers, and an output layer. The input layer receives the numerical values ​​of each dimension of the multi-dimensional demand feature vector, the hidden layers extract the interaction features between the dimensions through non-linear transformations, and the output layer maps the output of the hidden layers to the target value range of each dimension.

[0048] After the multidimensional demand feature vector is input into the dynamic demand feature model, the output layer outputs the mapping results of each dimension.

[0049] Job level range: Based on the job type "electrician", skill "electrician", number of people X, and historical job level distribution data for the same job, the model divides the job into three levels: "junior", "intermediate", and "senior". The "intermediate" level is taken as the core level, and an interval with one sub-level before and after it is generated. That is, the job level range is "intermediate to senior".

[0050] Working hours fluctuation range: Based on the time "T1 to T2", the seasonality coefficient "S" and the actual working hours statistics of similar positions in historical employment data, the model calculates the standard working hours as H0 hours. Combining the work intensity coefficient "high" and the urgency level "urgent", the working hours fluctuation range is set to fluctuate downward by H1 hours and upward by H2 hours, that is, the working hours fluctuation range is "(H0-H1) to (H0+H2) hours".

[0051] The qualification baseline threshold is set as "holding a certificate of C" based on the job type "electrician", qualification "certificate C" and scenario risk level "medium" and combined with the company's internal regulations on qualification management for electrician positions. This means that the minimum qualification requirement for candidates is to hold a certificate of C, which must be valid.

[0052] Scene adaptation weights: The model generates the weight allocation of each scene feature in the subsequent matching calculation based on the work environment type "factory workshop", the operation risk level "medium" and the seasonality coefficient "S". The scene adaptation weights are output in vector form, and the sum of the weight values ​​of each dimension is 1.

[0053] The dynamic demand feature model combines the above-mentioned job level range, working hour fluctuation range, qualification threshold and scenario adaptation weight to generate a structured employment demand benchmark, which is then output to the matching engine.

[0054] 2. Construction and matching calculation of dynamic personnel capability profiles The system retrieves the profile data of each candidate from the dynamic competency profile database. The matching engine then performs matching calculations between the employment demand benchmark and the dynamic competency profile of each candidate.

[0055] The employment demand benchmark includes job level, working hours, minimum qualification requirements, and scenario-appropriate weights for each dimension. The system extracts the standard values ​​of each dimension from the employment demand benchmark to construct a benchmark vector B; and extracts the actual values ​​of the candidates, such as skill proficiency, available working hours, geographical location matching degree, and fatigue index, to construct a profile vector P. The dimensions of B and P correspond one-to-one; for example, the first dimension represents skills, the second dimension represents working hours, the third dimension represents geographical location, and so on.

[0056] Taking a candidate P1 as an example, the matching calculation process between its profile vector P1 and the baseline vector B is as follows: Step 1: Construct the deviation vector D The system calculates the absolute value of the difference between the baseline vector B and the image vector P dimension by dimension.

[0057] Let the value of the k-th dimension of the baseline vector B be Bk, and the value of the k-th dimension of the image vector P be Pk.

[0058] In the k-th dimension, the system calculates the deviation value Dk: Dk = |Bk Pk∣ For example, if the first dimension is the skill dimension, the baseline value B1 is 80 points, and the skill value P1 of person P1 is 75 points, then the deviation D1 is 5; if the second dimension is the working hours dimension, the baseline value B2 is 8 hours, and the working hours P2 of person P1 is 9 hours, then the deviation D2 is 1.

[0059] The system iterates through all n dimensions, calculates D1, D2, and so on up to Dn in sequence, and arranges these deviation values ​​in dimensional order to form a deviation vector D = (D1, D2, ..., Dn).

[0060] If the k-th dimension involves qualification baseline verification (such as certificate validity), and personnel P1 does not meet the requirements, then Dk is directly set to a maximum value M (such as 10 to the power of 6) so that the personnel can be directly excluded in subsequent calculations.

[0061] Step 2: Calculate the weighted Euclidean distance Dist The system obtains the dynamic weight coefficients Wk of each dimension output by the arbitration agent.

[0062] The system calculates the weighted Euclidean distance Dist based on the bias vector D and the weight coefficients Wk.

[0063] This distance value quantitatively represents the overall gap between candidates and employment needs. The smaller the Dist value, the closer the candidate is to the standard in all dimensions, and the higher the match.

[0064] Step 3: Calculate the matching parameter Mp To prevent errors such as zero denominator during calculation, the system introduces a very small anti-overflow constant E (e.g., E = 0.000001).

[0065] The system calculates the matching parameter Mp: Mp = 1 / (Dist + E) Normalize Mp to the interval [0, 1] to obtain the final matching parameters. If Dk is a maximum value M due to qualification mismatch in the previous steps, then the calculated Mp here will approach 0, and the system will automatically determine that the person is unqualified.

[0066] The matching engine performs the above vector construction, deviation calculation, distance measurement and parameter mapping operations on each candidate, and outputs the matching parameters corresponding to each candidate.

[0067] 3. Work assignment plan generation The dispatch plan generation module receives each candidate and its matching parameters output by the matching engine, and sorts all candidates in descending order of matching parameters from high to low.

[0068] Starting from the first candidate in the sorting, the work assignment generation module verifies two constraints for each candidate: First, verify whether the candidate's qualifications meet the minimum qualification requirements in the employment needs benchmark, that is, whether the candidate's valid qualification certificate includes "Certificate C".

[0069] Second, verify whether the candidate's available time period overlaps with the employment time T1 to T2 in the employment demand benchmark, that is, whether the candidate's available working time period overlaps with T1 to T2.

[0070] Taking P1, the person with the highest matching parameter in the sorting results, as an example, the work assignment generation module first checks P1's qualification certificate list to confirm that it contains "Certificate C" and is within its validity period, thus passing the qualification verification; then it checks P1's available time period to confirm that it intersects with T1 to T2, thus passing the time verification. If both conditions are met, P1 is selected into the work assignment plan.

[0071] After removing P1 from the candidate list, the dispatch plan generation module continues to perform the above two checks on the remaining candidates one by one in the original sorting order, and selects the next person P2 who meets both conditions to be added to the dispatch plan. The above check and selection process is repeated until the number of people selected into the dispatch plan reaches the required number X.

[0072] Finally, the dispatch plan generation module outputs a dispatch plan containing X candidates. These candidates are the top X individuals who meet the minimum qualification requirements and available time constraints, and whose matching parameters are optimal. The dispatch plan is then sent to the execution system for dispatching.

[0073] Example 4 This embodiment builds upon the above-described dispatching scheme, further illustrating the specific process of collecting actual execution results and updating closed-loop feedback.

[0074] The system outputs the dispatch plan generated in Example 3 to the execution system. The execution system dispatches X candidates to Project A according to the dispatch plan, and they complete the electrician's work tasks within the time period T1 to T2. After the dispatch execution is completed, the system enters the feedback and update phase.

[0075] 1. Collection of actual execution results The system collects the actual execution result data of this dispatch from multiple data sources.

[0076] The attendance data collection module obtains the actual attendance records of each dispatched worker from the attendance system, including whether they arrived on time and whether they were absent or late or left early.

[0077] The task completion quality data collection module obtains the task completion quality score of each dispatcher from the project supervisor's acceptance evaluation system. This score is entered by the project supervisor through the system interface after the task is completed, and covers sub-dimensions such as workload completion rate, work qualification rate and number of rework.

[0078] The collaboration evaluation data collection module obtains the collaboration evaluation score of each dispatcher from the project team member peer evaluation system. This score is generated by other members of the same project team scoring based on communication efficiency, cooperation level and team contribution, and then summing and averaging the scores.

[0079] The work duration data acquisition module obtains the total actual working time of each dispatcher from the work time recording system, which is recorded in hours.

[0080] Assume that X people are dispatched in this job assignment. Take one of them, P1, as an example. The data collection system obtains P1's attendance data as perfect attendance (no absences, no lateness or early departures), task completion quality score as Q1 (score range 0 to 100), collaboration evaluation score as E1 (score range 0 to 100), and actual working hours as Hr1 (in hours).

[0081] 2. Update of personnel dynamic capability profile After the data collection is completed, the system will update the profile of person P1 as an example.

[0082] The dynamic capability profile of personnel is stored in the personnel profile database in the form of a multi-dimensional vector. The dimensions are skill proficiency, historical performance score, available time period, geographical location matching degree, and recent fatigue index. Each dimension is represented numerically.

[0083] The system extracts the expected standard values ​​for each dimension from the current employment demand benchmark. For personnel P1, the corresponding dimension standard values ​​in the current employment demand benchmark include the skill benchmark value B1 (this value is derived by the dynamic demand feature model based on the job level "intermediate to advanced") and the working hour benchmark value B2 (the standard working hours calculated by the dynamic demand feature model based on historical employment data).

[0084] For the skill proficiency dimension, the system calculates the difference between P1's task completion quality score Q1 and the skill benchmark value B1 in the employment demand baseline, resulting in the deviation Dq1 = Q1 - B1. If the deviation is positive, it indicates that P1's skill performance is higher than the job benchmark requirement, and the skill proficiency should be adjusted upward; if the deviation is negative, it indicates that P1's performance is lower than the benchmark requirement, and the skill proficiency should be adjusted downward.

[0085] For the historical performance evaluation dimension, the system will comprehensively assess the current execution result as "meets standards" or "fails standards". The comprehensive evaluation rules are as follows: attendance data is perfect, task completion quality score Q1 is not lower than the preset standard Tq, and collaboration evaluation score E1 is not lower than the preset standard Te. If all three conditions are met, it is considered "meets standards"; otherwise, it is considered "fails standards". Assuming P1 meets all three conditions, the current performance result is considered "meets standards", and the historical performance score will be adjusted upward accordingly.

[0086] For the available time period dimension, the system calculates the difference between P1's actual working time Hr1 and the baseline working time value B2, resulting in the deviation Dt1 = Hr1 - B2. If the deviation is positive, it means that P1's actual available working time is greater than the baseline value, and its available time period score is adjusted upward accordingly; if the deviation is negative, it is adjusted downward.

[0087] For the geolocation matching dimension, the actual commuting situation of P1 in this execution was obtained by the system from the location or check-in data, and the deviation Dl1 between the actual commuting time and the expected commuting time was calculated. This deviation was used to adjust the geolocation matching score.

[0088] Regarding the recent fatigue index dimension, the system updates the fatigue index according to the preset fatigue calculation formula based on the actual working hours Hr1 and the number of consecutive working days of P1. The longer the working hours and the more consecutive working days, the greater the increase in the fatigue index.

[0089] After calculating the deviations for each dimension, the system performs a weighted update on the original portrait vector. Specifically, for a certain dimension of the portrait vector, the new value equals the original value plus the product of the deviation value and the preset update step size, multiplied by the time decay factor F. The time decay factor F ranges from 0 to 1, and its value decreases as the time interval since the last update increases. That is, the older the historical data is, the smaller its impact on the current portrait, and the greater the impact of recent data. Taking the skill proficiency dimension as an example, if the deviation Dq1 of P1 is positive, then its new skill proficiency value is the original value plus the product of Dq1 and the update step size, multiplied by the current time decay factor F. The skill proficiency field in the portrait database is updated to the new value.

[0090] The system performs the above data collection and profile update operations on each of the X dispatched personnel, updating the corresponding dimension values ​​of each personnel in the profile database.

[0091] 3. Weight coefficient update and matching decision threshold update Meanwhile, the system updates the weight coefficients and matching decision thresholds of each field based on the deviation of the actual execution results.

[0092] The system compares the actual execution results with the expected goals dimension by dimension, categorized into four groups: basic fields, implicit fields, scenario fields, and baseline initial fields. The basic field group includes seven dimensions: job position, task, time, location, number of people, skills, and qualifications. The implicit field group includes four dimensions: workload coefficient, team collaboration needs, urgency level, and overtime preference. The scenario field group includes three dimensions: work environment type, operational risk level, and seasonality coefficient. The baseline initial field group includes three dimensions: hourly rate benchmark, average performance benchmark, and rating distribution benchmark.

[0093] Taking the "Number of People" dimension in the basic field group as an example, the expected number of people needed in this project is X, and the actual number of people on duty is Xr, with a deviation of Dn = Xr - X. If Dn is 0, it means that the number of people predicted is accurate and no adjustment is needed; if Dn is not 0, it means that there is a deviation in the number of people predicted.

[0094] The system calculates a deviation vector for each group, which is composed of the deviation values ​​of each dimension within that group. Taking the basic field group as an example, its deviation vector is composed of job matching deviation, task completion deviation, time deviation, location deviation, number of people deviation Dn, skill deviation, and qualification deviation.

[0095] After calculating the deviation vectors for all groups, the system determines whether the magnitude of each deviation vector exceeds the preset tolerance range Ro. The magnitude of the deviation vector is calculated by taking the square root of the sum of the squares of the deviation values ​​in each dimension. The preset tolerance range Ro is a pre-set threshold. If the magnitude of the deviation exceeds Ro, it indicates that the overall deviation of the group is large, and weight adjustment is required.

[0096] Taking a basic field group as an example, if its deviation vector magnitude exceeds the preset tolerance range Ro, the system proportionally allocates weight adjustment amounts to the deviation values ​​of each dimension within that group. Specifically, for the number of people dimension, the ratio of the absolute value of its deviation Dn to the sum of the absolute values ​​of the deviations of all dimensions in that group is the weight adjustment ratio for the number of people dimension. This ratio is multiplied by the preset total adjustment amount Ao to obtain the weight adjustment amount for the number of people dimension, which is then added to the current weight coefficient Wn corresponding to the number of people dimension. Dimensions with larger deviations receive more weight adjustment amounts.

[0097] The system adjusts the matching decision threshold based on the statistical results of the deviation vector for each field. Specifically, the system calculates the overall deviation distribution between the matching parameters and actual performance of all candidates during the current execution. If the overall deviation is generally high, it indicates that the current matching decision threshold is too low, resulting in too many qualified candidates, and the threshold should be appropriately increased. Conversely, if the overall deviation is generally low, it indicates that the current matching decision threshold is too high, resulting in too few qualified candidates, and the threshold should be appropriately decreased. The adjustment range is calculated based on the statistical characteristics of the deviation vector.

[0098] 4. Arbitration Rules Update The system also dynamically adjusts the conflict resolution parameters within the arbitration agent based on the historical performance success rate of each agent.

[0099] The historical success rate is calculated as the ratio of the number of times each agent's historically adopted outputs were actually executed to the total number of outputs. See Example 5 for the specific update process. When an agent's historical success rate changes, the arbitration agent adjusts its internal conflict resolution parameters accordingly.

[0100] The conflict resolution parameters consist of two parts: the initial weight assignment of each agent's output value and the priority offset during conflict resolution.

[0101] The initial weight allocation is used to set the baseline weights in the weighted fusion phase. Assuming the historical success rate of the base extraction agent is updated from the old value Rs1 to a new value Rs2, if Rs2 is higher than Rs1, it indicates that the agent's output reliability has improved, and the arbitration agent increases the initial weight allocation value of that agent in subsequent fusion; if Rs2 is lower than Rs1, its initial weight allocation value is decreased. Specifically, the new historical success rate value Rs2 is used as the weight adjustment coefficient, multiplied by the original initial weights, and then normalized again so that the sum of the initial weights of all agents is always 1.

[0102] Priority offsets are used in the compliance priority screening phase. When multiple conflicting values ​​exist for the same field and there are no mandatory legal provisions or hard constraints from internal company policies, the historical performance success rate of each agent determines the priority offset of its output value. Agents with higher historical performance success rates receive a positive priority offset in conflict adjudication, meaning their output values ​​are adopted first without violating the highest priority rule; agents with lower historical performance success rates receive a negative priority offset, meaning their output values ​​are adopted later without violating the highest priority rule. Specifically, the arbitration agent maps each agent's historical performance success rate to a priority offset Op, using either a linear or piecewise mapping. The offset Op is then added to the baseline priority of each agent's output value, and the final priority order for conflict adjudication is determined after reordering.

[0103] The arbitration agent writes the updated initial weight allocation and priority offset into its internal configuration for use in the subsequent conflict resolution and weighted fusion stages of the dispatching tasks.

[0104] This completes the full closed-loop iteration of this task assignment. Based on the actual execution results, the system simultaneously completed four feedback operations: updating the dynamic capability profile of personnel, adjusting the weight coefficients of each field, optimizing the matching decision threshold, and correcting the conflict resolution parameters of the arbitration agent. This allows the system to gain experience from this execution and provides data support for improving the decision-making quality of subsequent task assignments.

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A labor management method based on Chinese semantic parsing and multi-agent collaborative decision-making, characterized in that, Includes the following steps: Receive employment demand text, extract job positions, tasks, time, location, number of people, skills and qualification constraints through a semantic parsing model, and generate the demand semantic structure; The required semantic structure is distributed to the basic extraction agent, the implicit mining agent, the scene recognition agent, and the benchmark calibration agent to extract basic fields, implicit fields, scene fields, and benchmark initial fields, respectively. The arbitration agent filters out non-compliant values ​​in the conflict fields based on compliance priority, and uses the product of each agent's field confidence and historical performance success rate as weight to perform inverse weighted fusion on the remaining values ​​to generate a multi-dimensional demand feature vector. The multidimensional demand feature vector is input into the dynamic demand feature model to generate the employment demand benchmark; the employment demand benchmark and the dynamic ability profile of personnel are mapped to the same multidimensional feature space, the weighted Euclidean distance of the deviation values ​​of each dimension is calculated, the reciprocal of the weighted Euclidean distance is mapped to the matching parameter, and the work assignment scheme is generated based on the matching parameter. Collect actual execution results, update the weight coefficients and arbitration rules of each field according to the execution deviation, and update the matching decision threshold at the same time.

2. The method according to claim 1, characterized in that, The specific generation of the demand semantic structure is as follows: The semantic parsing model employs a sequence labeling network and a relation classification network. The sequence labeling network uses a BIO labeling strategy to identify the start and end boundaries of each entity in the requirement text. The relation classification network uses each pair of entities as input to determine its relation type, and constructs the requirement semantic structure using the identified entities as nodes and the relation types as edges.

3. The method according to claim 1, characterized in that, After distributing the aforementioned demand semantic structure to each agent: The basic extraction agent extracts explicitly appearing job positions, tasks, times, locations, number of people, skills, and qualifications as the basic fields using regular expression matching. The implicit mining agent queries a preset implicit feature mapping table based on job type and task description, and outputs work intensity, team collaboration requirements, urgency level and overtime preference as the implicit fields. The scene recognition agent inputs the semantic structure of the demand into the scene classifier and outputs the work environment type, job risk level and seasonality coefficient as the scene field. The benchmarking agent retrieves historical employment data using job type and region as indexes, and outputs the hourly rate, average performance, and rating distribution of similar jobs as the initial fields of the benchmark.

4. The method according to claim 1, characterized in that, The specific method for generating the multidimensional demand feature vector is as follows: The arbitration agent filters the output values ​​of each agent according to compliance priority to construct a compliance candidate set; The product of the field confidence level and the historical performance success rate of each agent is used as the initial weight of the corresponding value; Calculate the average deviation between each candidate value and other values ​​in the compliant candidate set, and perform exponential decay correction on the initial weights based on the average deviations to obtain the final decision weights; The values ​​within the compliance candidate set are weighted and fused based on the final decision weight to generate the multidimensional demand feature vector.

5. The method according to claim 4, characterized in that, The historical performance success rate is updated as follows: After each fusion of the multidimensional demand feature vector, the arbitration agent compares the fusion result with the actual execution result dimension by dimension. If the execution meets the target, each agent contributing to the fusion result is incremented by one effective count; otherwise, no increment is made. The historical performance success rate of each agent is updated by the ratio of the effective count of each agent to the total number of outputs.

6. The method according to claim 1, characterized in that, The specific method for generating the labor demand benchmark is as follows: The multidimensional demand feature vector is input into the dynamic demand feature model, which maps the continuous values ​​of each dimension in the vector to the corresponding job level range, working hour fluctuation range, qualification baseline threshold and scenario adaptation weight, and combines them to generate the employment demand benchmark.

7. The method according to claim 1, characterized in that, The matching calculation is performed in the following manner: The employment demand benchmark and the personnel dynamic capability profile are mapped to feature vectors in the same multi-dimensional feature space; Calculate the deviation values ​​of the two feature vectors in each dimension, and calculate the weighted Euclidean distance based on the dynamic weight coefficients of each dimension; The initial matching degree is determined based on the reciprocal of the weighted Euclidean distance, and the initial matching degree is normalized to obtain the matching parameters.

8. The method according to claim 1, characterized in that, The specific process for generating the dispatch plan is as follows: Arrange the candidates in descending order according to the matching parameters. Starting from the first candidate, verify whether the candidate's qualifications meet the minimum qualification requirements in the employment demand benchmark and whether the available time period overlaps with the employment time in the employment demand benchmark. Select the first candidate who meets both of the above conditions into the dispatch plan, remove the candidate from the candidate list, and repeat the above verification and selection process.

9. The method according to claim 1, characterized in that, The update method for the dynamic capability profile of the personnel is as follows: The dynamic capability profile of the personnel includes: skill proficiency, historical performance score, available time period, geographical location matching degree, and recent fatigue index, and each dimension is stored in vector form; After each work assignment is completed, data on the actual attendance, task completion quality, collaboration evaluation, and working hours of the personnel are obtained. The deviation between the data and the corresponding dimension of the current employment demand benchmark is calculated, and the original profile vector is updated by weighting according to the time decay factor. The time decay factor decreases as the time interval between the current update increases.

10. The method according to claim 1, characterized in that, The specific process of updating the weight coefficients and arbitration rules of each field, and simultaneously updating the matching decision threshold, is as follows: The actual execution results are compared with the expected goals dimension by dimension, including basic fields, implicit fields, scenario fields, and baseline initial fields, and the deviation vector is calculated for each dimension. If the magnitude of the deviation vector exceeds the preset tolerance range, the weight adjustment amount is allocated according to the proportion of the deviation values ​​of each dimension and superimposed on the current weight coefficient of the corresponding field. Simultaneously, the matching decision threshold is adjusted based on the deviation vector statistics of each field; the conflict resolution parameters inside the arbitration intelligence are dynamically adjusted based on the historical performance success rate of each agent, including the initial weight allocation of the output values ​​of each agent and the priority offset during conflict adjudication.

Images